Paul Graham

A successful startup needs only three things: good people, a product customers actually want, and minimal spending — all achievable without genius. The idea itself matters little; startups win by making something less bad than what exists, as Google did with a search site that didn't suck — indexing more of the web, ranking by links, keeping pages clean — now worth a billion dollars a year; dating sites today are just as backward. Ideas aren't transferable (VCs won't even sign an NDA to hear one), and Microsoft's original plan, selling programming languages, had nothing to do with the business IBM later handed it. What matters is people, not ideas.

Good people pass the "animal" test — obsessive about their work, whether a salesperson who won't take no or a hacker who won't sleep with a bug unfixed. For programmers, three checks apply: genuinely smart, able to get things done, and bearable to work with — the last filters out few, since true smartness (comfort saying "I don't know") usually precludes bad attitude, as Robert Morris exemplified. Startups form through friendship, hence their cluster near universities; students should work on their own projects rather than network deliberately. Ideal founder count is two to four — one person can't bear the weight alone, but beyond four, disputes harden into factions instead of forcing resolution. Founders should include technical people, since non-technical founders can't evaluate hackers or technology; business isn't a specialized discipline — only 5 of the Forbes 400's top 50 have MBAs, while most (Gates, Jobs, Ellison, Dell, Bezos) came from technical backgrounds.

Most business failures, like most restaurant failures, trace to not giving customers what they want — no startup with a massively popular product has failed. The right approach is rapid prototyping, not the "Hail Mary" of elaborate planning that burns millions building something unwanted. Viaweb pivoted from targeting web consultants and catalog companies, who resisted, to small individual online merchants, and learned a 10% gain in ease of use can double sales (per an anecdote about Stephen Hawking's editor). Watching users at trade shows is essential. Niche markets beat consumer brands as targets, since cheap products at the low end eventually swallow the high end, as Intel did to Sun, Word did to Interleaf, and Henry Ford did to earlier carmakers.

Raising money starts with seed capital from "angels" — Viaweb got $10,000 from a friend, Julian, plus legal help and an investor introduction. A business plan need only state what you'll do and how you'll profit. A near-fatal crisis hit when a founder's old employer claimed IP rights to the code mid-acquisition; the deal collapsed, but Viaweb secured new funding and was later bought by Yahoo for more than the original offer. Later VC rounds are slower, larger, and sometimes push a "newscaster" CEO — mature-seeming but shallow — which Graham avoided, also skipping an IPO; Google's own brand-averse IPO vindicated that instinct. VCs range from Sequoia and Kleiner Perkins down to weaker firms; founders hold more leverage than they think and shouldn't over-disclose.

Money, once raised, should not be spent: Bubble-era "get big fast" orthodoxy is often wrong, since brand matters little online — Google beat well-funded incumbents (Yahoo, Lycos, Excite, Altavista) without advertising. Viaweb grew slowly, about 70 users by end of 1996, letting founders master the business, unlike Yahoo, which let search "wither" while Google understood it better. Cheapness matters (Yahoo's David Filo policed disk usage); avoid impressive offices (Aeron chairs signaled trouble) and avoid hiring, a recurring cost.

A good hacker aged roughly 23 to 38 is the right founder — old enough to have learned a real job's negative lessons, young enough for the stamina and risk tolerance it demands, since a startup squeezes 40 years of work into about four with almost no leisure. It's compatible with grad school. Final message: build something users love, spend less than you make.

Smart people often build bad startups not because they lack ability but because of three correctable habits in how they pick ideas. Paul Graham draws on Y Combinator's first Summer Founders Program: sorting 227 applications, they found a need for a third bucket beyond "promising" and "unpromising" -- promising people with unpromising ideas.

He traces the pattern to his own first startup, Artix (January 1995, with Robert Morris and friends), which built websites for art galleries -- a market that turned out not to want them; dealers were technophobic and didn't want their inventory visible online. Only after abandoning Artix did the same team start Viaweb, software for building online stores, which succeeded. Microsoft likewise wasn't Bill Gates and Paul Allen's first company; that was Traf-o-data.

Three mistakes explain the failure. First, the "still life effect": founders chase the first idea that occurs to them, then mistake time already invested for evidence the idea is good -- as when painting a still life set up in minutes but worked on for a month. The fix isn't to avoid plunging into ideas but to remember later that sunk time doesn't confer merit, as with Viaweb's own name change from "Webgen," which felt essential until they lived with the replacement for three days. Second, "muck": people gravitate to work that feels cool rather than work that pays -- an old Yorkshire saying holds that "where there's muck, there's brass," and pleasant work like programming-language design pays badly because people do it for free. A startup must aim to make money first and be cool second. Third, timidity ("hyenas"): Artix targeted an unthreatening niche out of fear of VC-backed e-commerce rivals, who turned out to write bloated, ineffective software; "business" isn't one intimidating skill but separable, learnable tasks -- selling, pricing, incorporation, raising money.

Applying this to the 2005 applicant pool, Graham finds the same errors: most proposed blog/calendar/dating-site/Friendster hybrids or multiplayer games rather than mining obvious unsolved problems, citing micropayments and newspapers' economic collapse as opportunities visible in the Wall Street Journal. He attributes this to schooling, which trains students to solve assigned problems, not choose them.

The essay disputes an "entrepreneurship expert" who claimed 80% of MIT spinoffs need a business person as "voice of the customer," countering that Larry Page and Sergey Brin needed no such intermediary. Learning what customers want, Graham argues, is the real skill separating hackers from founders -- captured in the essay's bronze metaphor: raw intelligence (copper) plus empathy for others' perspective (tin, learnable partly from Dale Carnegie's How to Win Friends and Influence People) yields something far stronger than either alone.

Because starting a company now costs almost nothing, talented people in their early twenties no longer need an employer's or investor's permission to be valued at what they're worth — hiring as the default path from college is becoming obsolete.

The mechanism is cost collapse: a web startup's main expense is food and rent, so $10,000 of seed funding and a willingness to live on ramen can launch one. Cheap founding means less need for investor approval, opening the door to people who have everything backers want except experience — especially the young. Graham argues 20-year-olds are wildly undervalued: some are more capable than most 30-year-olds, but organizations, unable to distinguish them from the average (and misled by young people's individual inconsistency into filtering out false positives), route everyone through entry-level tracks priced at the mean. A company hiring you acts as a proxy for the customer, guessing your value — but you can appeal that judgment directly to users by starting your own company. Most successful startups exit not via IPO but by being bought, often before profitability, essentially as a hiring bonus (a startup acquired for $2–3 million six months in is really just recruiting).

Big companies are structurally bad at product development, for several compounding reasons: turf protection warps decisions (Microsoft's ambivalence toward web apps like Hotmail, developed outside the company); disruptive people lack internal power; big companies build only one version of each product instead of letting many competing startups try (ten startups building browsers beats one internal team); there are too many ideas for any single company to pursue (Microsoft couldn't run 500 in-house projects, roughly the number of startups gunning for it); and pay is decoupled from outcomes, so startup employees simply work harder since they get rich or get nothing. Graham predicts accelerating acquisition of ever-earlier-stage startups, once companies overcome the pride that treats buying as an admission of failure — fusing recruiting and product development, and yielding teams that already work smoothly together.

Cheap startups also shift power from investors to founders. Fairchild Semiconductor (1959), the first VC-backed startup, needed real capital for factories; a web startup today needs money mainly for sales, since word of mouth handles marketing — so founders need investors less and can resist ceding control. Google illustrates this: investors wanted an outside CEO, but the founders delayed a year and ultimately chose a computer-science PhD, retaining power; Graham predicts VC-installed executives will increasingly be COOs, not CEOs.

Despite this shift, undergraduates remain conservative, still following parents' "get a good job" advice, which fit an earlier era. Graham applies investment logic to careers: risk and reward are proportionate, and the young, with the most time to recover, should take the biggest risks. Roughly 9 of 10 startups fail, but the one that succeeds can pay more than 10 times an ordinary salary — and failing at 22 costs little. To test whether a failed startup hurts job prospects, Graham asked engineering managers at Yahoo, Google, Amazon, Cisco, and Microsoft to compare two equally able 24-year-olds — one with a failed startup, one with two years at a big company — and all preferred the failed founder; Yahoo's Zod Nazem said he'd value the failed-startup candidate more.

What young founders actually lack isn't technical skill or financial discipline but the realization that you must build something people want (illustrated by PowerPoint's real function being to relieve public-speaking anxiety, not present ideas) — a shift Graham thinks can be taught quickly by simply telling hackers to go find users. Grad school works as a launchpad, since it clusters smart people with free time (David Filo and Jerry Yang built Yahoo's directory in grad school for over a year before dropping out). Graham closes half-seriously: the "responsible" advice is to finish college and work for a few years first — advice he admits nineteen-year-old Bill Gates would have ignored, correctly.

Startup ideas only seem hard to generate because people overvalue them — they assume a good idea is a "million dollar idea," handed down complete like a blueprint, so they never bother trying. But there's no market for startup ideas; you can't sell one, which proves they're not intrinsically valuable. Most startups end up nothing like their initial idea; the idea's real use is as a starting question ("could one build a collaborative, web-based spreadsheet?") rather than an assertion, since questions don't invite the objections assertions do, and a wrong-but-productive question — even a partial solution — is fine as long as it leads somewhere.

Generating such questions requires exposure to new technology plus the right friends, which is why startups cluster around universities: research forces novelty, and fellow students are elastic-minded. Graham advises staying "upwind" — favoring work that maximizes future options over well-paid but closed-off "downwind" jobs like coding Java at a bank. Co-founders matter because ideas develop through the resistance of explaining them to someone else (even Einstein needed this); Y Combinator won't fund single-founder startups. Graham speculates this explains why only 1.7% of VC-backed startups are founded by women: best friends tend to be same-sex, so founder pairs drawn from a minority group become a minority squared.

Ideas arise like doodles — undirected products of "habits of mind" built from sustained work in a field, which is why novices can't generate ideas before they can even judge them. Importing habits from unrelated fields yields more novelty than obvious pairings (computer science and electrical engineering); Graham finds math and painting especially good sources of metaphor, since hackers and painters are both "makers." Randomly combining concepts rarely works — you need a real problem to anchor the wandering. People are in denial about problems because problems are irritating; Graham's 2002 spam-filter breakthrough succeeded not because the statistical approach was hard, but because no one had seriously tried it (Bill Yerazunis reached similar results independently via a general-purpose file classifier).

Since companies must make money, and wealth is simply what people want, startup ideas should aim at things people want, even though "good" and "valuable" ideas aren't identical (new theorems vs. celebrity gossip magazines). Fixing something broken — dating sites, or the deeper problem behind Windows, which can't be attacked frontally since a monopoly will just copy a direct competitor rather than buy it — is one route; Graham bets a small startup, not Google, will eventually "transcend" Windows by redefining the problem. Turning a luxury into a commodity is another, as Henry Ford did with cars and Roman water mills did with mechanical power; redefining the problem, à la Michael Rabin's near-certainly-prime numbers, often unlocks a cheaper version (the Model T only came in black). Making things easier to use is undervalued — Graham cites a stove that displayed only "Err" instead of physical knobs — because simplicity takes real design effort.

Because success now effectively means getting acquired rather than going public, founders should openly treat their work as product development on spec for larger companies, but must target something with multiple potential acquirers, not just one (don't fix Windows — only Microsoft would buy it). Finally, the most productive path to ideas is accidental: Lotus, Apple, and Yahoo all began as side projects (Mitch Kapor's program for a friend, Steve Wozniak's hobby machine Hewlett-Packard wouldn't sponsor, David Filo's personal link collection) — so the best way to find a "million dollar idea" is to build fun things with friends.

Startups fail less often from moving too fast than from moving too slow, and the founders Y Combinator funds keep needing the same counterintuitive points repeated. First: release early. Ship a minimal but non-buggy version 1, since users tolerate thinness but not bugs, and you can't guess what a userbase you don't yet know will want -- Reddit had half a million unique visitors its founders couldn't describe. Wufoo released its form-builder before the underlying database; 83,000 people showed up anyway, and Linux users' complaints about Flash led to a rewrite that a later, bundled release would have missed. An early release also doubles as a shakedown cruise for fatal problems (a bad idea, feuding founders) and forces harder work, since bugs in a live product feel urgent in a way bugs in an unreleased one don't. Second, and inseparable from the first: keep pumping out features -- "feature" meaning any unit of hacking that improves users' lives, not added complexity. Improvement compounds like exercise, and it doubles as marketing, since users expect visible progress (Gmail, per Paul Buchheit, showed how much better webmail could be, once someone bothered). If a product feels finished, assume that's a failure of imagination, not evidence of completion.

Third: make users happy, since a startup can force no one to use it or deal with it. Design for the median visitor, finger poised on the Back button -- state in one or two sentences what the site does (contrast the corporate site describing "enterprise content management solutions... to minimize business risk"), and front-load the best material rather than hiding it, since showing beats telling. Growth rate is the honest scorecard. Fourth: fear the right things. Most disasters -- a co-founder quitting, a patent surprise, a failed deal -- are survivable. Big incumbents like Google aren't the real threat: they're no smarter, less motivated since one product's failure won't sink them, and slowed by bureaucracy. The real danger is startups you don't know exist yet. And competitors generally aren't even the top killer -- internal disputes, inertia, and above all ignoring users are what hose most startups; Microsoft and Apple both survived by abandoning their original plans (selling languages, selling circuit boards) once customers redirected them, illustrating Feynman's point that nature's imagination exceeds any founder's a priori idea.

Fifth: determination matters more than intelligence, which is why VCs mistakenly chase eminent professors when they should back students -- Microsoft, Yahoo, and Google were all founded by dropouts. Losing intelligence won't kill a startup; losing commitment will, since investors and acquirers judge by perceived commitment (driven by fear of missing out, not expected returns) and treat you differently once they sense you're not going anywhere. The needed quality is determined-but-flexible, like a running back, not stubborn. Sixth: there is always room for more startups, since the ceiling isn't how many companies Google or Yahoo can acquire but the total wealth that can still be created -- effectively unlimited.

Seventh: don't get your hopes up. Treat optimism like a reactor core -- fine to apply to your own effort, dangerous applied to other people or deals. Assume any pending investment, acquisition, or big customer deal will fall through, both to avoid disappointment and, more practically, to keep hunting rather than lean on something that collapses; keeping other options open also yields better terms once the first deal is finally struck. Finally, Graham reframes the whole stressful enterprise: a startup isn't primarily a way to get rich but a way to compress the boring necessity of making a living into the shortest possible time, given how short life is -- speed, not money, is what makes the ordeal worthwhile.

Y Combinator's official advice to undergraduates is to wait until after graduation to start a startup — not because anything biological changes then, but because something social does: once out of school, everyone (family, friends, yourself) stops crediting you with a "student" identity and starts judging you as having one occupation, and the opinion of one's peers is the most powerful motivator there is, more powerful than the desire to get rich. A startup begun the summer before senior year reads as a summer job with an escape hatch; the same startup a year later reads as your life, and failing at it reads as failure. Sam Altman, who co-founded Loopt just after his sophomore year and became one of YC's most promising founders, is the exception — YC's actual policy is to fund only undergrads it can't talk out of it, the same trick a violinist uses in telling every hopeful they lack talent: determined people ignore the warning, uncertain ones heed it, and the advice is right either way. The sweet spot, based on who YC is most excited to fund, is the mid-twenties.

Young founders have five real advantages, none of them raw talent. Stamina: every successful startup Graham knows of ran on long hours, and young founders especially need them because they're less efficient than they'll later become. Poverty: since most startups fail by running out of money, a low burn rate — easy without a mortgage — buys time to recover from the mistakes startups inevitably make; Graham and Robert Morris, 29 and 30 when they started Viaweb, deliberately priced it low ($300/month, an order of magnitude below the norm) because they thought cheaply — the same instinct behind the cheap Apple II, built around cassette-tape storage and a TV monitor because Wozniak was broke. Rootlessness: young people move easily, and startups constantly require moving to a hub — per-capita startup output in Houston, Chicago, or Miami is negligible next to the Bay Area, and Graham doubts geography-agnostic exceptions like 37signals will become the rule. Colleagues: co-founders are the scarcest resource in startups, and school offers the highest concentration of smart, ambitious peers you'll ever have — Google, Yahoo, and Microsoft were all founded by people who met in school. Ignorance: founders interviewed for Jessica Livingston's book of interviews with startup founders repeatedly said that if they'd known in advance how hard it would be, they'd have been too intimidated to start; Steve Wozniak credited Apple's best work to "not having money and not having done it before, ever."

Young founders' disadvantage is that they build "class projects" instead of real startups, missing two things: an iteratively-discovered real problem (professors can't assign real problems, and class timelines don't allow an idea to evolve through building, the way startups do), and intensity (professors grade effort — distance traveled from the start; the market grades only distance remaining to what users need, indifferent to how hard you worked). What "work experience" actually supplies is elimination of the childhood "flake reflex," plus a visceral understanding that money comes only from doing what other people want — which is why 24-year-olds who grasp that getting rich means escaping "the brutal equation" of trading time for survival outwork 20-year-olds who think it just buys Ferraris.

Practically, Graham tells students to learn startups by working at one (sneaking in informally, since business-school classes and most books on the subject are worthless — successful founders don't need to write books), to pick co-founders by genuine admiration rather than convenience, since a startup is a stress test for friendship, to learn languages founders actually choose (Ruby, Python) rather than ones employers want (Java, C++), and above all to build something with real users while still in school — since college teaches programming the way one might teach grammar without ever mentioning that writing exists to communicate to an audience.

There is really only one mistake that kills startups: not making something users want. Everything else is a way of failing to do that, and Graham catalogs eighteen specific versions.

The first cluster concerns founders and idea choice. A single founder is a bad sign — Oracle and other "solo" successes usually had more than one, and one person alone lacks colleagues to brainstorm with, talk them out of bad calls, or sustain them through the low points. Location matters enormously: startup density falls off sharply outside Silicon Valley, then Boston, Seattle, Austin, Denver, and New York — in Houston, Chicago, or Detroit it's too small to measure, because that's where the experts, standards, and chance encounters are. Choosing a marginal niche to dodge competition is the startup equivalent of a child flinching from a fly ball — avoiding competition means avoiding good ideas. Derivative ideas (copying Facebook with a tweak) are weaker than ideas drawn from a problem the founders personally have, as with Wozniak and Apple, Larry and Sergey at Google, or Sabeer Bhatia and Jack Smith at Hotmail. And obstinacy — clinging to the original plan — kills, since most successful startups end up doing something quite different than intended; the test for whether a pivot is healthy is whether each new idea reuses what came before, and users are the best guide.

A second cluster is execution. Hiring bad programmers killed many 1990s e-commerce startups run by "business guys" who couldn't judge programming talent. Choosing the wrong platform is related: Bubble-era startups built on Windows often died, as FreeBSD-based Hotmail would have; PayPal nearly switched to Windows post-merger until Max Levchin showed it scaled at only 1% of Unix performance, and Java applets destroyed nearly everyone who bet on them. Slowness in launching is dangerous because software is "always 85% done," and only shipping forces real completion and real user feedback — though launching too early can burn your reputation, so the fix is to launch the smallest useful, expandable core. Having no specific user in mind — building for vague "teenagers" or "business users" rather than someone concrete — means flying blind.

A third cluster concerns money. Raising too little leaves you without runway to reach the next milestone (prototype, launch, growth), so Graham advises spending little and keeping the initial bar low. Spending too much, now rarer, is classically caused by hiring too many people, which both raises burn and (per Fred Brooks) slows you down. Raising too much brings its own cost: VCs expect the money "to go to work," pushing you toward offices, hires, and office politics, and makes it harder to change direction once a team is built around one plan; VC-scale raises are also a huge time sink, so Graham advises taking the first reasonable offer rather than shopping around. Poor investor management — letting investors run the company, or provoking needless fights — is dangerous in both directions; even Apple's board nearly destroyed the company by firing Steve Jobs.

The last cluster: sacrificing users for a business model too early is a mistake, because making something people want is far harder than monetizing it — Google built search first and worried about revenue later. Not wanting to get your hands dirty — avoiding sales because it's unpleasant — is fatal, since acquirers pay for users, not cleverness alone. Fights between founders hit about 20% of Y Combinator startups, though a founder leaving isn't always fatal (Blogger survived down to one person); most disputes trace to suppressed misgivings at founding. Finally, a half-hearted effort — not quitting the day job — is the most common failure of all, quieter than flameouts: most founders of failed startups keep their day jobs, most founders of successful ones don't, and Graham suspects many who could have succeeded never tried hard enough to find out.

Most people who would benefit from starting a startup are held back not by real obstacles but by a jumble of reasons, justified and bogus alike, that only become clear once separated out — Graham lists all sixteen.

The case for trying: of Y Combinator's first batch (summer 2005, eight startups), at least four succeeded — Reddit (a merger with Infogami), Loopt, and one unnamed company were acquired; three others died, but even Kiko, crushed by Google Calendar after a year, sold its software on eBay for $250,000, leaving founders a year's salary each before launching Justin.tv. Zero percent of that batch had a genuinely bad experience. Graham won't claim a lasting 50% success rate but thinks 25% achievable, against an oft-cited (probably invented) 10% baseline.

The sixteen reasons and verdicts: (1) Too young — real for some (world median age is 27), but the cutoff is adulthood, not years: the test is whether you still flake ("I'm just a kid") or meet a challenge with submission rather than "Why do you think so?" Sam Altman, funded at 19, passed easily. (2) Too inexperienced — he used to require age 23 plus prior work; he now thinks starting young cures inexperience fastest, since a job tames you into needing a boss, whereas the Kiko founders grew more in one failed year than Microsoft or Google would have produced. (3) Not determined enough — the single best predictor of success, and hackers underestimate their own; the tell is whether you already drive your own side projects, as Page and Brin did. (4) Not smart enough — irrelevant for most startups, which run on effort not brains; worrying about it is evidence you're smart enough. (5) Know nothing about business — should carry zero weight; build what people want first, monetize later, and acquirers buy for strategic value, not revenue. (6) No cofounder — a real problem, since every investor prefers funded teams; the fix is to find one, even by moving or switching ideas — easiest while still in school. (7) No idea — not fatal, since ~70% of a YC startup's idea changes within three months; YC was about to accept "we have no idea" applicants, since what matters is what you've built, not the pitch; the recipe is fixing a genuine need in your own life, as Wozniak did by building himself a computer. (8) No room for more startups — a fallacy: there's no cap on salaried jobs at big firms, so none on equity slots at small ones, since demand for material wealth (health care alone is a "black hole") is unbounded. (9) Family to support — genuinely real; Graham won't advise parents to do it, though he will 22-year-olds; alternatives are a consulting business pivoted into product, or joining a startup as roughly a "1/n² founder." (10) Independently wealthy — his own excuse, though the itch returns since people worth working with still need income. (11) Not ready for commitment — legitimate given the three-to-four-year ask, but a regular job takes similar time with less freedom. (12) Need for structure — real for some, but good startups have near-zero hierarchy below twelve people, like a soccer team of skilled players who rarely need to talk. (13) Fear of uncertainty — self-resolving, since most startups fail anyway, and failure carries no stigma with employers or investors here. (14) Don't realize what you're avoiding — people without a real job (summer internships don't count) underestimate how demoralizing fixed hours and boring work are. (15) Parents want a safe career — they're rationally more risk-averse for kids than themselves, and "fight the last war," valuing careers like medicine as they were decades earlier. (16) A job is the default — only a ~100-year-old norm succeeding farming; startups may mark a shift as large as farming's replacement by manufacturing, no less frightening than a serf fleeing to the city.

For a startup, merely not dying is close to sufficient for success, since success means the founders get rich and staying alive long enough tends to produce that. Y Combinator expects a third to half of funded startups to succeed by this measure; Viaweb, Graham's own company, nearly died even while being acquired, once having to talk down a wavering investor in a borrowed Yahoo conference room. Across five YC batches, about ten funded startups have died, almost always quietly. The most reliable predictor of death is silence: a startup that stops emailing or showing up is, 100% of the time, dying, while one that stays in regular contact — weekly dinners, other YC founders nearby in San Francisco — tends to survive, because the contact itself creates pressure to keep producing.

Official causes of death (running out of money, a co-founder quitting) mask the real cause: demoralization. Startups rarely die "in mid keystroke," so persistence beats any single fix. Feeling like nothing is working is normal, since first launches usually fail; the answer is to iterate, per Paul Buchheit's advice to find even a small group of users who love the product and expand from that core — the path Blogger and Delicious both took over years.

Distraction is fatal: any sentence ending "but we're going to keep working on the startup" (grad school, consulting, moving away) is quitting in disguise, offering a face-saving excuse to fail. Founders are driven more by fear of public humiliation than by hope of riches — shown by an Octopart founder who, after dropping out of grad school and appearing in Newsweek as a "billionaire," became unable to quit without shame; Graham estimates that treatment would push success rates to 90%. Disaster is near-certain (roughly 1000-to-1 odds against avoiding one), so the instruction is simply: don't give up.

Startups are now undergoing the same historical transformation that any technology undergoes once it becomes cheap to produce — the pattern seen with computers, steel in the 1850s, power in the 1780s, and thirteenth-century cloth manufacture — and this cheapening will reshape startups in ten predictable ways.

First, there will be far more startups, since the only remaining barrier is courage rather than investor permission. Second, cheap high-volume production forces standardization: YC, though still just four people, has standardized its investment paperwork and commissioned generic angel agreements, though Series A rounds above a million dollars will stay custom deals. Third, acquisitions will become standardized and routine, following Google's lead — Google buys far more startups than it announces, partly because Larry and Sergey once pitched search engines to indifferent buyers themselves and know the quality of people available this way; Graham predicts companies will need a "chief acquisition officer" merging the currently separate roles of spotting and negotiating deals, since at present no one is penalized for buying at $200 million something that could have been bought for $20 million earlier.

Fourth, because founders can start companies younger and repeatedly, they can adopt investors' portfolio logic and take riskier, higher-return strategies. Fifth, founders will skew younger and more technical, since cheap seed funding — a few tens of thousands of dollars from YC or "your uncle" — lets hackers prove a product with data (pointing investors to Alexa rankings) rather than "selling" a business plan to investors who are often poor judges. Sixth, startup hubs like Silicon Valley will matter more, not less: succeeding still depends on face-to-face contact even though starting no longer does; seed funding is a national, possibly international, business, and mobile two-person startups relocate easily, so local "clone" hubs fail because top talent keeps migrating to the real hub ("Atlanta is just as hosed as Munich"). Seventh, investors and acquirers must get better at judging as volume rises — YC already receives about 1,000 applications a year and must prepare for 10,000. Eighth, college will change: degrees and school prestige will matter less since startups are judged by users, not credentials, and students may prioritize meeting future cofounders over chasing grades or internships. Ninth, more competitors won't erase the advantage of cheapness, since wealth creation isn't zero-sum, but good ideas can no longer be sat on. Tenth, faster iteration follows, illustrated by a founder who measured his team's output as one-thirteenth as productive after being acquired by a big company — the "cost of bigness."

Graham closes with a plumbing metaphor: financing and hiring today are a leaky maze of narrow, twisty pipes insulating people from real performance, producing a chain where high schoolers chase grades to enter elite colleges, college students chase grades to impress employers, and employees then waste time on office politics. As this consolidates into one big, straight pipe, the force of being measured by actual performance will propagate back through the system, flushing out the arbitrary measures — which he calls the future of web startups.

Startups succeed by behaving like nonprofits: making something people want while treating profit as a secondary problem to solve later, since revenue is far easier to engineer than a great product. Craigslist proves the model at the extreme—wildly popular with a startlingly small staff, "upwind" of revenue it doesn't bother collecting, the way Patrick O'Brian's naval captains hold the weather gauge over an opponent. Google looked identical to a nonprofit for its first ad-free year. Graham recalls shelving his own idea for a spam-free webmail service, convinced it had to be a company to survive, not a grant-funded project—and notes whoever built it would now be rich, since spam filtering was terrible everywhere else during a two-year window.

How far does benevolence extend as a strategy? Possibly even to people without money: Y Combinator profits by helping founders who arrive broke, and Graham speculates an organization curing malaria could similarly ride a country's resulting growth. Microsoft, in its early years breaking IBM's hardware monopoly and crashing PC prices, acted like Robin Hood and its stock rose like Google's; grown mean and stagnant since, its stock has gone flat. Small companies must charm users; big ones can bully them, until conditions shift and users flee—making Google's "Don't be evil" (credited to Paul Buchheit) a hedge against that fate, one rivals like tobacco or record companies can't credibly copy.

Benevolence also works internally, in three ways. It sustains morale: a startup's belief it's helping people keeps founders working through the "roller-coaster" lows that otherwise become self-fulfilling failure—as when Blogger's Evan Williams, left alone after the company ran out of money, kept going because thousands of users depended on him. It attracts help: investors leaned in once Chatterous's founders stopped needing their money, and idealistic top hackers, who can work anywhere, gravitate toward benevolent missions—illustrated by Octopart, fighting distributor Digi-Key to keep component pricing transparent. And it acts as a decision compass, stateless like truth-telling: "do whatever's best for your users" is the rule Y Combinator applies uniformly across the 57 (of 80) funded startups still alive, since no one can track ulterior motives at that scale. Graham closes by disclaiming personal goodness—he's offering benevolence as effective strategy and design spec, not moral exhortation.

Startup success comes down to thirteen compact rules. Pick good cofounders: founders are to a startup what location is to real estate—nearly impossible to change once set—and a startup's success is almost always a function of them. Launch fast, because real work only starts once users engage; then let the idea evolve through iteration rather than treating the first concept as fixed. Understand your users, since wealth equals number of users times how much you improve their lives, and satisfying a few users completely beats satisfying many partially—it's harder to fool yourself about depth than about headcount. Offer customer service so good it doesn't scale, since that's how you learn about users. You make what you measure: tracking a number, like Joe Kraus found, tends to push it up. Spend little, since running out of money is the top cause of death; get "ramen profitable" (covering founders' living costs) to change your leverage with investors. Avoid distractions, especially paying ones like consulting or fundraising. Don't get demoralized, and don't give up—effort carries startups further than fields like mathematics. Expect deals to fall through; ignore them until they close. Forced to keep one rule, Graham picks understanding users, since it underlies launching early, iterating, and sustaining morale.

The essential quality of a good startup founder is being relentlessly resourceful. Graham arrives at this by first defining its opposite: hapless, wrongly glossed by dictionaries as merely "unlucky," when it actually means passivity — letting circumstances batter you rather than bending the world to your will. Since English has no antonym for hapless, and metaphors like "be a good running back" (determined but adaptive) don't travel outside the US, he needed a direct phrase.

Relentless alone isn't enough, because in any interesting domain the obstacles are novel — you don't know in advance whether you're plowing through foam or granite — so you must also be resourceful, continually trying new approaches. This differs from the recipe for success in writing or painting, where obstacles are internal (your own obtuseness) and the operative quality is active curiosity instead.

Graham reports that after four years of trying, this trait can be taught to many people, especially the young, though some are constitutionally passive. Because there's no economic ceiling on how many startups could exist — demand for new wealth is as unbounded as the supply of provable theorems — the real limiting factor is the pool of people who are relentlessly resourceful, making it a practical test for founders and cofounders alike.

Ramen profitable means a startup earns just enough to cover the founders' living expenses — a newly feasible threshold that matters because it frees a company from needing investors, buying it time rather than proving it has succeeded. This differs from traditional profitability, where a hardware company might spend $50 million over five years before turning a profit at $50 million in annual revenue. A ramen-profitable startup might hit $3,000 a month after two months, sustainable only because its founders are young and can live cheaply — feasible for software, not for capital-intensive fields like biotech.

Beyond avoiding the obvious risk of investors exploiting a cash-strapped founder, three subtler benefits follow. It makes a company more attractive to investors by demonstrating people will pay, the founders solved the right problem, and expenses stayed disciplined — addressing the three classic causes of startup failure. It boosts morale: once revenue covers living costs, survival becomes the default rather than death. And least obviously but most importantly, it removes the need to interrupt building the company to fundraise — Graham notes raising money cut Y Combinator's own productivity to a third of normal for months, since fundraising monopolizes attention.

Ramen profitability isn't bootstrapping (few startups succeed without eventually taking investment) nor Joe Kraus's rule of monetizing from day one — Facebook didn't, and thrived. The money needn't come from the eventual business model; Google initially profited by licensing search to Yahoo. The main danger is drifting into consulting, which can't scale like a product company; some early consulting-type work is fine, but ramen profitability is only a way of not dying en route to growing big.

Graham asked YC founders what had surprised them about running a startup — where his own advice fell short — and over 100 responses clustered around the same patterns, capped by one explanation. Most-cited: choose cofounders for character and commitment, not ability, since the relationship becomes as intense as a marriage (several used that word) and must be built on trust from the start. Second, a startup takes over your life — time seems to stretch, and immersion stops feeling like "work" only because no boundary remains between work and everything else. Third, founders face an emotional roller-coaster, swinging between fantasies of buying islands and rehearsing failure within hours; the energy to keep going, one noted, "is siphoned from the founders themselves." Fourth, the highs compensate — many said startups were more fun than any job — though Graham downplays this, preferring surprise-by-fun to disillusionment-by-grimness.

Persistence mattered more than expected, and more than raw intelligence; persistent founders found even uncontrollable-seeming obstacles (immigration was cited) resolving themselves. Everything also takes longer than expected — 2-3x by one estimate, successful startups typically taking 3-5+ years — from friction wherever a startup touches a big company or VC fund, and because founders wrongly expect to be the rare instant hit like YouTube (true for maybe 1 in 100). Success comes from grinding through many small things, not one brilliant insight or "killer feature," so spreading effort is necessity, not caution, since you don't know which basket is best. Launch the minimal version fast — pride, not practicality, drives over-building — and treat product development as a conversation only truly begun after launch; the first version is a device for getting users talking, not a finished product. That demands willingness to change the idea itself: determination without flexibility is "a greedy algorithm" stranding you at a mediocre local maximum, whereas fast iteration was the real key to success. Competitor fears are almost always false alarms — execution matters more than ideas, and hype-driven competitors without real users vanish fast.

Acquiring users proved unexpectedly hard, especially where a product requires other companies' developers to integrate it; one founder criticized YC for treating "make something people want" as a purely engineering task that underweights customer-acquisition cost. Deals fall through constantly since startups have little leverage, so the advice is to assume no investment is coming and take money whenever offered. VCs are often startlingly clueless — one couldn't switch on a hardware device they'd funded — while angels, usually ex-founders, fare better; VCs need only seem confident enough to persuade the limited partners supplying their capital. Because investors judge so poorly, founders play games: "feigning certitude" impresses them, in a chain where VCs sell confidence to LPs and founders sell it to VCs. Luck consequently looms larger than hackers, trained to believe skill determines outcomes, expect; outcome is modeled as skill times determination times luck, so a zero on luck zeroes the result — and founders in the middle, neither quick failures nor successes, see this most clearly.

Founders were also surprised by the value of community — the tight peer group of fellow YC founders, and Silicon Valley's broader culture of accessible, often disinterested help, evidence startup wealth creation isn't zero-sum. Outside that world, though, founders get little social respect, even in dating, since most good startup ideas look like bad ideas to outsiders (95%, by one estimate, peg a pitched idea as a flop). Roles also keep changing with growth: founders code less and manage more every 6-12 months, learning an "employee equation" distinct from the founder one, though stress drops sharply — one estimated 75% — once cruising altitude is reached. Graham's closing point ties everything together: every surprise is something he'd already told founders, yet still surprises, because everyone's baseline model of work is a job, and a startup differs from a job far more than people believe until they've lived it.

Y Combinator has settled on five qualities, not raw intelligence, as what actually predicts founder success. Determination matters most: startups hit constant obstacles, and easily-demoralized founders fail, exemplified by Bill Clerico and Rich Aberman of WePay, who out-persist bureaucratic companies in deal negotiations. This determination must pair with flexibility, since startups are too unpredictable for rigid dreams — like a running back who may need to cut sideways to advance; Daniel Gross of Greplin embodied this by abandoning a bad ecommerce pitch and cycling through two more ideas within days of Demo Day. Where intelligence does matter, it's imagination over problem-solving speed: good startup ideas usually look bad at first, as with Airbnb, which Y Combinator funded despite doubts, swayed mainly by the founders (who'd financed themselves selling Obama- and McCain-branded cereal). A fourth trait, naughtiness — a "piratical gleam," not dishonesty — shows up in Sam Altman of Loopt's suggested interview question about times applicants "hacked" a system to their advantage. Finally, since most successful startups have two or three founders, not one, the founders' friendship must be genuinely strong enough to survive strain, as with childhood friends Emmett Shear and Justin Kan of Justin.tv.

Great startup ideas sit unexploited because founders suffer "schlep blindness" — an unconscious aversion to tedious, unpleasant work that keeps them from even perceiving ideas requiring it. Hackers especially want to build a company by writing clever code alone, but schleps (a Yiddish-derived term for grinding tasks) are inevitable and, Graham argues, largely constitute what business is. Founders should treat schleps like a cold pool: jump in rather than seek them out.

Stripe is the prime example: for over a decade every hacker who processed payments online knew the pain, yet applicants kept pitching recipe sites and event aggregators instead of fixing payments, because dealing with banks, fraud, and regulation felt too intimidating to even consider. This scariness makes ambitious ideas more valuable — like undervalued stocks, since fewer founders compete for them.

Ignorance is the best antidote: founders who knew the obstacles upfront might never start, which may explain why the most successful startups often have young founders — they misjudge both how much they'll need to grow and how much they can, and the errors cancel out. Older founders only make the first mistake. Where ideas' schleps are too obvious to miss, Graham's fix is to stop asking "what problem should I solve?" and instead ask "what problem do I wish someone else would solve for me?" — the question that would have surfaced Stripe. He closes: it's too late to be Stripe, but plenty remains broken for those who can see it.

The biggest startup ideas provoke instinctive fear rather than excitement, and that repulsion — not just the work involved — is why most people never think of them: their scale threatens the founder's sense of identity, so the subconscious filters them out, leaving them invisible except to those willing to approach obliquely.

Graham illustrates with seven ideas. (1) A new search engine: Microsoft's panicked move into search showed it had lost its way, and now Google itself shows cracks — cluttered, unpredictable results — so a search engine built for the top 10,000 hackers, unapologetically hackerish (e.g., strong code search), could become dominant the way Facebook's Harvard-only launch eventually did. (2) Replace email: email is really a bad todo-list protocol; a replacement should give recipients more control over what lands on their list and, crucially, be fast — Gmail has grown "painfully slow," and Graham says he'd pay $1000/month for something better given how many hours he spends in his inbox. (3) Replace universities: not eliminate them, but strip their monopoly on learning; credentialing may separate from education entirely (Y Combinator already resembles a piece of this), though high schools are a harder target due to bureaucracy. (4) Internet drama: the Internet has beaten cable as a delivery mechanism (Graham's family replaced their TV with a wall-mounted iMac); scripted drama will persist alongside social apps and games, needing either an "app store" gatekeeper like Netflix or Apple, or independent streaming/payment infrastructure. (5) The next Steve Jobs: a well-connected source confirmed post-Jobs Apple won't produce major new products; since product visionaries emerge only by founding, not by being hired, the next hardware giant must be a startup — helped by Jobs's example proving it possible, as Roger Bannister's four-minute mile emboldened later runners (Jim Ryun ran 3:59 as a high schooler ten years later). (6) Bring back Moore's Law: since about 2002, density gains have gone to more cores rather than faster clock speeds, forcing painful manual parallelization; a "sufficiently smart compiler" that auto-parallelizes code, sold as a web service, could recapture the value Intel once delivered via clock speed — short of full automation, options include parallel-programming toolkits like Hadoop and MapReduce, or a human-assisted optimization marketplace. (7) Ongoing diagnosis: continuous, symptom-free monitoring will replace reactive diagnosis — illustrated by Bill Clinton in 2004, who learned his arteries were over 90% blocked only once he felt short of breath. Resistance comes from doctors wary of "incidentalomas" (a friend's brain scan falsely flagged what turned out to be a harmless cyst, costing her days of terror), but Graham predicts constant scanning will eventually make such findings as unremarkable as knowing one's weight.

Tactically, Graham advises never announcing the big goal directly — don't say you're replacing email, call it "todo-list software" — and to start deceptively small, as Bill Gates did with a Basic interpreter and Mark Zuckerberg did with a Harvard-only site, since neither knew at first how big the outcome would become. Aim in a general direction, like Columbus heading west, rather than plotting a precise future course, since any detailed blueprint of the future is almost certainly wrong.

A startup is a company designed to grow fast — the only defining trait; funding, technology, age, and "exit" are consequences of growth, not the thing itself. Founders should treat growth as a compass for nearly every decision.

Most new companies aren't startups: a barbershop satisfies plenty of demand (a) but can't reach a wide market (b), since customers won't travel far. Software often solves reach but can still fail on demand-size — software teaching Tibetan to Hungarians reaches everyone who wants it, but there's no market; software teaching English to Chinese speakers is a real startup idea. Markets are efficient, so ideas satisfying both (a) and (b) attract ferocious competition, while niche constraints (a bar tied to one neighborhood) both limit and protect ordinary businesses. Startup ideas thus tend to come from founders who, being unusually different, see problems others miss — often unconsciously at first. Steve Wozniak wanted his own computer in 1975; chip advances soon made that mainstream, yielding Apple. Larry Page and Sergey Brin wanted better search; as the web grew, their niche problem became everyone's, and Google was already entrenched by the time others noticed. Two links to technology follow: change uncovers newly soluble or newly large problems, and startups create new technology (broadly defined) themselves.

"Startup" is a pole, not a threshold — like "actor," it starts as a declaration of ambition. Growth follows an S-curve: a flat period figuring things out, a rapid-growth phase defining the company, then slowing as it matures and hits market limits. The number every founder should track is growth rate as a *ratio* — a constant number of new customers each month actually signals a declining rate. YC tracks it weekly: 5–7% is good, 10% exceptional, 1% means the founders haven't found what they're doing. Revenue is the best metric; active users the best proxy pre-revenue.

YC's method: pick a weekly growth target and try to hit it, reducing starting a startup to a single optimization problem, like optimizing code around one variable. Missing the target motivates actions (like hiring) that won't help this week's number but prevent future misses. This rarely traps founders on a local maximum, since inaction is the real risk and good ideas cluster near better ones; optimizing for growth can even evolve a company into a better idea than the one it started with.

A table converting weekly to yearly growth rates makes the stakes concrete: 1%/week compounds to 1.7x/year, 2% to 2.8x, 5% to 12.6x, 7% to 33.7x, 10% to 142x. A company making $1,000/month at 1%/week reaches only $7,900/month after four years (less than a good programmer's salary), while at 5%/week it reaches $25 million/month. This variance explains why startups behave unlike ordinary companies and why they fail so often: if success is worth $100 million, even a 1% chance yields $1 million in expected value, and for the most capable founders that probability can run 20–50%. Most "startups" are therefore doomed, but Graham isn't bothered — critics alarmed by this are wrongly judging by the median rather than the average in a high-variance domain.

Growth also explains investor and acquirer behavior. VCs favor high-growth companies not just for returns but because capital gains are easy to oversee — founders can't enrich themselves without enriching investors too, unlike the profit-skimming risk in dividend-based investing. Founders take VC money even when profitable because it lets them *choose* a faster growth rate, insurance against being outgrown, especially with network effects. Successful startups also draw acquisition offers, since rapid growth is both valuable and threatening to incumbents (eBay's purchase of PayPal, later ~43% of eBay's sales) — acquirers pay a premium from fear of what a competitor could do with the company. Understanding growth is what starting a startup consists of: a founder is an economic research scientist searching for a rare idea that generates rapid growth, most finding nothing remarkable, a few discovering the equivalent of relativity.

Startup ideas should not be invented; they should be noticed, as problems you have yourself. The best ideas share three traits: founders want them, can build them, and few others see their value (Microsoft, Google, Facebook). Working on your own problem guarantees the problem is real; the most common startup failure is solving problems nobody has. Graham made this mistake in 1995, spending six months on an online art-gallery startup because galleries didn't actually want to be online. Y Combinator calls plausible-but-invented ideas "sitcom" ideas: a social network for pet owners sounds reasonable but draws zero real users, since no one urgently needs it.

A viable startup needs a few users who want it desperately, not many who want it a little — anything broader a v1 could serve would probably already exist. Demand can be pictured as a hole: Google is a vast crater of dependent users; a startup instead needs a "well" — narrow but deep. Microsoft's Altair Basic served a couple thousand machine-language programmers who badly needed it; Facebook began with only Harvard's few thousand students. Narrowness is a byproduct of depth, not the goal. Depth alone isn't sufficient, though — the idea needs a path outward. Facebook worked because colleges are similar enough that a Harvard product spread college to college, then opened to everyone; Microsoft's path ran from Basic to other machines, languages, and operating systems.

Predicting that path is usually impossible — Airbnb began as merely renting floor space during conventions. The real answer is becoming someone whose hunches are reliable, by living at a fast-changing field's leading edge — as a builder or heavy user, like Zuckerberg. Pirsig's "make yourself perfect and paint naturally," combined with Paul Buchheit's point that such people "live in the future," yields the formula: live in the future, build what's missing. Gates and Allen heard about the Altair; Drew Houston built Dropbox after forgetting a USB stick. The verb is "notice," not "think up." Anyone can reach a leading edge in about a year — programming is surest, since "software is eating the world." To notice ideas, turn off every filter except "what's missing," question the status quo, and note what chafes you; a good idea should feel obvious, as it did to Viaweb's founders about software-generated stores. Projects dismissed as "toys" — early microcomputers, BackRub, "the Facebook" as a stalking tool — often matter most. The refined rule: live in the future and build what seems interesting.

For students this favors building over studying "entrepreneurship"; clashing domains generates ideas (a CS major should study genetics, not business), undergrad side-projects beat PhD research, and Microsoft and Facebook both launched during Harvard's Reading Period. Feeling "late" on a good idea is normal — ten minutes of searching usually settles it, and startups are almost never killed by competitors. Something urgent that no rival offers is a beachhead; a crowded market is a good sign if you have a thesis on what incumbents overlook, as Google did against search engines afraid of their own logic's implications.

Absent an organic idea, deliberate recipes can substitute, used with discipline since most generated ideas are bad: stay within your expertise (a database expert's judgment on chat apps is worthless), recall moments you said "why doesn't someone build x," and note what's unusual about you, especially youth — only students could have built Facebook. Study others' unmet needs consultant-style — Rajat Suri waited tables to build restaurant software. Deliberately seek schleps and unsexy problems: the schlep filter, overcome by Stripe with payments, blocks more good ideas than the unsexy filter, overcome by Viaweb with ecommerce; the fear is usually overblown. Look at dying industries like journalism for what might replace them, and at markets incumbents disdain, as Hewlett-Packard dismissed Wozniak's TV-monitor computer. These recipes are plan B; the durable method is becoming, through time and exposure, someone to whom good ideas are simply obvious.

Startups don't take off by themselves; founders have to push them into motion, and the necessary push is usually a set of laborious, unscalable actions early on — like the hand crank that started car engines before electric starters.

The most common unscalable task is recruiting users manually rather than waiting for them to arrive. Stripe, despite solving an urgent problem, is famous within Y Combinator for aggressive early acquisition: the "Collison installation," where instead of emailing a beta link the founders said "give me your laptop" and set people up on the spot. Founders resist manual recruiting out of shyness/laziness and because early numbers look trivial — but they underestimate compound growth: 10% weekly growth turns 100 users into 14,000 in a year and 2 million in two years. Airbnb's founders went door to door in New York recruiting hosts and improving listings by hand; that roughly 30-day push was the difference between success and failure. Early startups are inherently fragile, and people wrongly judge them by the standards of mature companies (even Bill Gates briefly returned to Harvard after starting Microsoft). The right question isn't "is this taking over the world" but "how big could this get if the founders did the right things." Pinterest's Ben Silbermann noticed early users skewed toward design and deliberately recruited more at a design-blogger conference.

Founders should also take extraordinary measures to delight the users they get — Wufoo mailed hand-written thank-you notes to every new signup. Three things make this counterintuitive: engineering training emphasizes robust systems over hand-holding; founders fear over-attention "won't scale" (in fact it usually scales better than expected once it becomes cultural); and founders' own standards were set by big companies incapable of such service — Tim Cook can't send a hand-written note, but a small company can. Graham reframes Steve Jobs's "insanely great" for the earliest stage: it's not the (necessarily unfinished) product that should be insanely great, but the experience of being an early user, achieved through attentiveness. Direct engagement with the first users also supplies the best product feedback a company will ever get, before it's reduced to focus groups.

Sometimes the right unscalable move is deliberately narrowing the market to build critical mass, like containing a fire before adding logs — Facebook launched Harvard-only, then school-by-school, so students felt it was "their" site; other startups are the best early adopters for B2B products. Hardware startups can "pull a Meraki" (named for Meraki's founders) by assembling products by hand before a factory run — Pebble hand-built its first watches, learning things like "how valuable it is to source good screws," before selling $10 million worth on Kickstarter. B2B founders can act as a consultant to one user, tailoring the product to their exact needs, as long as they aren't paid hourly for it; Graham's own Viaweb built e-commerce stores by hand for reluctant merchants, which taught them what to build. Some startups can even fake automation by doing tasks manually behind the scenes — Stripe's "instant" merchant accounts were initially set up manually by the founders.

By contrast, the "Big Launch" — embargoed press, simultaneous coverage, an optimal Tuesday — rarely works, since long-term success depends on how happy the initial users were, not how many showed up; big-company partnerships similarly tend to disappoint. Graham concludes founders should treat a startup idea as a vector: what you'll build, plus the unscalable thing you'll do to get it going — and in the best cases, that early unscalable behavior (aggressiveness, hands-on manufacturing, obsessive attentiveness) becomes permanent company DNA rather than a phase to outgrow.

Preparing to start a startup requires distrusting much of your intuition and seeking not startup expertise but expertise in your future users and their problem. Startups are counterintuitive, like skiing: the instinct to lean back when scared makes things worse, and at Y Combinator the partners' real function, Graham says, was to tell founders things they'd ignore — founders make the predicted mistakes, then return a year later saying "I wish we'd listened." The one exception is people: if a potential cofounder, employee, investor, or acquirer gives you misgivings, trust that instinct rather than being swayed by how impressive they seem.

The second counterintuitive point is that startup knowledge itself barely matters. Zuckerberg succeeded not because he knew startup mechanics but because he understood his users; an undergrad fluent in convertible notes and class FF stock would worry Graham, not impress him. That fluency often signals "playing house" — imitating a startup's outward forms (a plausible idea, money raised at a good valuation, a cool office, employees hired) while skipping the one essential task, making something people want. Graham traces this to a lifetime of training in gaming systems: extracurriculars to get into college, and exam prep that, he admits of his own college years, meant guessing the 20-30 likely exam questions rather than mastering material. Founders arrive at YC still hunting for the trick, asking "How do we..." only to be told "Just..." Startups are where gaming the system stops working: there's no boss to trick, only users, and success depends purely on the product. Faking can fool investors for a round or two, but the company is doomed regardless; a founder with genuine user love raises money more easily than one who knows every fundraising trick but has flat usage.

Startups are also all-consuming, taking over a founder's life for years, possibly a decade, possibly the rest of their working life — Larry Page has been running flat-out since 25, and every YC founder of a big success reports the worry never lets up, only changes shape. Like having kids, starting a successful startup is an irrevocable button-push, so Graham's advice is: don't start one in college. It's "a brutally fast depth-first search" when most 20-year-olds should still search breadth-first, using their early 20s for the exploration — travel, whim-driven projects — that success forecloses (Zuckerberg, he notes, will never get to bum around a foreign country). There's no tradeoff in waiting, since you're more likely to succeed later anyway.

Whether you're suited to it can't be known in advance. After nine years predicting which YC founders would succeed, Graham found he could judge intelligence easily but toughness and ambition barely at all — founders' initial confidence (or lack of it) showed little correlation with outcomes, much as swaggering military recruits are no tougher than quiet ones. Only actually trying reveals the answer — just not yet.

The final counterintuitive point concerns ideas: deliberately trying to think of startup ideas produces plausible-sounding bad ones. The best companies — Apple, Yahoo, Google, Facebook, and the anecdote of a Harvard grad student who wrote VoIP software just to call his girlfriend in Taiwan cheaply — began as side projects never intended as companies. The formula: (1) learn a lot about things that matter, (2) work on problems that interest you, (3) with people you like and respect — which also solves the cofounder problem. This isn't necessarily technical: Airbnb's Chesky and Gebbia excelled at design and organizing, not technology. College should stay classic liberal education rather than vocational "entrepreneurship" programs, since domain expertise built on real curiosity — the path Larry Page took to search — is what matters. Graham's advice for would-be founders, boiled down: just learn.

Among the most successful people Paul Graham knows — founders, programmers, professors — meanness is rare, despite the internet showing how common it is generally. Allowing for selection bias (hedge fund managers, drug lords might be mean), he cites his wife, Y Combinator cofounder Jessica Livingston, an ex-banker with "x-ray vision for character," who has seen good people consistently succeed as founders and bad people fail.

Three reasons why. Meanness makes you stupid: fights reward narrow tricks over general ideas, so your brain "goes fast but you get nowhere, like a car spinning its wheels" — and startups win by transcending rivals, not attacking them. Mean founders can't recruit the best people, who have other options. And benevolence helps: the richest founders aren't money-driven (those sell early); the ones who persist are trying to improve the world.

Historically, success meant winning zero-sum fights over scarce resources — nomads versus hunter-gatherers, Gilded Age railroad barons — where meanness paid off. Archimedes, killed by a Roman soldier in the third century BC despite orders to spare him, shows why idea-driven, non-zero-sum success instead requires civil order and secure property. That mode, long true of thinkers, is now spreading — reversing meanness's historical edge.

A startup's health reduces to one question: given current expenses and revenue growth, will it reach profitability on the cash it has -- is it default alive or default dead? Half the founders Graham talks to don't know which they are; Trevor Blackwell built a calculator (growth.tlb.org) to find out. The question feels meaningless early on and turns critical later, catching founders off guard, and many assume fundraising will save them -- an assumption that grows falser the more they depend on it. Waiting too long risks the "fatal pinch": default dead, slow growth, no time left to fix it. Graham's fix: ask early, and state the truth outright ("we're default dead, counting on investors"). Investors respond to growth (5x/year can attract money pre-profit), but they're fickle, so fundraising should be plan A only, with an explicit plan B. Spending and growth are largely uncorrelated -- fast growth comes from a product hitting a real need, high spending from expensive products or waste. The biggest killer of funded startups is hiring too fast, driven by overestimating workload, copying successful companies' headcounts, avoiding the real problem (weak product appeal), and VC pressure, since VCs' kill-or-cure incentives, protected by the portfolio effect, diverge from founders' need to survive. A common death spiral: moderate growth, an easy first raise, overhiring to force growth, runway exhaustion, and a failed second raise. Airbnb waited four months after YC to hire its first employee, evolving the product through founder overwork.

Paul Graham argues that YC's value comes from two things: an ever-expanding pattern library of startup problems, and the deliberate creation of a founder community, with focus as the thread tying them together.

The pattern-recognition insight: nearly all startups hit the same problems regardless of what they build; after advising 100 companies you rarely see something new. This wasn't obvious when YC started — Graham only had his own startup and friends' as data points — and later-stage investors, seeing far fewer companies in a career, may never learn it. But knowing the problems doesn't mean advising can be automated: each startup still needs specific partners who know it well. YC learned this the hard way in the "batch that broke YC" (summer 2012), when a shared pool of partners worked fine at 60 startups but broke at 80 — an O(n²) problem — fixed by sharding the batch into groups with dedicated partners.

Founders, Graham finds, are often poor judges of their own problems: they arrive worried about fundraising when the real issue is that the company is doing badly, or ask "would you use this yourself?" and admit "no." They also misjudge which problems matter most, fixating on a minor worry while ignoring something fatal. And they frequently don't listen — not from stubbornness alone, but because startup advice is counterintuitive and so sounds wrong until experience proves it right; many return a year later saying "we wish we'd listened."

Because founders have no one but themselves to fix a hundred problems, focus is essential: YC's method is identifying the biggest problem, generating a weekly-resolution fix, testing it, and measuring results — correcting course often enough to be decisive short-term and flexible long-term, like a running back's winding but fast path downfield. "Speed defines startups. Focus enables speed. YC improves focus." Founders' uncertainty stems from novelty, counterintuitiveness, and schooling that rewards "hacking the test" rather than solving real problems — habits YC spends a year retraining.

Equally important is the colleague effect, possibly more valuable than advice: like Renaissance Florence, Göttingen, Bell Labs, or Xerox PARC, YC deliberately built a cluster of ambitious founders, on the belief that great colleagues — unlike seed funding, now a commodity — can't be replicated at scale, and that founders help each other more generously than partners' advice alone provides.

To reach the point of starting a company with Google's odds of success, Paul Graham argues you need exactly three things — skill at some technology, an idea, and cofounders — and all three come from the same source: working on your own projects.

Starting a company is, he says, the standard route to getting rich and to escaping a boss, though not to escaping work; you'll work harder than at a normal job. He and his wife Jessica Livingston started Y Combinator in 2005, which has since funded over 4000 startups, giving him data on what founders actually need. Since no one could have predicted Google's trillion-dollar value when Larry Page and Sergey Brin began, "starting Google" really means reaching a position with the same odds Google once had, not guaranteeing the outcome.

Skill at technology comes from building projects, not from classes: a computer science degree doesn't guarantee real programming ability, which is why tech companies still give coding tests regardless of pedigree. "Technology" broadly means anything involving "make" or "build" — welding, clothing, video — the dividing line is producing versus consuming. Steve Jobs' teenage study of calligraphy, later crucial to the Macintosh's typography, shows any pursued interest can pay off unpredictably. Games are a fine entry point; roughly 90% of programmers start out building them.

Ideas follow automatically once you're skilled: expertise makes "missing" things as visible as a shop's permanently sticking door (an actual sign Graham cites near his house). His examples: paper student directories called "facebooks" at pre-2002 Harvard prompted Mark Zuckerberg to build an online version overnight; in 1997, when most search engines just returned every page containing a searched word unranked, Larry and Sergey built ranking instead. Apple began the same way, as Steve Wozniak's personal computer project, before Jobs suggested selling the plans. The lesson: build for yourself and friends — things they'd genuinely miss if you shut them down — rather than guessing what strangers want, since a young founder's own taste predicts demand better than market research would.

Cofounders, the third requirement, are likewise found through projects, not searches — you learn who's good and compatible only by working alongside them. This is why doing well in school still matters: getting into a selective university matters not for prestige or teaching quality but because admissions difficulty concentrates smart, determined people, who become both cofounders and early employees (Google's first hires came straight out of Stanford). Graham's closing prescription: build things, and do well in school.

The conventional advice given to scaling founders—hire good people and give them room to do their jobs—is really advice for running a company you didn't found, and it damages the companies of founders who follow it. Paul Graham traces this insight to a YC talk by Brian Chesky: Airbnb's results turned disastrous when Chesky followed standard scaling advice, and only recovered once he invented his own approach, partly modeled on Steve Jobs, after which Airbnb's free cash flow margin became among the best in Silicon Valley. Other successful founders at the event reported the identical experience. Graham calls the standard approach manager mode: treating org-chart subtrees as black boxes, never engaging below direct reports. In practice this trains founders to hire "professional fakers" who drive the company into the ground, while founders feel gaslit by VCs who've never founded anything pushing manager mode, and by C-level executives—some of the most skillful liars around—who exploit it. Founder mode instead breaks the rule that CEOs work only through direct reports, normalizing "skip-level" contact, as in Jobs's annual retreat for Apple's 100 most important people regardless of rank. It's messier, requiring delegation to expand as a company grows, but works better. Graham predicts the concept will soon be misused as an excuse to avoid necessary delegation, and closes by contrasting running a company like Jobs versus like John Sculley.

Procrastination can't be cured, only redirected well, because there are always infinite things you could be doing instead. Graham distinguishes three types by what replaces the neglected task: doing (a) nothing, (b) something less important, or (c) something more important. Only type C is good — the "absent-minded professor" who forgets to shave because he's absorbed in a real problem.

"Small stuff" means anything with zero chance of appearing in your obituary — shaving, laundry, thank-you notes, errands generally. Good procrastination is avoiding errands to do real work, which requires annoying the people who want the errands done. Some errands vanish if ignored (replying to letters); others worsen (mowing the lawn, taxes), yet deferring even these can pay off because real work needs large, uninterrupted time blocks and the right mood — an interruption costs not just its own time but breaks the surrounding time in half. This is why startups are most productive as just a couple of guys in an apartment, before hires who are type-B procrastinators drag everyone into their interrupt-driven rhythm. The most dangerous procrastination is unacknowledged type-B: crossing off a to-do list that is itself avoidance of the biggest possible problem.

Citing Richard Hamming's "You and Your Research," Graham recommends asking what the best problem you could work on is, and why you aren't. Big problems are avoided because rewards are distant, failure feels like wasted time, and they're genuinely terrifying — like a vacuum cleaner sucking out every idea you have. The fix is approaching obliquely: start small, grow the project, split the load with collaborators, and let delight pull you rather than a to-do list push you.

To do good work you must like it — but "do what you love" is deceptively simple, since finding and sustaining loved work is hard. Childhood conditioning misleads people: school teaches that work is inherently tedious, and adults reinforce a work/play split, so kids absorb three lies — schoolwork isn't real work, grownup work isn't necessarily worse, and most adults claiming to enjoy their jobs are simply following an upper-middle-class convention of pretending to. Parents who take boring jobs to support their families risk teaching their kids that work is inherently unpleasant.

Once work separates from mere income (a shift Graham dates to college; Einstein's patent-office job is the classic case of the two diverging), the real question becomes what to work on. He sets bounds on "loving" work: an upper bound (you needn't prefer it to every activity at every instant — even Einstein wanted coffee breaks) and a lower bound (you must like it more than any unproductive pleasure, so "spare time" feels like a mistaken category, or procrastination follows). The love should hold over a week or month, not a given second, since pleasures like lying on a beach eventually pall. The test of genuine work — a rule credited to "Gino Lee" — is whether it makes your friends say "wow," a test that doesn't function until about age 22, once you have friends enough to judge by; reading books doesn't qualify, since there's no test of how well you've read one.

The biggest hazards are prestige and money. Prestige — the opinion of people you don't even know, versus friends whose judgment you trust — warps beliefs about what you enjoy, luring people into wanting to want something: aspiring novelists are drawn by the idea of a Nobel Prize rather than the actual grind of "making up elaborate lies." Prestige is "fossilized inspiration": do anything well enough and it becomes prestigious (jazz is his example), so the safer strategy is to do what you like and let prestige follow; between two equally admirable options, pick the less prestigious. Money alone rarely corrupts ambitious people, who aren't tempted by well-paid, low-status work like telemarketing; money combined with prestige — corporate law, medicine — is the dangerous mix. His test: would you do the work unpaid, moonlighting after a day job as a waiter? By that measure, math would persist without math departments, while most gender-and-identity literary criticism exists only because English-teaching jobs do.

So few people escape both childhood's pain-conditioning and the prestige/money traps — "a few hundred thousand, perhaps, out of billions" — that finding loved work takes real discipline, even though doing the work, once found, takes less. Careers tend to zigzag rather than follow tracks laid at age 12; a good self-check is to "always produce" — an aspiring novelist should actually be writing pages even at a day job — which steers you toward what you truly like, not what you wish you liked. Two paths lead to getting paid for loved work: the organic route, where growing eminence lets you shed disliked parts of a job (his example: an architect who gradually gets to choose projects), and the riskier two-job route — earning money at one thing to fund another — dangerous because tedious work "rots your brain" and expenses creep up with age. He warns against deciding too early (a doctor friend who committed to medicine in high school now complains constantly about her job) and argues financial security alone doesn't produce happiness — lottery winners and heirs, freed of constraints, often have no idea what to do with themselves — so better to pick flexible careers, like his own choice of computers, that leave both routes open. Expect a struggle either way: most people fail, and even those who succeed are rarely free to work on what they love before their thirties or forties.

Great new ideas overwhelmingly come from the margins, not simply because outsiders outnumber insiders but because insider status carries structural disadvantages that outweigh its perks. Famous companies started in garages — Hewlett-Packard (1938), Apple (1976), Google (1998) — using the tinkering space mild climates allow; cold places raise the "activation energy" for new projects, forcing every idea to be officially sanctioned first. Apple's garage myth is exaggerated (Wozniak actually designed the Apple I and II in his apartment and HP cube), but Jobs and Wozniak really were marginal: college dropouts with about three years of school between them, whose prior venture was illegal, unprofitable phone-hacking "blue boxes." Graham imagines the government commissioning a Great American Novel to show why insiders struggle: infighting excludes the best writers, a committee hands the job to someone past his prime with a checklist of themes, and twelve years later the result is a derivative "Gone with the Wind plus Roots," a brief bestseller then forgotten. The exercise isolates insider liabilities: wrong selection, excessive scope, inability to take risks, crushing expectations, vested interests, and work curdling into duty.

Whether outsiders can beat insiders depends on how honest a field's tests are. PhD admissions in hard sciences are relatively honest, since professors inherit their own choices as grad students; undergrad admissions are more hackable. A rough diagnostic is the overlap between a field's top practitioners and its teachers: high in math and physics, middling in medicine, law, and computer science, near zero in business, literature, and art — hence "those who can't do, teach." Where tests are corrupt — high-school popularity, corporate politics (Bill Gates could never have risen through GE's hierarchy) — most good people are outsiders, why startups blindside big companies. But winning a corrupt contest has a catch: Chardin, painting amid the flattering portraiture standards of the 1700s, ranks below Leonardo, Bellini, and Memling, who worked under honest ones. Worth competing in a corrupt test, though, if an honest one follows — college admissions matter less once harder-to-game measures of real work take over.

Outsiders hold further advantages: they take risks because failure goes unnoticed, whereas "eminence is like a suit" — impressive but constraining (Lord Acton: judge talent at its best, character at its worst, so one great book redeems ten bad ones). Eminent people, starved for time, delegate — losing problems best solved in one head: glass artist Dale Chihuly hasn't blown glass himself in 27 years, while Wozniak did all the Apple II's hardware and software alone and claims no bug was ever found in it. Expertise also narrows focus, so outsiders win by working broadly across fields or by colonizing something so new no elite gatekeeps it yet — Durer's engravings, once dismissed as devotional bric-a-brac; new computing platforms, always called "not ready for real work" until proven otherwise. HP rejected the Apple II when Wozniak first offered it, partly because using a TV as a monitor seemed "declasse."

Constrained resources push outsiders toward cheap, lightweight, small work, which spreads and evolves faster and can be perfect in a way big institutional projects — forced toward scale to justify budgets and keep staff busy — rarely are. Eminence's other tax is responsibility, dangerous because it doesn't feel like avoidance the way idle procrastination does. Audiences, not money, are insiders' real advantage, and the internet has made audiences newly stealable — bloggers, Reddit, and amateur video already out-compete journalists and sitcoms.

Graham's closing advice: "just try hacking something together," in the spirit of Raymond Chandler's rule for thriller writers — "when in doubt, have a man come through a door with a gun in his hand." Work the margin of the margin — essay-writing, once gatekept by roughly eight New Yorkers, is now wide open — and treat being called "unqualified" or "inappropriate" as a homing beacon: the null criticism insiders resort to once they have no real objection left.

Programmers and other "makers" run on a fundamentally different clock than managers, which is why meetings cost them more. The manager's schedule slices each day into hour-long slots, so booking a meeting is trivial. The maker's schedule needs units of half a day or more — you can't write or program well in an hour — so one meeting can wreck an entire afternoon by splitting it into two pieces too small for real work, and even the anticipation of it makes a maker less likely to start something ambitious that morning.

Trouble starts when the schedules collide, since powerful people mostly run on the manager's schedule and can force others to their rhythm unless they deliberately restrain themselves. Y Combinator is unusual: Paul Graham, Robert Morris ("Rtm"), Trevor Blackwell, and largely Jessica Livingston keep the maker's schedule despite being investors, managing it through "office hours" — founder meetings clustered at day's end so they compress rather than interrupt it. Graham once ran two workdays daily: programming from dinner to 3am, then "business stuff" from 11am to dinner. Speculative "grab coffee" meetings, free for managers, are especially costly for makers — the piece asks only that managers understand that cost.

Determination, not intelligence, is the best predictor of which startups will succeed. Talent gets credit because it makes a better story and excuses onlookers' laziness, but plenty of people as smart as Bill Gates achieve nothing; talent matters more in "purer" fields like math than in messier ones like organized crime, where determination counts for more.

Determination breaks into willfulness and discipline. Willfulness is largely innate (siblings vary sharply); discipline must balance it, or base impulses win and you end up on a local maximum like drug addiction. The model: two fingers squeezing a melon seed — equal pressure sends it far, unequal pressure sends it sideways (footnote: determination proportionate to wd^m − k|w−d|^n). Success breeds temptation, so determination erodes unless discipline keeps rising — why Shakespeare's Caesar feared thin men, untempted by power's perks. Excess discipline can also crush willfulness; the young sometimes succeed by not yet knowing how incompetent they are, gaining confidence through "deficit spending."

A third component, ambition, aims the other two; it's malleable, boosted by exposure to ambitious peers and by each achievement raising the bar further. A separate factor, loving the work, can substitute for determination entirely.

At any given time, most people have one "top idea" that their mind drifts to whenever it's free — and that idea alone gets the benefit of ambient, non-deliberate thought, the kind that quietly solves hard problems (the "answer in the shower" phenomenon). Graham now believes this indirect thinking isn't just helpful but necessary, and since you can't control it directly, letting the wrong thing become your top idea is costly.

He learned this raising money twice, for Viaweb and for Y Combinator: the drain isn't the hours spent meeting investors but that fundraising colonizes the shower-thoughts, starving everything else — the same trap he sees in professors turned full-time fundraisers, and in startups mid-acquisition talks.

Two things are especially worth guarding against: money and disputes, both "velcro-like" attention sinks. Newton, tangled for years in arguments over his 1672 theory of colors with Linus's students at Liege, vowed to stop publishing rather than keep "defending" his work — a reaction Graham thinks his biographer Westfall underrates. Since injuries cost you twice — the injury, then the rumination — Graham cultivates selfish forgiveness, forgetting disputes on purpose. His test: notice what your thoughts return to in the shower, and change it if it's not what you want.

Confidence in a belief should grow monotonically with the experience it survives only if the world is static; since most things change, experts are usually wrong because they're experts on an earlier version of the world. Paul Graham draws on a decade of startup investing—where the best ideas often look bad until some change flips them—for techniques to guard against obsolete beliefs. First, hold an explicit belief in change rather than implicitly treating the world as fixed. Second, don't try to predict the future, since change is inherently unforeseeable; instead stay "aggressively open-minded," holding working hypotheses loosely enough that they never harden, especially once others start identifying them with you. Third, treat weird hunches from people expert in a field as worth exploring—at Y Combinator, calling an idea "crazy" is a compliment. Fourth, force real stakes onto opinions: investors must say yes or no and learn if they were right, and anyone can gain the same discipline by publishing durable, public commitments rather than talking casually. Fifth, bet on people over ideas—Graham funded Airbnb despite thinking it a bad idea, because its founders were earnest, energetic, and independent-minded. Surrounding yourself with such people is the best early-warning system, since expertise will only get harder to hold onto as change keeps accelerating.

Great work requires a third ingredient beyond natural ability and determination: an obsessive, disinterested interest in something that happens to matter.

Graham illustrates the obsession itself using bus ticket collectors, who track minutiae in bus tickets purely because they love it, not to impress anyone or get rich. Darwin's absorption in natural history aboard the Beagle and Ramanujan's hours spent working out series on his slate follow the same pattern — not "laying groundwork" for later discoveries, but pursuing what they found fascinating for its own sake. The difference from bus ticket collecting is that series and species matter and bus tickets don't. Hence the recipe: a disinterested obsession with something that matters. This obsession isn't separate from ability and determination — it's a proxy for the first (you won't obsess over math unless you have some aptitude) and a substitute for the second (curiosity pulls rather than requiring willpower), and per Pasteur, it prepares the mind to get lucky.

Disinterestedness matters most because it lets people find paths to new ideas, which by definition look unpromising — if they looked promising, others would already be exploring them. Darwin and Ramanujan didn't choose their paths for strategic reasons; they simply couldn't turn the interest off, as did Tolkien (inventing an elvish language) and Trollope (visiting every household in southwest Britain). This refines Carlyle's "genius is an infinite capacity for taking pains": the source is infinite interest, not diligence, and it must attach to something that matters. Since no one can know in advance what matters, useful heuristics are: prefer creating to consuming, prefer things difficult for others but not you, and trust talented people's obsessions more. Graham raises an unsettling corollary — reward may be proportionate to risk, meaning great work requires wasting time on paths that turn out to be dead ends, as Newton's alchemy and theology were, though his mathematical physics paid for both; some people surely make all bad bets and simply stay unknown. Returns also shift with era: 1830 rewarded natural history in a way 1709 wouldn't have.

Facing this uncertainty, one can hedge into conventional paths (lower risk, lower reward) or diversify across genuine interests (at the cost of depth). The theory explains uneven achievement across fields (interest is unevenly distributed, ability less so), reduced output after having children, and suggests remedies: relax and follow fun rather than only "important" problems, stay "irresponsible" with side projects to avoid decline with age, and let children go deep into whatever obsesses them rather than staying broad and shallow.

The fear of making something lame keeps people from doing great work — a rational fear, since ambitious projects look unimpressive early on, but most people never get that far; they're too scared to start. The reflex isn't deeply rooted: making new things is itself new to humanity, so we've never developed customs for judging early-stage work and instead judge it by finished-work standards. Silicon Valley is already ahead here, not dismissing unknown people with strange ideas — it has learned to "switch polarity," hunting for reasons an idea might work rather than listing reasons it won't. Y Combinator partners see thousands of ideas every six months and must find the rare power-law winner, making optimism an urgent, practiced skill likely to spread as a custom because it's lucrative. People also dismiss ideas to seem clever (in startups, dismissers are often right — just not when weighted by outcome), or for a darker reason: they don't want you to rise above them, sometimes as national culture. But even self-interested encouragement (investors hoping to get rich alongside you) has hardened into genuine custom. Critics citing impostors like Theranos overlook that its cap table lacked Silicon Valley firms — journalists were fooled, not SV investors — and by revenue, SV's optimism outperforms.

A harder obstacle: your own harsh judgment. Fixes: exaggerate your subject's and your own importance (per G.H. Hardy's A Mathematician's Apology), so overrating your goal's value cancels underrating your progress; be slightly overconfident, armoring against both others' skepticism and your own; exploit the young's advantage of being a lax judge of finished work; find colleagues (not mere cheerleaders) who can spot a baby swan, or a rare good teacher; rely on discipline, though standards rise with age; track your rate of improvement rather than your current state; call early work "just a sketch" or a "quick hack" (Patrick Collison: make implausibility itself a feature, as YC did; John Carmack: crude media like Quake mods, then Minecraft and Roblox, let gameplay ideas dominate); treat risky projects as experiments where failure is still knowledge (Lisa Randall); or lean on curiosity, as Graham did with YC and his Lisp dialect Bel. Since early work truly is undervalued, studying great people's often-feeble first steps (easier now, per Michael Nielsen, given public first commits) trains real judgment — our own harsh standards being themselves early work, a crude "version 1" still evolving.

Earnestness—caring about the right thing and trying as hard as possible—is Paul Graham and Jessica Livingston's highest compliment for founders; paired with being "formidable," it makes someone nearly unstoppable. Motives function like vectors, needing both correct direction (genuine interest in the problem) and magnitude (effort), and the two reinforce each other. Silicon Valley obsesses over motives because so many founders there chase money and fame instead of caring about the problem for its own sake—the same trait that defines a "nerd," someone willing to sacrifice looking cool for what they care about. That caring is also a vulnerability: the earnest can't easily counter mockery with cool nihil admirari, a liability in high school that becomes an asset later, which is why former nerds end up as their former mockers' bosses.

Earnestness doesn't matter everywhere—it's largely irrelevant in politics, crime, gambling, patent trolling, and the more bogus corners of academia, a useful heuristic for which fields to avoid. Earnest people often seem naive, both about others' uglier motives and about a problem's real difficulty; that naive optimism lets them attempt things wrongly assumed impossible. This same naivete is why sophisticated observers misjudge Silicon Valley, dismissing founders' claims of wanting to improve the world as implausible even when many are sincere. Historically, making money was rarely intellectually interesting, hence the blind spot; but since Henry Ford, more business has become genuinely interesting, making it easier to get rich doing what you care about—an alignment Graham calls startups' most important change, and part of why they move so fast.

Nerds seem quiet and diffident, but that's an illusion from seeing them only in ordinary social settings where they're out of their element; a fierce minority are instead highly competitive, taking losses personally because their competitions involve less randomness than most. They run overconfident, which is self-fulfilling up to a point, and needs real intelligence to sustain it. Their independent-mindedness is the aggressive kind — they're annoyed by rules rather than oblivious to them — and they interrupt in conversation, a habit tied to a deeper impatience about solving problems that Graham suspects is the same drive as their competitiveness. James Watson's The Double Helix, opening "I have never seen Francis Crick in a modest mood," is the clearest case: Crick and Watson (both dismissed as "clowns" by one insider) fit the type, and their overconfidence, independent-mindedness, and impatience let them beat two rival groups to DNA's structure. The same traits now pay off in business — 7 of America's 8 richest people qualify — since running a large company demands fierceness even where scholarship (Darwin) didn't. Unexercised, fierceness curdles into bitterness — the troll, the hater, the grumpy sysadmin — which only ambitious work, not meditation, neutralizes. Graham sees this favoring nerds broadly: power has shifted for a century from dealmakers to technicians, a trend he expects to continue until nerds end it via the singularity.

Paul Graham argues that work chosen and directed by oneself — "skating," as he calls it — is categorically different from ordinary work: not necessarily happier (he describes his own essay-writing as mostly worry and puzzlement) but more engaged, and far more productive. He suspects a disproportionate share of great work has been done this way.

Kids naturally experience this excitement (his nine-year-old racing home to write his story; a treehouse as proto-engineering), but adult custom severs the connection: we treat "hobbies" as categorically separate from "real work" and route kids instead through school, where work is usually neither a project nor one's own. Graham says he'd choose ambitious projects over good grades for his kids, citing his time picking startups for Y Combinator: grades didn't predict success, but past projects did.

He distinguishes two senses of ownership. First, voluntary vs. assigned — a sharp edge, since assigned work can still be owned (his mathematician father treated math-homework problems as puzzles, not compliance). Second, solo vs. collaborative — a soft edge, shading into collaboration either by sharing one project (two mathematicians co-developing a proof) or by fitting separate projects together like a jigsaw puzzle (a writer and a graphic designer). Organizations survive partly by preserving individual ownership inside collaboration — exemplified by the original Macintosh team (Burrell Smith, Andy Hertzfeld, Bill Atkinson, Susan Kare), whom Jobs "launched like rockets" rather than commanded, working late nights not from exploitation but excitement — which is why Graham rejects rigid "work/life balance" as a false dichotomy for people who love their work.

This dynamic explains programmers' compulsion to rewrite existing software from scratch (wasteful in characters typed, but more rewarding, and a quiet benefit of capitalist redundancy) and the outsized payoffs of startups, which double productivity by both attracting skaters and letting them skate fully. Autonomy is essential, achieved by being one's own boss or working outside a job — hence open-source projects are a "wormhole" into startup ideas. The hardest constraint is morale at the start: high standards are actively harmful early on, since far more people are deterred by fear of failure than waste time starting too much. Graham closes urging adults to recover children's careless confidence in beginning things, while keeping adults' greater deliberateness in choosing what to work on.

Great work needs three things at once — natural ability, practice, and effort — and no amount of brilliance lets you skip the hard work. Bill Gates ("I never took a day off in my twenties. Not one."), Lionel Messi (whose coaches remember dedication over talent), and P. G. Wodehouse, who at 74 said each book felt like "a lemon in the garden of literature," forcing him to rewrite sentences "ten times. Or in many cases twenty times," show what that combination looks like. Because talent and drive rarely coincide, and you can't change how much talent you have, doing great work mostly comes down to working hard.

School teaches the easy version: pursuing goals someone else set. The harder skill is working toward goals that are neither externally imposed nor clearly defined — learning to feel, unprompted, that you should be working, and to feel awful when you're not. Graham dates his own shift to age 13, when he stopped watching TV; Patrick Collison recalls getting serious at the same age, staring outside and "wondering why I was wasting my summer holiday." School itself was the biggest obstacle, making "work" look boring and pointless. Two kinds of fakeness must be discounted: school-distorted versions of real subjects (G. H. Hardy, in *A Mathematician's Apology*, admits he only wanted to "beat other boys" until reading Jordan's *Cours d'analyse* showed him what mathematics really meant), and work that's inherently bogus busywork. Real work instead has a hard-to-define "solidity" — it feels necessary.

Once you recognize real work, you must gauge daily hours, since quality declines past some point — about five hours for Graham on hard writing or programming, versus almost unlimited hours during the three years he ran his startup, Viaweb. That limit is found only by crossing it, which demands honesty in both directions: noticing laziness but also overwork, since working too hard is often just showing off to yourself. What sustains effort varies by project — fear of failure drove him at Viaweb; the visible flaws in a draft drive his essays now.

Hard problems have a core of important questions surrounded by easier peripheral ones; working hard means aiming at the actual center — not the current consensus, which is often mistaken — both daily and in life's bigger choices of field. Ambitious work is usually harder, but not always: some of the best work comes from finding an easy way to do something hard, often through unusual interest rather than raw talent. Interests mature later than talents, not until your twenties in some cases, and must be distinguished from wanting money or others' approval. Some people's calling converges early, like Mozart's; others, like Newton, who split time between physics and alchemy, never converge, adding a third, ongoing variable — what to work on — alongside judging effort and quality. The best test of whether work is worth doing is simply whether you find it interesting. Working hard, Graham concludes, isn't a dial turned to 11 but a dynamic system tuned continuously; sustained honesty with yourself lets it settle into an optimal shape.

What made Einstein special wasn't sheer intelligence but the new ideas he produced: intelligence is necessary for discovery but not identical to it, and the gap is wide — many smart academics discover little. Offered a choice between being very smart but discovering nothing, or less smart but prolific in new ideas, most people would rightly pick the latter, yet the choice feels uncomfortable. That discomfort is a holdover from childhood, when intelligence is easy to measure and constantly judged while discoveries come later, making intelligence seem like the only thing that matters; it also wins arguments and dominance hierarchies, so society hasn't absorbed that new ideas, not intelligence, are the real destination.

Since intelligence is believed largely inborn, treating it as what matters implies a Brave New World-style fatalism: find your aptitude and work hard. But other ingredients in discovery are more cultivable, trading fatalism for control and a more complicated life. These include obsessive interest in a topic; independent-mindedness; general techniques for pursuing projects and surviving early-stage work, plus startup- and essay-specific methods; mundane factors like sleep, low stress, and good colleagues (youth's link to new ideas may reflect health and fewer responsibilities more than youth itself); and, surprisingly, writing ability, since some ideas can only be discovered by writing them, not recorded afterward. Reframed this way, the intelligence/new-ideas gap becomes fertile ground for discoveries about discovery itself.

The essay proposes that doing great work follows a definable, if not overly narrow, recipe, one Paul Graham built by intersecting techniques used across many fields. The first move is choosing work you have both natural aptitude for and deep interest in; scope for greatness rarely needs separate worry, since ambitious people already err toward projects large enough. Because you cannot know what a field is really like except by doing it, the way to find your subject is to work: develop your own projects rather than only what others assign you, follow "excitingly ambitious" over safe, and let excited curiosity — curiosity to a degree that would bore others — both drive you and point the way. From there, four steps recur across painters and physicists alike: pick a field, learn enough to reach a frontier of knowledge, notice the gaps in it (knowledge looks smooth from afar but is full of cracks up close), and chase the most promising gaps, especially ones others ignore. Graham warns that schools mislead people into thinking a field can be chosen early and definitively; in reality choosing is often years of coevolving with the problem, guided by luck, chance encounters, and casting a wide net.

On technique: hard work yields diminishing returns, so even for the hardest tasks four or five contiguous hours a day may be the ceiling; interruption-proof blocks of time matter more than raw hours. Starting is harder than continuing, so it's acceptable to lie to yourself ("I'll just reread what I have") to get past the activation-energy threshold, and to underestimate a project's difficulty just enough to begin it — though you should still try to finish what you start, since much of the best work happens in the final stage. Per-project procrastination (postponing an ambitious project year after year while looking busy elsewhere) is far more dangerous than daily procrastination because it disguises itself as work. Consistency beats intensity: a page a day yields a book a year, and work that compounds — like learning or audience-building — produces exponential growth that feels flat early on and is chronically underrated. Undirected thinking during walks or showers can solve what frontal attack cannot, but only if it's fed by prior deliberate engagement with the problem.

Taste matters: consciously cultivate a sense of the best work in your field, because aiming merely to be good, rather than the best, usually produces less than good. Avoid affectation — pretending someone more impressive is doing the work — and instead be earnest, which rests on intellectual honesty, willingness to admit error, and informality (focusing on substance over appearance). Be willing to redo or cut material you're attached to; elegance, though it may look like an art-world metaphor, is a real and valuable standard, and the best work often feels discovered rather than invented. Originality is a habit of mind, not a process — original thinkers "throw off ideas like an angle grinder throws off sparks" when engaged with something slightly too hard. It's fed by talking/writing about interests, changing physical or topical context, and distributing curiosity across topics by something like a power law rather than evenly. New ideas require noticing broken models (Einstein's strictness about Maxwell's equations is the example) and a willingness to break rules — either aggressively (delighting in defiance) or passively (indifferent outsiders and novices often discover things precisely because they don't know the rules). Originality in choosing problems matters more than in solving them; good questions are often partial discoveries in themselves. The practical corollary is to start many small things — successive versions, not grand plans — since planning only works for pre-describable goals, and to take on risk proportionate to potential reward.

Youth offers time, energy, optimism, and freedom; age offers knowledge and efficiency — use whichever you have. Unlearn school-induced passivity and the habit of "hacking the test." Copying is fine, even valuable, if done openly rather than unconsciously ("great artists steal"), though copying features you admire that are actually flaws (or copying an eminent person's manner) makes you ridiculous. Seek a few excellent colleagues over many mediocre ones, since you become like the people you work with. Morale, ultimately physical, is the foundation of ambitious work: a small, genuinely enthusiastic audience suffices; avoid people (including romantic partners) who treat your work as competition for attention; treat setbacks as normal backtracking, not proof of failure. If a single word had to answer what drives great work, Graham concludes it would be curiosity — it selects the field, drives you to the frontier, and exposes the gaps worth exploring.

The returns for performance in the world are superlinear, not linear as teachers and coaches imply with "you get out what you put in." A product half as good as a competitor's wins no customers, not half as many — and the company fails. The same pattern holds in fame, power, military victories, knowledge, and benefit to humanity: the rich get richer, and understanding why matters for anyone ambitious.

Superlinear returns reduce to two causes: exponential growth and thresholds. Exponential growth appears wherever how well you do depends on how well you've already done — bacterial cultures, and startups, where Y Combinator has founders focus on growth rate rather than absolute numbers, since if growth rate tracks performance p over time t, reward is proportional to pt. Humans have few customs for handling this because history offered few instances: herding was capped by grazing land, empires compounded through conquest but touched too few people to shape custom, and scholarship compounded too (more knowledge made learning easier) but had little practical effect until the Industrial Revolution let "emperors of ideas" build weapons that beat emperors of territory.

Thresholds are the other source — a step function, as in sports, where the winner gets one win regardless of margin. Thresholds don't require competition; proving a theorem qualifies too. The causes often combine: crossing a threshold can trigger exponential growth (winning battles reduce future losses), while exponential growth can help cross thresholds (network effects letting fast growth shut out rivals). Fame combines both — fans attracting fans, capped by the threshold of limited room on the mental A-list. Learning is the most important combined case: reading is a threshold that accelerates everything after it, and pushing at a field's boundary can crack open an entirely new one, as Newton, Dürer, and Darwin did.

To find superlinear work, seek work that compounds — directly (infrastructure, an audience) or through learning, which is why Silicon Valley tolerates failure so long as founders are learning. Rule: always be learning, but don't restrict yourself to what's already known to be valuable. Thresholds are harder to target — "seek competition" fails if the prize isn't worth it (Russian roulette has a threshold but no upside). A better test: replace things that are mediocre yet still popular. For research, favor curiosity over careerism, since fields that yield new discoveries tend to look mystifying rather than "legit but boring."

This territory is expanding because organizations matter less than in 1970, when prestige was almost entirely institutional; that erosion means more variance in outcomes — good for those confident they're above average, or the young who can afford to gamble. The way to exploit it is exceptionally good work, since competition thins dramatically at the far end of the curve. The condensed recipe: work you're suited to and curious about, hard work without burnout, noticing gaps at knowledge's frontier, affordable risk, good colleagues and taste, honesty, health, and following curiosity — plus luck, best handled by taking multiple shots.

Science is the clearest example — Newton's discoveries reportedly outweighed all his contemporaries' combined — and superlinear returns always imply inequality. Fields with big winners (sports, politics, art, science, startups, investing) mostly share one trait once fame and externally imposed thresholds are set aside: independent-mindedness, since ideas must be novel, not just correct, to pay off. Initial effort in these fields feels wasted ("do things that don't scale"), and the work itself is carried across jobs, unlike the 20th-century equation of "your work" with employment. Ambition climbs existing peaks; curiosity, followed far enough, can grow an entirely new one.

Persistence and obstinacy are not one behavior labeled differently after the fact by who turns out right — they are genuinely different behaviors. The obstinate won't listen: pointed out problems make their eyes glaze over, and they answer like ideologues discussing doctrine. The persistent, by contrast, listen with almost predatory intensity — Graham cites the Collison brothers, who want to know immediately if there's a hole in their boat. Both are "hard to stop," but differently: the persistent are boats whose engines can't be throttled back, the obstinate boats whose rudders can't be turned. On simple problems with only one path, the two look identical, which is why culture conflates them; on complex problems the difference shows — persistence attaches to the goal high in the decision tree, while obstinacy sprays "don't give up" over the whole tree, often anchoring on the first, least-informed idea.

Graham first suspected obstinacy came from being overwhelmed, but rejects this: handing the Collisons a harder problem makes them less obstinate, not more, so the trait must be a fixed personality feature — reflexive resistance to changing one's ideas, closely related to stupidity, and analogous to the anaerobic respiration inherited from distant ancestors: primitive, but useful for preventing panic during a sudden setback. The optimal amount of obstinacy isn't zero.

Persistence, though, has real structure: five qualities — energy, imagination, resilience (setbacks can't destroy morale, though they should change your mind), good judgment (rational focus on expected value), and attachment to a fairly specific goal (too broad and it can't guide action; too narrow and you miss adjacent discoveries). Oddly, persistent people turn irrational choosing among goals of near-equal expected value at the top of the tree, deciding by personal preference — which works, Graham argues, because unconscious preference often tracks real importance others have missed.

Moral fashions are as arbitrary and invisible as clothing fashions, but far more dangerous: dressing oddly gets you laughed at; violating one can get you fired, ostracized, imprisoned, or killed. Every era has believed things later generations find absurd — Galileo was punished for saying the earth moves — so it would be remarkable if ours were the first to get everything right. The goal is general methods for locating what can't be said now.

First, the Conformist Test: holding no opinion you'd be reluctant to voice among peers isn't evidence you're right — it's evidence of unthinking conformity. Like a mapmaker's deliberate error, sharing all the "mistakes" of your era's moral map is no coincidence; you'd have believed the same things as a 1930s German or a pre-Civil War plantation owner.

Second, track what actually gets people in trouble. No one is punished for obviously false claims ("2+2=5"); people are punished for statements that might be true. Since many past taboo statements now look harmless, today's Galileos likely exist too.

Third, follow the labels: "blasphemy," "heresy," "unamerican," "divisive," "inappropriate," "sexist." Graham traces how "defeatist" was weaponized by Ludendorff in WWI Germany and by Churchill's supporters in 1940 to silence dissent without argument. Pick a current label and list ideas it would condemn — the first ones that surface are usually already half-formed in your mind, as shown by a 1989 study finding radiologists' eyes paused on lung-cancer lesions they consciously missed.

Fourth, diff present beliefs against other eras and cultures: anything harmless elsewhere but taboo here flags a likely mistake. Example: an early-1990s Harvard brochure instructing staff not to compliment a colleague's clothing — a norm rare across world cultures, and likely a local, temporary taboo.

Fifth, study "prigs," especially children, whose minds are deliberately sanitized (parents hiding profanity, Santa Claus) and so embody current taboos in pure form. Imagine subtracting the worldview of a well-traveled mercenary-doctor-nightclub-manager type from that of a sheltered suburban teenager; the remainder is what can't be said.

Sixth, the mechanism itself. Ordinary fashion spreads by imitating an admired figure (broad-toed shoes followed Charles VIII's six toes), but moral fashions are usually manufactured deliberately by a group. Taboos are strongest when a group sits "halfway between weakness and power" — strong enough to enforce a taboo, nervous enough to need one; a confident group (Americans, the English) needs none, and a marginal one ("coprophiles") lacks the power to create one. Power struggles get recast as struggles of ideas — the English Reformation was really about wealth, and WWII is remembered as freedom over totalitarianism though the Soviet Union was among the winners. Adoption follows two waves: ambitious early adopters seeking distinction, then a larger group driven by fear of standing out.

Why bother? Curiosity, the wish not to be mistaken, and — most practically — because unthinkable ideas are where great work hides: natural selection was obvious in retrospect, yet Darwin tiptoed around its implications; the U.S. car industry's decline has an obvious cause ("they make bad cars") no one inside the industry dares say. Practicing forbidden thought is like stretching before a run.

Finally, tactics for coping. Don't say what you can't safely say — think it, but keep "i pensieri stretti, il viso sciolto" (closed thoughts, open face), Sir Henry Wootton's advice to Milton before he visited Inquisition-era Italy. Forced to pick a side, answer "I haven't decided," as Larry Summers did rather than take a litmus test. To resist taboos without becoming their target, attack at one remove: fight a label with a meta-label (the term "political correctness" itself helped kill it), use metaphor (Arthur Miller's "The Crucible" indicted McCarthyism through the Salem witch trials without naming HUAC), or use humor, which zealots can't counter. The only real safeguard against fashions invisible from inside is deliberate distance — questioning your own reflex labels, like "hate speech," as hard as anyone else's.

Disagreement online can be sorted into a seven-level hierarchy (DH0–DH6) that measures the form of a rebuttal, not whether it's correct. The web has turned writing into conversation, and disagreement dominates because agreeing gives you little to add while the author has usually already covered the interesting implications of agreement — a structural shift, not rising anger, though more disagreement risks producing more anger, especially online where face-to-face restraint disappears.

The ladder: DH0 Name-calling ("u r a fag," or its pretentious equivalent); DH1 Ad Hominem (attacking the arguer, e.g. dismissing a senator's pay proposal because he's a senator, or an author for lacking authority); DH2 Responding to Tone (objecting to how something was said); DH3 Contradiction (stating the opposite view with little evidence); DH4 Counterargument (contradiction plus reasoning, though often aimed at a slightly different claim than the original); DH5 Refutation (quoting a specific "smoking gun" passage and explaining its error); DH6 Refuting the Central Point (identifying and demolishing the argument's actual core, not minor errors like grammar).

DH levels set no lower bound on convincingness — a DH6 can still be wrong — but do set an upper bound: DH2 or below is never convincing. The hierarchy helps readers spot demagoguery, helps writers self-correct, and, most importantly, correlates with meanness: DH1 is far meaner than DH6, since people with a real point don't need to be mean.

Political and religious arguments are uniquely useless not because those questions lack definite answers, but because the topics have become part of people's identities, and no one can think clearly about something tied to their identity. Graham notes religion and politics have no expertise threshold—anyone with strong convictions feels qualified to opine, unlike on Javascript or baking. He rejects the alternative theory that these subjects are just unanswerable: some political questions (like a policy's cost) do have definite answers yet still provoke the same reflexive fighting. What triggers identity-driven arguments depends on the people, not the subject: a modern battle involving one's own country turns political, while a Bronze Age battle doesn't, since no one has a side. This explains why programming-language debates become "religious wars"—programmers identify as "X programmers"—wrongly suggesting all languages are equally good, when in fact anything designed can be well or badly designed. Even Ford-vs-Chevy trucks can be a minefield with the wrong audience. His prescription: since few labels into your identity are safer than many, don't just tolerate other views—stop considering yourself an "X" at all. A partial exception: "scientist," which commits you only to following evidence, not to any specific belief.

Political moderation comes in two forms, and they are not the same trait: intentional moderates deliberately pick a position midway between left and right, while accidental moderates simply judge each question independently and land in the middle on average because the far left and far right turn out to be roughly equally wrong. The two are distinguishable by the shape of their opinions: an intentional moderate scores near 50 on every single issue, while an accidental moderate's opinions scatter widely but still average to 50. In this respect intentional moderates resemble ideologues more than they resemble accidental moderates — their positions are acquired in bulk and must shift whenever the political median shifts. Accidental moderates go further, choosing their own questions too, so they may ignore issues both sides consider crucial, making "if you're not with us, you're against us" often literally false rather than just rhetorical bullying. Moderates get called cowards, especially by the far left, but accidental moderates need more courage than anyone, since they're attacked from both sides with no group to shelter in. Independent-mindedness only matters for the ideas one actually works with — one could be a doctrinaire Marxist and still be a fine mathematician — so anyone whose work touches contemporary politics faces a choice: become an accidental moderate, or be mediocre.

People sort along two independent axes -- conventional-minded versus independent-minded, passive versus aggressive -- into four durable types, and which quadrant someone occupies depends more on personality than on their society's rules. Children show this: the aggressively conventional-minded are tattletales who want rule-breakers punished; the passively conventional-minded are sheep who obey but only worry when others don't; the passively independent-minded are dreamy kids indifferent to rules; the aggressively independent-minded instinctively question them. School rules are arbitrary yet these types recur, so the sorting is personality, not rules. As kids age, rule-setters shift from adults to peers, so teens who flout school rules in unison are conforming, not independent. Adults show signature "calls": "Crush <outgroup>!", "What will the neighbors think?", "To each his own," "Eppur si muove." Passive and conventional types dominate, making passively conventional-minded the largest group and aggressively independent-minded the smallest. Citing Princeton's Robert George, whose students insist they'd have been abolitionists in the antebellum South, Graham counters that today's aggressively conventional-minded would instead have defended slavery.

The aggressively conventional-minded cause disproportionate harm, Graham argues, and Enlightenment norms -- replacing heresy-punishment with free debate -- exist to restrain them, since new ideas come only from the independent-minded (scientists and startup CEOs must be right when everyone else is wrong). Banning "bad" ideas is dangerous: any banning process is fallible and run by mediocre enforcers competing to out-purify each other, eliminating the margin for error; and ideas are interconnected, so restricting one topic chills unrelated fields with implications there, like soccer played with a minefield in one corner. The independent-minded once sheltered in self-governing courts and universities, but this intolerance began inside universities (mid-1980s, receding by 2000, reignited by social media -- an "own goal" by independent-minded Silicon Valley), though perhaps universities decline because the independent-minded already left for startups and quant jobs. Graham can't predict the outcome but trusts they'll build new institutions if needed.

Conventional-minded people suffer from what Paul Graham calls "orthodox privilege": because their own opinions simply track whatever is currently acceptable to believe, it's safe for them to say what they think — and they wrongly generalize that this safety is universal. They literally cannot imagine a true statement that would get them in trouble. Yet at every point in history there were true things that were dangerous to say; assuming the present is the sole exception would be an extraordinary coincidence, though most people go with their gut instead. The tell is the reflexive question "Why don't you just say it?", sometimes paired with guessing which heresy ("xist or yist") the person must be hiding — and these accusers are sincere, not performing. Unlike other privileges, this one can't be dissolved by learning more; it requires becoming more independent-minded, which doesn't happen in one conversation. Naming the phenomenon helps a little, as does appealing to politeness — the way you'd take someone's word for hearing an inaudible high-pitched noise. The concept also explains the "cancel culture" divide: if you believe nothing true is unsayable, anyone punished for speech must deserve it.

Some kinds of work require not just being right but being right in a way no one else has realized yet — scientists, public-market investors, startup founders, and essayists all fail if their ideas merely duplicate the consensus. Most work isn't like this: an administrator only needs the first half, being correct, with no penalty for being unoriginal. Knowing which category your ambitions fall into matters because independent-mindedness seems to be inborn rather than learned, so picking the wrong kind of work makes you miserable — the independent-minded chafe as middle managers, the conventional-minded struggle at original research. People are also bad judges of their own position on this spectrum: conventional-minded people don't identify as such and genuinely feel they think for themselves, while the independent-minded often don't notice how unusual their views are until they state them aloud, producing a Dunning-Kruger-like mismatch.

Independent-mindedness can nonetheless be cultivated. Nerds do it unintentionally by simply not tracking what's conventional. Surroundings matter enormously: conventional-minded company constrains which ideas you can even express, while independent-minded company encourages more of them — which is why the independent-minded self-segregate once able to, making high school (no chance yet to segregate) uniquely bad, and why growing startups dilute their founding spirit as conventional-minded hires come to outnumber the founders, to the point where founders end up speaking more freely with founders of other companies than with their own staff. Only one or two independent-minded friends are needed, and good universities remain a reliable place to find them. It also helps to cultivate many different types of peers (differing in thought, not demographics) to dilute any one group's influence and import ideas across worlds, and to read history to get inside the heads of people who thought differently. More directly: cultivate a habit of silently asking "is that true?" of everything you're told, treating it as a game whose prize is the novel idea hidden behind a broken conventional one — and watch for intellectual fashions spreading through people like waves, since unfashionable ideas are disproportionately likely to be interesting.

Independent-mindedness breaks into three components. Fastidiousness about truth means calibrating degree of belief carefully rather than letting it rush to the extremes, and produces a horror of ideologies (accepting a whole bundle of beliefs at once, like eating a sandwich of indeterminate ingredients); it can be grown just by thinking about it. Resistance to being told what to think is not mere immunity but a positive delight in subversive, counterintuitive ideas, correlates with having a sense of humor, and is the least improvable — largely fixed by adulthood. Curiosity, the most individual of the three, is what people feel right before having novel ideas, and unlike other appetites it grows when indulged rather than sated. The three substitute for each other like muscles: fastidiousness and resistance clear space in your head, and curiosity fills it. The essay's closing prescription: pursue not "do what you love" but "do what you're curious about."

The archaic religious concept of heresy has returned in secular form: an opinion whose public expression gets punished regardless of truth, and which outweighs everything else the speaker has done. Graham cites Richard Westfall's biography of Newton, who as a new Trinity fellow had only to avoid crime, heresy, and marriage — a line that once sounded medieval but now describes modern employment. Two markers distinguish heresy from ordinary disagreement: calling a statement "x-ist" ends debate rather than testing its truth (no one hedges with "probably x-ist" as they would with "probably fallacious"), and the same statement is judged x-ist depending on who said it. Like a crime, a heresy voids a lifetime of good conduct — you could spend a decade saving children's lives and still be fired for one remark, whereas ordinary opinions are judged as an average, "like dropping a chunk of uranium onto the scale."

Why the resurgence? Intolerance needs two ingredients: aggressively conventional-minded people, who always exist (per his essay "Conformism"), and a unifying ideology with religion-like purity rules, as with Mao's Cultural Revolution. Such an ideology arose in US universities in the late 1980s, producing a wave "eerily similar" in form though smaller. Graham names no specific heresies, partly because heretic-hunters accuse critics of heresy in turn, and so the essay stays applicable to whatever ideology comes next, left or right.

He compares the trend to measles: absolute numbers remain low, but the derivative matters — freedom of expression, after widening for centuries, has narrowed since roughly 1985. He ends optimistic: independent-minded people seem more confident, some on the left now question whether things have "gone too far," and youth culture has moved on, suggesting the wave is peaking.

Wokeness is not new but the latest incarnation of the "prig": a self-righteously moralistic person who proves purity by attacking rule-breakers. Every society has such people; only the rules change — Christian virtue in Victorian England, Marxism-Leninism in Stalin's Russia, social justice today. Wokeness and its predecessor, political correctness (PC), share one definition: "an aggressively performative focus on social justice." The real problem is the performativeness, not the underlying concern: racism is genuine, just not at the scale the woke believe, and PC erred by policing language instead of quietly helping marginalized people.

PC began in the late 1980s in humanities and social-science departments, where politics could permeate scholarship as it couldn't in physics, because 1960s radicals landed there once hired as professors. Those protests hadn't produced PC — students lacked power — but from the early 1970s the radicals joined faculties, gaining numbers and seniority until they too held power. Personally, PC wasn't a thing in 1982 or 1986, was "definitely a thing" by 1988, and pervaded campus life by the early 1990s, once radicals had tenure to enforce ideas, not just voice them. The rules that resulted were elaborate and arbitrary: difficult, shifting orthodoxy substitutes for real virtue, attracting bad people.

Two more factors mattered: religion and sex, the old targets of moral policing, had become dead letters among elites, starving prigs of rules; and the 1989 Berlin Wall's fall discredited Marxism as a rival purity project. PC also skewed female at first, since a 1986 Supreme Court "hostile environment" ruling, via Title IX, let harassment's definition creep from advances to merely voicing uncomfortable ideas — as when Larry Summers was ousted as Harvard president for citing Darwin's greater-male-variability hypothesis (one attendee said it made her "physically ill"), exposing the comfort-versus-truth clash.

PC seemed to die by the late 1990s once comedians made it a joke, but its infrastructure survived — founders became deans, social-justice departments and administrators multiplied — and it reignited in the early 2010s as a more virulent second wave, spreading beyond campuses and multiplying its -isms and -phobias (first "cancel culture," renamed "wokeness" in the 2020s). Graham blames social media (Tumblr and Twitter made outrage shareable — on his own forum, 2007–2014, outraging posts got 3x more upvotes) and group chats, which made cancellation mobs easy to organize; a polarizing press, freed from print-era geographic neutrality to chase ideological markets (in October 2020 the Times announced moving from "stodgy paper of record" to "a juicy collection of great narratives"), feeding a cheap outrage-to-news-to-outrage loop; and a professional "inclusion" bureaucracy, incentivized like Soviet commissars to find problems. Accelerants: Black Lives Matter (2013), Me Too (2017, after Weinstein), Trump's 2016 election (headlines named him at 4x prior presidents' rate), and, largest, the 2020 killing of George Floyd — wokeness's peak.

Wokeness spreads virally: people fear breaking rules they don't know and default to believing accusers; zealots invent an impropriety, more zealots adopt it to signal virtue, a larger fear-driven group follows, and the cycle accelerates — especially in leaderless organizations, which run on "best practices" and can't risk delay. Since 2020 it has retreated: Brian Armstrong and other CEOs renounced it, Chicago and MIT affirmed free speech, Musk's Twitter takeover neutralized it without right-wing censorship, Bud Light suffered a backlash, and disgust with wokeness helped Trump's 2024 win. Graham's prescription: treat wokeness as a religion (as Marxism was) and apply the norms already used for religion in institutions — no orthodoxy-enforcing jobs, no censoring contradictory work, no DEI-statement loyalty tests, no indoctrination sessions. The harder, permanent problem: "aggressively conventional-minded" people are born that way and always want a new orthodoxy; since they give themselves away by inventing new heresies, the defense is a standing bias against banning speech, with the burden of proof on whoever claims harm. His closing principle: the number of true things people aren't allowed to say should never increase.

Programming style holds that a component grown too large becomes a "mass of complexity" that conceals errors, so large programs must be divided into small, comprehensible pieces. The standard method, top-down design, recursively splits a program's purpose into subroutines. Experienced Lisp programmers add a second axis: bottom-up design, building the language itself up toward the program -- writing a needed operator, then noticing it simplifies other parts, until language and program converge on the problem's natural shape. This yields a different artifact, not just a reordered one: a larger language and a smaller program (an arch instead of a lintel).

Four benefits follow: smaller, more agile programs with fewer inter-component connections and thus fewer errors; heavy code reuse, as utilities from one program become a substrate for the next; easier reading, since grasping one general-purpose operator beats parsing many special-purpose subroutines (Graham rejects the objection that readers must first learn all the new utilities); and clearer design, since bottom-up work surfaces repeated patterns worth abstracting. Some bottom-up design happens in any language with libraries, but Lisp's power here makes it "a whole different way of programming." Graham links this to Brooks's Mythical Man-Month: since per-programmer productivity falls as teams grow, small Lisp-equipped teams can "win outright."

Paul Graham argues the next generation of software will run on servers rather than desktops, a huge opportunity for startups, since startups — not big companies — will build it. He traces the pattern to Viaweb, the online store-builder he started with Robert Morris in 1995: planned as desktop software, they switched within a week to running it server-side, browser as interface. It became Yahoo Store, with 14,000 users; the term "Application Service Provider" came only once Hotmail popularized the model.

Desktop software forces ordinary users into becoming amateur system administrators (Graham's mother, alarmed by an Apple letter offering an OS upgrade she didn't understand, is his example). Web-based software removes this: no installation, so no OS incompatibility; free test-drives before "buying"; invisible upgrades instead of yearly shocks; simultaneous multi-user editing; safer data, since the provider handles backups; lower virus exposure, since nothing executes on the client. The cost: tenth-second latency, mattering only for heavily interactive software like Photoshop.

A Web application isn't one binary but a "city" of cooperating programs — editors, background monitors, restarters, statistics compilers, credit-card interfaces — plus server hardware controlled directly, down to building its own servers. Owning the stack means using any language, not the C/C++ desktop software forces; Viaweb's editor, written in Lisp, felt more desktop-like than competitors' stateless CGI.

Because releases are incremental — three to five a day at Viaweb, versus one or two a year for boxed software — bugs stay few and traceable to the last change; "version numbers" persisted only as a press-relations fiction (Viaweb reached "4.1" this way). Bugs are reproducible, since user data sits on the developer's servers, and fresh ones are cheaper to fix than old; Viaweb's purely functional scripting language, RTML, kept code testable. This reshaped support: staff sat thirty feet from programmers, fixing bugs mid-call; after Yahoo moved support elsewhere, morale and bug-catching both suffered.

Fast releases raise morale: ideas ship the day they're conceived, not shelved, and ownership was individual, with no approval beyond peer judgment. Web software needs fewer programmers — Viaweb had three, versus a comparable desktop firm with over 100 engineers, only 13 in product development — inverting Brooks's Mythical Man-Month: smaller teams get exponentially more efficient. The tradeoff: programmers must also be system administrators, on call for a live system that "never ships."

Server access lets you watch real usage instead of guessing: Viaweb optimized until its editor became memory- rather than CPU-bound, got capital cost per user to about $5, and let RTML usage guide UI defaults, e.g. button-bar placement. One fix — a message added after noticing users hitting Back mid-signup — raised test-drive completion from 60% to 90%, lifting revenue growth 50%. Subscription billing fits Web software naturally, is easy to buy, and ends piracy, itself a crude form of price discrimination.

Big companies will eventually adopt Web-based software too, despite it meaning outsourcing IT — Graham argues this is safer, since a startup's existence depends on security in a way an in-house admin's doesn't; merchants who paid roughly $500,000 for custom servers fared worse than Viaweb's $300-a-month customers when Christmas load hit.

This repeats the mainframe-to-microcomputer transition: mainframe software needed too much startup capital, so hackers on cheap microcomputers won — VisiCalc, the PC's breakthrough spreadsheet, was written by two guys, Dan Bricklin and Bob Frankston. Cheap Intel hardware now plays that role for servers, and Microsoft has no automatic foothold; Graham predicts a desktop/server hybrid hamstrung by reluctance to cannibalize Windows. Web-based startups face more stress, since the product never ships and 16-hour days become a default arms race, programmers absorbing both developer and sysadmin anxiety — Trevor Blackwell argues this favors small dedicated teams, since capital is no longer the bottleneck. Graham closes urging hackers not to fear business ignorance or competition: Viaweb launched on under $10,000, and well-run Yahoo proved only a tenth as productive as a small startup.

Programming languages are not interchangeable in power, and language choice is a decisive competitive weapon most managers ignore. The "pointy-haired boss" picks Java because it's a "standard" and assumes all languages are equivalent — but his own history refutes him: in 1992 he'd have insisted on C++, and Java exists only because James Gosling built it to fix specific problems with C++. A language whose developers felt compelled to replace it cannot be equivalent to its replacement. Past the industry's "coolness" hierarchy — Java, then the hacker-favored Perl (which runs Slashdot), then the newer Python that looks down on Perl — lies a pattern: each successive language edges closer to Lisp, invented by John McCarthy in 1958. Lisp isn't stale because it was never really engineered as a language; it began as a theoretical alternative to the Turing machine, and McCarthy's "eval" function was meant only to be read, not run. In late 1958, grad student Steve Russell hand-translated eval into IBM 704 machine code anyway, accidentally producing the first Lisp interpreter. Since Lisp is math, not 1950s technology, it ages like Quicksort (1960), not like 1950s hardware. Fortran (1956) represents the opposite lineage — assembly language with math, no subroutines — and the two trunks have been converging ever since.

Lisp introduced nine ideas: conditionals, functions as a data type, recursion, dynamic typing, garbage collection, expression-based (not statement-based) programs, a symbol type, code represented as trees of symbols, and the collapse of read/compile/run-time boundaries. Ideas 1–5 are now mainstream; 6 is emerging; Python has a partial version of 7. Ideas 8 and 9 together enable macros — programs that write programs — and remain unique to Lisp, partly because adding that power basically forces a language to look like Lisp.

Language choice matters least for small "glue" programs and most for hard problems under competitive pressure. ITA Software's 200,000-line Common Lisp fare-search engine let it out-search Travelocity and Expedia, still using mainframe-era techniques. Graham dismisses the boss's three real worries: server-side control over your whole stack removes interoperability risk; libraries matter mainly for small jobs, not multi-hacker six-month-plus applications; and since good teams run under ten people, hiring is rarely the bottleneck — a more powerful language may need fewer, sharper hackers. Viaweb used Lisp, FreeBSD, and generic Intel hardware despite VC skepticism, on the principle: design for users, not orthodoxy.

Code size is the best proxy for power, since program-writing time scales with length and (per Fred Brooks's Mythical Man-Month) can't simply be bought back with more hires. Lisp-to-C ratios are typically cited around 7–10x; ITA's president is quoted saying one line of Lisp replaces twenty of C. At that ratio, a competitor writing in C would need a year to match what ITA ships in under three weeks, and a five-year effort to match ITA's three-month project — a timeline long enough that, institutionally, the project likely never finishes at all. Even a milder 2–3x gap is enough to guarantee permanent competitive disadvantage.

"Industry best practice," borrowed from accounting, shields managers from blame but yields only average results, not best ones — a case Erann Gat documented at JPL. The recipe Graham offers: attack the hardest problem available, use the most powerful language available, and let competitors' cautious managers revert to the mean.

An appendix demonstrates the power gap with an "accumulator generator" function across languages: trivial one-liners in Common Lisp, Scheme, and Perl; awkward workarounds in Smalltalk and JavaScript; genuine limitations in Python (no lexical closures without a list hack or class simulation); and near-impossibility in Java, which only handles integers via a verbose anonymous class. This illustrates Greenspun's Tenth Rule — sufficiently complex C or Fortran programs end up reimplementing an ad hoc, buggy half of Common Lisp — and Graham suggests OO "design patterns" are often evidence of programmers manually generating, by hand, what a macro should generate automatically.

Succinctness and power are effectively the same thing in programming language design: a language's job is to let programmers say more with less, and the measure of how well it does that job is how small it makes programs. Graham begins from a challenge to Paul Prescod's claim that "Python's goal is regularity and readability, not succinctness" — if succinctness equals power, that reads as "not power," an odd goal to admit to.

High-level languages exist to compress work: ten lines of a good language do what a thousand lines of machine code would. Measuring "small" is tricky — lines of code vary by per-line convention (C wastes many on delimiters), and character counts reward short identifiers (Perl) rather than real compression. Graham proposes counting "elements" instead: every node in the parse tree — variable/function names, numbers, literal text segments, pattern or format-directive pieces, each new block — as the truer measure of the mental work a program demands.

This metric matters most as a design tool: when building a language, comparing "with this feature" against "without it" reveals whether the addition helps, refuting the claim that all languages are equivalent. Chasing succinctness surfaces real abstractions; Forth, Joy, and Icon are cited as languages worth studying for this reason.

On evidence, Fred Brooks (The Mythical Man-Month) found programmers write roughly the same volume of code per day regardless of language, implying more succinct languages ship software faster. Lutz Prechelt's comparative studies, while directionally consistent, use toy problems too small to be meaningful, and any study using a predefined spec tests the wrong thing — like judging embroidery against oil painting when the real test is discovering an image, not reproducing one. Field data are more persuasive: Ulf Wiger's Ericsson study found Erlang 4–10x more succinct than C++, with proportionately faster development and fewer bugs.

Beyond metrics, Graham invokes a "taste test": the felt sense of restrictiveness — like a detour when your street is blocked — is mostly just insufficient succinctness; a shorter forced path doesn't feel restrictive at all. Readability should be judged per-program, not per-line — total effort equals effort-per-line times number of lines, so Basic's readable lines can still cost more overall than Lisp's denser ones, the way a low monthly payment disguises a worse total price.

Finally, Graham distinguishes program-level over-density (a macro must save many multiples of its own length to justify itself) from language-level restrictiveness, and cannot identify any language that forces such crabbed code, since a longer alternative expression is normally always available.

When McCarthy designed Lisp in the late 1950s, it departed radically from Fortran through nine new ideas, most now absorbed into the mainstream. (1) Conditionals: McCarthy invented if-then-else, replacing Fortran's hardware-derived conditional goto, then pushed it into Algol as an Algol-committee member, from where it spread everywhere. (2) Functions as first-class objects, storable and passable like data. (3) Recursion, arguably implicit in first-class functions. (4) Variables as pointers, so assignment copies references, not values. (5) Garbage collection. (6) Programs built entirely from expressions rather than the expression/statement split Fortran imposed because punch cards forced a line-oriented, non-nesting format. (7) A symbol type, testable for equality by pointer rather than content, unlike strings. (8) Code represented as trees of symbols. (9) No firm boundary between read-time, compile-time, and runtime, enabling syntax reprogramming, macros, runtime extension (as in Emacs), and s-expression communication (later reinvented as XML).

Ideas 1-5 are now standard; 6 is entering the mainstream; Python has a partial 7 without real syntax. 8, which together with 9 makes macros possible, remains unique to Lisp -- adding it means designing a new Lisp dialect rather than a genuinely new language. Graham concludes Lisp is best understood not as a reaction against other languages' mistakes but as a byproduct of McCarthy's attempt to axiomatize computation.

Spam can be stopped outright, not merely dodged, because a spammer's message is the one thing they cannot avoid sending -- software that recognizes it leaves them no way around it. Content-based, specifically Bayesian, filtering is the way to do this: Graham's filter catches more than 995 of every 1000 spams with zero false positives.

Hand-written feature rules worked well at first -- flagging "click" alone catches 79.7% of his spam corpus with only 1.2% false positives -- but after six months of refining them, the last few percent of spam grew hard to catch and stricter rules produced more false positives, which he treats as costlier than missed spam ("like an acne cure that carries a risk of death"). Worse, better filters make false positives more dangerous, since users trust and stop checking the spam folder.

Switching to statistics, he builds two roughly 4000-message corpora (spam and nonspam), tokenizes on alphanumeric characters plus dashes, apostrophes, and dollar signs, and counts each token's occurrences. A token's spam probability comes from a formula that doubles the nonspam ("good") count to bias against false positives, ignores tokens under five total occurrences, and assigns .01 or .99 to words seen in only one corpus. A message's fifteen most "interesting" tokens (farthest from neutral .5) combine via Bayes' Rule; an unseen word defaults to .4, since unfamiliar words are usually innocent while spam words are "all too familiar"; scores above .9 count as spam. The results surprised him: besides expected terms like "virtumundo," the filter found "per," "FL," and "ff0000" (bright-red HTML) as strong an indicator as any pornographic word.

The Bayesian approach's edge over feature-scoring filters like SpamAssassin is that a probability is unambiguous where a "score" is not: "sex" indicates .97 and "sexy" .99, and together they imply a 99.97% chance of spam. It also weighs good evidence -- words like "though" push toward innocence -- so an ordinary email mentioning "sex" isn't wrongly flagged. Ideally probabilities are per-user: idiosyncratic words ("Lisp," his zipcode) act as passwords, and a "delete-as-spam" button builds each user's training corpus. Whitelists mainly save computation rather than improve accuracy, since a trusted sender can mail from a new address.

Because the filter retrains on new evidence, it evolves as spammers obfuscate words -- "c0ck" becomes more damning than "cock." Stress-testing his scheme, Graham asks what a spammer who knew the algorithm could get away with: to beat it, spam would have to become indistinguishable from ordinary mail in both body and headers, squeezing spammers toward bare pitches like "Thought you should check out the following: http://...". Spam persists because its response rate, though abysmal (roughly 15 per million versus 3000 for a catalog mailing), still justifies its near-zero cost (about $200, 1/50th of a cent, per million emails, against some five man-weeks of recipients' time). Filtering out 95% raises the spammer's effective cost twentyfold, potentially pricing marginal spam businesses out rather than just hiding their mail.

A four-example appendix shows the scoring: one spam scores all-.99 header/body noise; a second combines words near .99 ("madam," "promotion," "republic") with words near .05 ("shortest," "mandatory") to a final .9027, where "madam" traces to "Dear Sir or Madam" scam letters and "enter" is a genuine miss (innocent here, though usually unsubscribe language); a third slips through because it's full of programming terms (perl, python, tcl, all .01) matching Graham's vocabulary, a fluke he expects word-pair filtering ("cost effective," "setup fee") to catch; a fourth, legitimate email is correctly cleared despite mildly suspicious words like "color" and "California." Further ideas: word-pair/triple scoring for sharper estimates, decomposing domain tokens like "xxxporn," treating known spam features as "virtual words," a non-Bayesian layer against false positives, and a cooperative spam-URL corpus with trust metrics. Finally, spam is redefined not as "unsolicited commercial email" but "unsolicited automated email," since automation -- not commercial intent or an absent relationship -- is what makes a message spam.

Programming languages differ in real power, not just style, and a startup that chooses the most powerful language available can build a durable technical advantage its competitors literally cannot perceive, because habits of mind change far slower than technology does.

Graham grounds this in Viaweb, the online-store-building startup he co-founded with Robert Morris in the summer of 1995 — reportedly the first web-based application, and one of the first major end-user programs written in Lisp. He notes the usual advice, echoing Eric Raymond's "How to Become a Hacker," treats Lisp like Latin: worth learning for the "enlightenment" but not for actual use. Graham rejects the analogy — Latin is useless because no one speaks it, but computers speak whatever language the programmer chooses, so if Lisp makes you a better programmer, you should use it. This matters acutely for startups: a big company growing at the average ~10% a year is fine, but startups have a survival rate under fifty percent, so average performance means going out of business — a startup must do something unusual to survive. Because Viaweb ran entirely on their own server rather than as desktop software, Graham and Morris were freed from the usual pressure to match the OS's language, and could pick Lisp purely because it enabled rapid development. Their hypothesis: faster feature velocity and a smaller team would let them out-build and underprice twenty to thirty eventual competitors. It worked — Viaweb could sometimes clone a rival's press-released feature within a day or two, and Graham describes their advantage as a "secret weapon" nobody suspected, comparing their unthreatening reputation as an "AI language" shop to the disguised assassin in "The Day of the Jackal." They never publicized their language choice. User growth: 70 stores by end of 1996, 500 by end of 1997, 1070 when Yahoo acquired the company in mid-1998, and roughly 20,000 by the time of writing as Yahoo Store.

Graham then argues languages fall on a real continuum of power, not a flat plane of "high-level" equivalence — Cobol sits closer to machine language than Python, and Perl 5 is more powerful than Perl 4 because it added lexical closures, proving one high-level language can beat another. He illustrates the resulting blind spot with a hypothetical middling language, "Blub": a Blub programmer easily sees that lower languages lack features he has, but looking upward he perceives only "weird" languages he assumes are equivalent to Blub — the "Blub Paradox" — so only someone fluent in the most powerful language can see the whole gradient. He recalls his own teenage certainty coding in Basic, which lacked recursion yet felt complete to him at the time. Of the five languages Raymond recommends, Graham ranks Lisp highest because of macros — Lisp code is literally built from manipulable data structures (parse trees), letting programs write programs, a technique he says took until page 160 of his own book to fully explain. He notes roughly 20–25% of Viaweb's editor was macro code, functionality essentially unreplicable in other languages.

The piece closes by reassuring, not converting: he doesn't expect to persuade anyone over 25 to adopt Lisp, and argues its low adoption is precisely the advantage, since the "median language" moves like an iceberg — garbage collection (Lisp, ~1960) is now mainstream, lexical closures (1970s) are barely arriving, and macros (1960s) remain "terra incognita." Big-company programmers face a "pointy-haired boss" barrier, but startups can use this inertia like aikido against rivals. His practical tip: judge a competitor's danger by its job listings — wanting Oracle, C++, or Java signals safety; wanting Perl or Python is mildly worrying; wanting Lisp hackers would have been alarming.

Programming languages evolve like species, branching into evolutionary trees with dead ends, and predicting which branches survive a hundred years hence is the best guide to which languages are worth using or designing now. Cobol is already a "Neanderthal language" with no descendants, and Graham predicts Java will meet the same fate: not because it isn't currently successful by cheap metrics like shelf space or the number of undergrads who think they need it for a job, but because it is a dead end intellectually. Unlike species, language branches can converge -- Fortran is merging with Algol's descendants -- because the design space is smaller and mutations aren't random: designers deliberately borrow ideas from other languages.

Every language reduces to a small set of fundamental operators (its axioms) plus everything else built from them; the axioms matter most for survival, the way location matters most in buying a house, since everything else can be fixed later. Fewer axioms are better, mathematicians' aesthetic, and the main branches likely run through languages with the smallest, cleanest cores. Graham doubts natural language will replace programs; he expects people will still write something recognizable as programs in a hundred years.

Speed will keep changing what counts as an acceptable core. If Moore's Law holds, computers will be 73,786,976,294,838,206,464 (74 quintillion) times faster in a hundred years; even a "paltry million" times faster changes the ground rules, leaving room for inefficient-code languages. But some problems -- video processing tied to another computer's output rate, rendering, cryptography, simulation -- will always absorb unlimited cycles, so future languages must span an ever-widening range between "acceptable" and "maximal" efficiency, making profilers increasingly important as that gap grows.

Graham distinguishes bad waste from good: SUVs are gross because they solve a gross problem (making minivans look masculine), but treating long-distance calls as free like local calls is good waste -- spending resources to buy simpler designs. The real inefficiency to avoid is wasting programmer time, not machine time. So language axioms shouldn't encode efficiency hacks: strings are semantically just lists of characters and exist only for speed, so a cleaner language has only lists, with optional compiler hints for contiguous byte layout -- which is what Arc did, successfully. Pushed further, arrays are just hash tables keyed by integer vectors, and even numbers could be represented as lists, as in John McCarthy's 1960 Lisp paper (never meant for actual implementation) -- absurdly inefficient today, but maybe not "as t approaches infinity."

Layering interpreters is another way to trade speed for flexibility; Bill Woods' rule of thumb is that each layer of interpretation costs a factor of 10 in speed. Early Arc ran as a "metacircular" interpreter atop Common Lisp (CLisp), itself atop a byte-code interpreter -- two slow layers, barely usable but workable. This bottom-up layering, not object orientation, is what actually produces reusable software, despite reusability's having become wrongly associated with OOP in the 1980s; Graham still expects OOP to persist because it lets large organizations accrete "spaghetti code" as patches. Parallelism, promised for two decades, will likewise mostly go unused except when explicitly requested (e.g., forking processes), applied late as an optimization rather than built into version 1.

Language design is shifting from academic research -- where publishability, not usefulness, decides which topics (like static typing, which Graham thinks precludes true macros) get worked on -- to open-source hackers building languages like Perl, Python, and Ruby. Graham's proposal: since a language's core could plausibly have been designed as early as 1958, and might even be worth using today paired with optimization hints, design toward that hundred-year target now -- write the program you'd want with unlimited resources, using brevity (parse-tree size) as a proxy for ease, the way a driver aims at a distant point rather than the road ten feet ahead.

Hackers and painters are the same kind of worker: makers trying to make good things, not scientists. Treating hacking as a branch of science, or software design as "engineering," distorts how code gets designed, taught, and judged.

"Computer science" is a misnomer, Graham argues: it lumps mathematicians chasing grants, natural historians of algorithms, and hackers who treat code as a medium of expression into one department, "like Yugoslavia." "Software engineering" is equally wrong, since good designers decide what to build (the architect's job), not just how (the engineer's). Working inside science departments, hackers feel pressure to publish, but research demands originality and substance that push people toward awkward, ugly systems—Graham cites AI's predicate-logic tradition—rather than beautiful ones. Judging hackers by publication count is as crude as judging aptitude by standardized tests, or programmers by lines of code: easy metrics that "kind of work." The only real test is time; Samuel Johnson said a writer's reputation takes a hundred years to converge, since you must wait out both friends and followers.

Rather than mining theoretical computer science for ideas, hackers should look to painting. Graham describes discovering that he programs by sketching at the keyboard rather than planning fully on paper—debugging as the real substance of programming, not a final cleanup pass. This favors malleable, dynamic languages ("a pencil, not a pen") over static typing, and argues against "math envy," science's habit of dressing work up as formal rather than important.

Companies distort hacking as badly as academia: after Yahoo bought his startup Viaweb, Graham found hackers were expected merely to implement designs from product managers, since spreading out design authority narrows outcome variance—good for avoiding disasters, bad for greatness, because "big companies win by sucking less than other big companies." Startups win by letting the same people design and implement, fighting "design wars" in new, uncontested markets, as Microsoft, Apple, and Hewlett-Packard once did. But startups have their own problem: little time actually goes to hacking (a quarter of his own, at Viaweb), and profitable software rarely overlaps with software that's fun to write. His fix, borrowed from painters and musicians, is the day job: earn money one way, build beauty on the side—essentially what open source is.

More lessons come straight from painting. Hackers, like painters, learn mainly by doing rather than coursework, and should restart projects from scratch rather than endlessly revising one. Painters study masters in museums; writers like Benjamin Franklin (copying Addison and Steele) and Raymond Chandler learned by imitation; open source similarly lets hackers learn from real code, once scarce—Graham recalls illicit photocopies of John Lions's 1977 Unix commentary, unpublishable until 1996. Paintings develop through revision (x-rays reveal Leonardo's repositioned limbs), so programs should tolerate specs changing on the fly; "premature design" is as dangerous as premature optimization. Leonardo's obsessively rendered juniper bush behind Ginevra de Benci shows great work attends to details no one consciously notices—true of code's unseen parts too. Work comes in cycles, so hackers should save routine tasks, like debugging, for low-energy periods.

On collaboration, painters like Verrocchio, whose apprentice Leonardo painted one angel in the Baptism of Christ, or Michelangelo insisting on painting every Sistine figure himself, divided work into distinct, owned sections rather than overwriting each other—the model Graham recommends for software modules.

Finally, since software serves a human audience, empathy separates good hackers from great ones: source code should be "written for people to read, and only incidentally for machines to execute," quoting Structure and Interpretation of Computer Programs—including empathy for your future self, since forgotten code (Graham cites people who swear off Perl) becomes unreadable even to its author. Prestige lags reality, as with the once-obscure Duke of Urbino, Federico da Montefeltro, now known mainly through Piero della Francesca's portrait; just as painting's 1430–1500 golden age, or Shakespeare's and Austen's eras, exhausted a new medium's possibilities early, hacking may be living its own glory days now.

Variation in productivity among programmers is so extreme that it amounts to a difference in kind, and understanding what drives the best of them matters economically because their output is disproportionate and growing. Low-tech work shows little variation (a stick-gatherer beats the worst by maybe 2x), but complex tools like computers magnify differences enormously. Fred Brooks noted in 1974 that top programmers solve a given problem in a tenth the time, but Graham argues Brooks understated it: the harder skill is deciding what problem to solve, not writing fast code, and that imaginative judgment dominates measured output. As technology's leverage grows in every field, a point is reached where 1% of a group produces 90% of its output, so dragging top performers down to average (whether via Viking raids or central planning) is costly — the point behind his book's much-criticized claim that wealth variation can signal productivity variation.

Great hackers are defined less by skill than by loving programming as play rather than a paycheck; companies therefore drastically underpay them relative to value (a 10-100x-productive programmer may accept only 3x the pay), partly because they don't know their own worth. What they actually want: good tools (they refuse bad infrastructure, prefer Python/Perl/open source over Java/Windows, and value the control open source gives them to fix what's broken); a quiet office with a door, since interruption kills hard thinking — cubicle culture (mocked via Dilbert, and even at Cisco, whose CEO also sits in one) misunderstands this, while Microsoft's "give you a door" recruiting ad shows it does; and interesting problems, which can be redefined into existence by ambitious managers (ITA turned boring airline-fare search into something hard and interesting; Google did the same to search; Steve Jobs did it to hardware design by demanding beauty). Managing hackers well requires being a hacker yourself — Graham calls this the "design paradox": you can't recognize good design, or good hackers, without taste of your own.

The worst work is "nasty little problems" — bug-ridden interfaces, ad hoc client customization — because unlike a compiler's coherent challenges, they teach nothing and make people stupid. Companies solve this by insulating talent: at startups, brilliant founders (Robert Morris doing sysadmin work) will grind through drudgery for their own company; big companies wall off an R&D group from customers; or firms can use "bottom-up programming," having elite toolmakers build for internal developers rather than end users.

Great hackers cluster (as at Xerox PARC), making attraction winner-take-all — roughly ten or twenty places draw them at any time, and missing that list means getting none, not fewer. Talent alone doesn't guarantee success (it didn't save Thinking Machines or Xerox), but Graham argues good software usually wins the market, contra 1990s VCs who chased "the next Microsoft" via brand and channel deals — Microsoft, he says, is an outlier explained by one lucky break with IBM, and what worries Bill Gates about Google is its hackers, not its brand.

Identifying great hackers is nearly impossible, even for hackers themselves, since work quality can't be judged from a resume and hackers (like novelists) can't benchmark against each other except through direct collaboration — Graham cites his friend Trevor Blackwell, who built a working Segway clone's software in a single day, as someone he initially misjudged as a fool. This is why tech hubs cluster around universities, where talented peers identify each other by working together.

On cultivating greatness: work only on what you love, since forced disinterest guarantees mediocrity. Interviewed friends cited curiosity and intense concentration (hackers report being unable to code after even half a beer) as hacker traits, which Graham compares to basketball player Bill Bradley's unusually wide 70-degree peripheral vision. He also notes great hackers tend toward political incorrectness — comfort questioning assumptions that carries into debugging fluid, complex programs. His closing recipe: never work on boring projects, and in exchange, never do sloppy work.

Holding an entire program in your head—the way a mathematician holds a problem, walking around it like a childhood house—is what lets a programmer freely reshape not just the solution but the problem itself; only then do you truly understand what you're building. This state is fragile: reloading a program you're actively working on can take half an hour each morning, and returning after months away can take days. Paul Graham lists eight practices that protect it. Avoid distractions—scheduled ones are often worse than unscheduled, since knowing a meeting is coming keeps you from starting hard work at all. Work in long stretches, since there's a fixed cost to loading a program each session (Graham manages up to 18 hours, works best in 12-hour chunks). Use succinct, powerful languages, ideally with layered "bottom-up" programming so only the topmost layer needs holding. Rewrite the program repeatedly, since rewriting forces complete understanding. Write "rereadable" code for yourself rather than merely readable code for others, favoring brevity over spread-out clarity. Work in small groups, since a lone programmer can freely redesign everything while code with several owners resists change. Never let multiple people edit the same code, since no one understands code they only read as well as code they wrote. And start small—a prototype beats a spec for keeping a big problem graspable. Graham notes programmers often hit all eight by accident on unsanctioned side projects, while officially sanctioned corporate projects routinely get all eight wrong, because organizations exist to make people interchangeable and having ideas resists that. The phrase "software company" is nearly a contradiction—which is exactly the gap small startups can exploit against big, individual-suppressing rivals.

The essay taught in school — thesis, supporting paragraphs, conclusion, defended like a legal case — descends from two separate historical accidents, not from what an essay really is; a genuine essay is an open-ended attempt to figure something out, and the two skills it requires are noticing surprises and following them the way a river finds the sea.

The first accident explains why school "essays" are about literature. Around 1100, Europe rediscovered classical texts and spent centuries assimilating them; studying ancient texts acquired prestige as the essence of scholarship, even after, by 1350, better teachers than Aristotle existed. When the German research-university model (imported to Johns Hopkins in 1876) forced writing professors to have a research subject, literature was the only one available, since composition itself yields no scholarship. English departments appeared first at newer schools — Dartmouth, Vermont, Amherst, University College London, in the 1820s — reaching Harvard only in 1876 and Oxford in 1885. In 1892 the National Education Association recommended fusing literature and composition in high school, leaving students three steps removed from real work, imitating professors who imitate classical scholars.

The second accident is the "defend a position" structure. Medieval universities were essentially law schools; rhetoric was a third of the undergraduate curriculum, and today's thesis defense preserves the old disputation, where "thesis" and "dissertation" were once distinct: a position and its defense. Defending a fixed position may suit a legal dispute, but it's bad epistemology: you can't change the question. Graham's own test on a draft is which parts bore readers and which seem unconvincing — fixed by rethinking, not arguing harder, since "as the reader gets smarter, convincing and true become identical."

The real form, he argues, descends from Michel de Montaigne, who in 1580 published essais — French for "attempts." An essay opens with a question, not a thesis, because writing itself, not just thinking, forms ideas; most of what ends up in his essays he discovered only while writing them. It still needs an audience — like having guests forces you to clean your apartment — since essays written purely for oneself "peter out," as do balance-obsessed newsmagazine pieces that flinch from controversial questions.

An essay should meander like the Meander river in Turkey, which winds not from frivolity but because meandering is the most economical route to the sea; Graham had to backtrack seven paragraphs while writing this very piece. The rule at each step is "flow interesting," and for him interesting means surprising — facts that contradict what you thought you knew. He learned to hunt surprise by asking well-traveled friends what surprised them, and later found it even in a teenage fast-food job: Baskin-Robbins kids could name the color they wanted ("yellow") but not the flavor.

Noticing surprises is trainable, like learning history: knowledge compounds because facts grow hooks for other facts — learning that Normans conquered England in 1066 primes you to notice they also conquered southern Italy, and that "Norman" means Viking "north man," arriving in 911, connecting to Dublin's founding by Vikings in the 840s. Useful habits: assume we're achieving only 1% of what's possible, study social and economic rather than political history, and attend to what seems wrong or funny. Coolness (nil admirari) is an obstacle, since being surprised means admitting you were mistaken.

Essayists should also write about whatever they're "not supposed to" — his example is comb-overs, which led him to a real insight: gradualness lets people trick themselves into monstrosities, or, constructively, into building great software or art incrementally from a small kernel. He closes optimistically: the internet, by letting anyone publish and be judged on content rather than credentials or age, is dismantling the old gatekeeping that once restricted essays on topic x to people over forty whose job title contained x — making this, he suggests, the essay's golden age, following the short story's golden age under mass-market magazines.

Writing to persuade and writing to discover are opposed goals, and an essayist must choose one. Social niceties like "pleased to meet you" are harmless lies, but the analogous padding writers use to placate readers is not: it gets woven into the ideas themselves, blurring into real distortion. Graham illustrates with two versions of a paragraph from his labor-unions essay: the blunt original says early union organizers weren't uniquely heroic, since unions are shrinking now for external reasons; a rewritten version makes the identical argument but frames it as sympathetic to unions, softening the same point into something inoffensive. He believes both versions equally but prefers the curt one, because pleasing readers who hold mistaken beliefs requires padding that protects those misconceptions from the truth. Most writers write to persuade, out of habit; Graham writes only to persuade a hypothetical unbiased reader, aiming to discover surprising truths rather than win over actual ones. The two goals diverge further as ideas grow more surprising: the more a conclusion contradicts readers' existing beliefs, the more energy goes into selling it rather than finding it, like accelerating against increasing drag until all effort is consumed by resistance.

Having good ideas is most of what makes writing good, but only a small part of what makes speaking good — the two skills pull in different directions. Graham traces this to two talks: at one conference, a speaker who had the audience roaring turned out, when Graham imagined the transcript, to have said almost nothing; at another, a famous, brilliant speaker made him think, within ten sentences, "I don't want to be a good speaker." The mechanism is structural. Prewritten talks divide a speaker's attention between audience and script; ad-libbing forces you to think about each sentence only as long as it takes to say it, unlike writing, where revision is unlimited. Rehearsing helps polish delivery but steals time from improving content — a tradeoff actors mostly escape since they rarely write their own material. Larger, "dumber" audiences reward flattery, jokes, and bullshitting more, and behave like an incipient mob whose cruder reactions (like the laughter that swept Graham along) spread contagiously. Still, talks aren't useless: they offer conversation-like access to interesting people and are unusually good at motivating action, for better or worse.

Writing improves when it sounds like talking to a friend: spoken language is easier to read, holds attention, and keeps writers honest about how much they are actually saying. Graham traces the problem to a habit of switching into a stiffer, formal register once people start writing -- using words like "pen" as a verb, or, as in Neil Oliver's A History of Ancient Britain, calling Picasso "the mercurial Spaniard," phrasing no one would use in conversation. Complexity and distance let readers' attention drift, and fancy words falsely convince the writer they are saying more than they are. Complex ideas do not need complex sentences: experts discussing hard topics use language no more complex than talking about lunch, partly because they have less to prove and cannot afford to let words obscure the ideas. Exceptions exist -- poetry, a few skilled stylists, and writing meant to obscure, like corporate bad-news announcements or some humanities scholarship -- but spoken language works better for nearly everyone. The fix: draft normally, then revise each sentence by asking whether you would say it to a friend, and read finished essays aloud, fixing what clanks. For writing too far gone to fix sentence by sentence, explain it to a friend and use that as the new draft -- enough alone to put a writer ahead of 95% of others.

The most valuable insights combine two rare qualities at once: generality and surprise, as in F = ma. That combination is hard to find because it's so valuable that the territory gets picked clean; most people manage only one quality without the other — surprising but narrow (gossip) or general but obvious (platitudes).

The interesting middle ground is moderately valuable insights built by adding a small dose of whichever quality is missing. Usually this means adding generality to gossip, so it teaches something about the world. A rarer approach is to start from the most general ideas and hunt for any small delta of novelty — since these ideas are already maximally general, even a tiny addition is valuable, though the result will often resemble existing ideas or turn out to be a rediscovery.

Because of this, Graham argues, thinkers working on general ideas shouldn't worry about repeating themselves: the brain and its stimuli stay similar year to year, so repetition is inevitable, but each retelling varies enough to raise the odds of hitting that critical delta. Ideas also compound — small novelty can lead to more, but only if pursued. It isn't true nothing is new under the sun; some domains yield almost nothing new, but "almost nothing" differs hugely from "nothing" once multiplied across the sun's whole area.

An essay's job is not to persuade but to be useful, and useful writing has four multiplicative components: correctness, novelty, importance, and strength. Correctness alone is cheap — vague academic hedging ("many factors must be considered") is always true and always empty. Real precision pushes a claim as far as it can go without becoming false: saying Pike's Peak is "near the middle of Colorado" is more useful than "somewhere in Colorado," though "exact middle" would be false. Novelty need not mean surprising everyone — telling readers something they knew unconsciously but never articulated counts too, and such insights are often the more fundamental ones.

Graham's method for hitting all four: the "Morris technique," named for his friend Robert Morris, who says nothing unless sure it's worth hearing. Applied to writing, this means deleting any sentence, paragraph, or whole essay that isn't right rather than publishing it — publication bias, condemned in science, is exactly right for essays. His process is "loose, then tight": a fast, exploratory first draft, then days of rewriting, rereading until no sentence still snags like a briar. Essayists have an edge over journalists here — no deadline forces a mistake into print. Importance and novelty both use the same trick: treat yourself as a proxy for the reader, the way founders build products they themselves want; if a topic matters to you or surprises you after deep thought, it will matter to a real number of readers too. Importance is people-affected times how much they care, "a ragged comb, like a Riemann sum." Strength comes from balancing confident claims against precise qualification — "I think" isn't weakness but calibration, and qualifiers can express dozens of distinct things (scope, source, confidence). Simplicity, while not one of the four components, aids correctness because errors show up more plainly in plain language.

This formula also breeds anger: novelty threatens cherished beliefs, creating "dead zones" of unexplored ideas around popular errors; strength plus brevity reads as rudeness ("brevity is the diction of command"); and confident, well-qualified claims become easy to misrepresent by slight exaggeration — critics who ask you to justify a claim rarely quote what you actually said. Graham advises forestalling honest misreadings in the text but relegating defenses against bad-faith ones to end-notes. To start writing, relax the audience constraint (write narrowly for ten readers) or the publication constraint itself — Graham wrote unpublished essays for fifteen years, as Steve Wozniak designed computers on paper before he could afford parts. Because print-era publishing was expensive, most essays, he argues, remain unwritten — and so do the ideas only essay-writing can uncover.

Writing about something you know well typically proves you knew it less precisely than assumed: putting ideas into words is a severe test, and roughly half the ideas in a finished essay only occur to the writer while writing it. Publishing creates the illusion these were fully formed thoughts merely transcribed, but writing itself changes ideas, discarding some as unfixably broken. The real test is rereading as a naive stranger who knows only what's written — if unsatisfied, you add what's missing, even at the cost of elegant sentences. Only in formal domains like chess or math can people form complete ideas purely in their heads — arguably still a kind of silent writing. Graham cites writing about two subjects he knew well, Lisp hacking and startups, as teaching him things he hadn't consciously realized. Talking can serve the same function, but writing is stricter — no tone of voice to lean on, commitment to one exact sequence of words, two weeks and fifty rereads per essay. The unsettling implication: anyone who has never written about a topic doesn't actually have fully formed ideas about it, however complete those ideas feel. Writing can't guarantee ideas are correct, but it is a necessary condition for them to be.

The best possible essay would cover the most important topic on which you could tell readers something they don't know, which makes essay-writing a question about science: the best essay at any moment describes the era's greatest discovery, and Darwin's 1844 manuscript on natural selection would have been the best essay of that year. Great essays, unlike great paintings, aren't timeless: writing about natural selection today would be worthless now that everyone knows it. So "write the best essay" reduces to "make the best discovery" — useless advice — leaving the real question: how to write essays well, meaning how to discover ideas by writing.

An essay starts from a "question," broadly defined; you need an "edge," some insight or way in, and curiosity about something seemingly minor (Darwin's "how can they all be finches?") can be edge enough. You commit a specific, usually wrong or incomplete, answer to words; rereading strictly exposes the gaps where new ideas hide, sometimes a false assumption beneath a claim. Each response should advance toward truth and spawn further questions; since essays are linear but responses form a tree, you follow the branch with the most generality and novelty. Heavy rewriting means you needn't guess right first: Graham cut a 17-paragraph subtree from this essay, reattaching five paragraphs and discarding the rest — the rule being to cut fast anything not right.

The initial question still matters, even though idea-space is highly connected, because in an essay you don't know your destination going in — you stay near the start, backtracking once you've wandered too far. It bounds the essay's quality but shouldn't be over-optimized, since the unpredictability of where a question leads is the point; the fix is to write many essays and take risks, since most yield good essays and only some yield great ones. Great questions tend to be outrageous — counterintuitive, overambitious, or heterodox, ideally all three — and must genuinely interest the writer, since only real caring sustains the stretch for novel insight. Graham rarely picks topics deliberately; he writes about whatever he's thinking about, so the lever is improving what pops into your head: breadth, from learning disparate things (his Hay-on-Wye book-buying trips), and depth, from real work and the right people — an afternoon with Robert Morris beats talking to twenty new smart people.

Every subtree of an essay is itself a smaller essay, like a subtree of a Calder mobile. A subtree ends when you feel sated, not lazy, since you could always start a new essay's question instead; discoveries made along the way are drag that keeps idea-space feeling less connected than it is, though writing many essays burns off that drag over time. You can then stop, or return, Cubist-style, to an earlier skipped branch, as Graham does here with his claim that essays aren't timeless. Timelessness has two senses that align in art but diverge in essays: being permanently important, versus always affecting readers equally. Essays teach, and you can't teach people what they already know, so "evergreen" timelessness needs an essay whose discoveries never get absorbed into shared culture — one about things people never learn except by living them, or about lies adults and institutions tell, like schools teaching students to hack tests, which fails in real life. Ideas that stick and become obvious later count as success, not failure — "Darwin territory." Timelessness is one case of a deeper aim: breadth across time and fields, along with novelty, which Graham says he's always chasing.

Short of discovering something as big as natural selection, essay quality comes down to procedure: casting a wide net for questions, which depends on inspiration, and being exacting with answers, which depends only on persistence, since you can rewrite until they're right. Inspiration is the real bottleneck, a pool with no bottom; the hardest, most important question, Graham concludes, is how to get more questions.

Paul Graham predicts that within a couple of decades few people will still be able to write, because AI has removed the pressure that once forced people to learn. Writing is fundamentally hard since it requires clear thinking, yet it pervades the most prestigious jobs. The tension between this pervasive expectation and the skill's real difficulty explains why eminent professors turn to plagiarism, typically stealing mundane boilerplate — proof they were never even halfway decent writers. Historically the only escape valves were paying someone to write for you (JFK) or plagiarizing (MLK); otherwise you had to learn. AI now removes that pressure entirely, in school and at work, so the middle ground of "ok writers" will vanish, leaving only good writers and non-writers. Graham argues this matters because, per Leslie Lamport, "if you're thinking without writing, you only think you're thinking" — writing is thinking, so the writes/write-nots split becomes a thinks/think-nots split. He compares it to physical strength: preindustrial labor made everyone strong, but now only those who choose to work out are. Likewise, there will still be smart people — only those who choose to be.

Writing that sounds good is more likely to express ideas that are right — not just true, but well-developed, with the right conclusions explored to the right depth. Paul Graham says he never has to choose between the best-sounding sentence and the best-expressed idea; fixing a sentence's sound tends to fix its logic too. He traces this to laying out his first book, where forcing a cut of one line always left the passage better — like shaking a bin of objects, which only packs them tighter since gravity forbids the reverse, a forced rewrite can't make ideas less true. Sounding good also helps directly: prose that's easier to read is easier for the writer — who rereads a draft 50 to 100 times — to check for anything wrong. Good writing's rhythm follows the natural shape of a train of thought rather than regular meter, so fixing the rhythm often means fixing the thought, as Kelly Johnson said of planes: if it looks good, it flies well. This bond breaks when ideas come from elsewhere, as in experiments or textbooks summarizing others' work, since the ideas live in the work, not the prose. A skilled liar can write something beautiful yet false by first half-believing it; such writing is internally consistent rather than true, though in an honest writer the two converge. The two senses of good writing are a rope, not a rod: hard to be right without sounding right.

Smart kids are unpopular in American schools not because intelligence is punished, but because popularity is itself a demanding, full-time skill, and nerds are too preoccupied with other things to compete for it.

Graham recalls that in junior high, he and his friend Rich mapped their school's lunch tables by popularity, grading A through E; they sat at a D table, and everyone in school would have graded them the same way. The puzzle is why smart kids don't just reverse-engineer popularity the way they do standardized tests. It isn't that other kids envy or punish intelligence directly — if being smart were genuinely enviable, girls would have liked the smart guys, which they didn't; intelligence simply didn't register much either way. The real answer is that nerds don't want popularity badly enough. They want something else more: to be smart, to write well, to program, to make things. Given a hypothetical trade of popularity for average intelligence, Graham says he'd have refused it, even though most kids would take that deal.

Popularity, he argues, is not something you have but something you constantly do — citing Alberti's line that "no art, however minor, demands less than total dedication if you want to excel in it." Teenagers work at it harder than Navy SEALs work at soldiering, mostly unconsciously: their clothes, speech, and behavior are constant, low-level bids for others' approval. Nerds don't grasp that popularity takes this kind of effort, just as most people don't realize drawing ability comes from practice, not innate gift. Popular kids are trained by their parents to please; nerds are trained to get right answers — and since attention is finite, nerds who can't spare enough for the social game fall to the bottom by default, "nerd" being a purely relative label to whatever the local standard demands.

This explains why ages roughly eleven to seventeen are worst: around eleven, kids start treating their families as a "day job" and build a peer world from scratch, which — left ungoverned by adults, as in Lord of the Flies — degenerates into cruelty. Unpopularity isn't just neglect but active persecution, for three reasons: children are intrinsically cruel before developing a conscience; insecure kids elevate their own rank by pushing others down (Graham compares this to poor whites' hostility to Black Americans); and most importantly, popularity runs on alliances, and nothing bonds a group like a shared enemy — so bullying is mostly a group phenomenon, not individual sadism, disproportionately driven by the "nervous middle classes" rather than the most popular kids, who don't need it. Because the unpopular tier is small (a pear-shaped distribution, not a pyramid), there are more kids who want a scapegoat than there are scapegoats, and even sympathetic kids ostracize nerds to avoid guilt by association.

Adult life differs, Graham argues, not because adults are more mature (prisoners and Manhattan society wives show otherwise), but because the real world is large and consequential — right answers matter (Bill Gates gets results despite poor social skills), and nerds can cluster into their own critical-mass subcultures (John Nash imitated Norbert Wiener's mannerisms to signal genius). Schools, by contrast, function like prisons: their real purpose is warehousing kids while adults work, a byproduct of industrial-era specialization stretching job training into the late twenties and making teenagers economically useless. With no real external test of merit, school status "degenerates into a popularity contest," unlike a sports team's hierarchy, where a rookie doesn't resent a veteran. This same purposelessness — not hormones, which Graham notes went unmentioned before the twentieth century and didn't afflict Renaissance apprentices — produces both the cruelty and the boredom, and drives some kids into drug-using "freak" subcultures as parallel rebellion. His conclusion: nerds aren't losers, just playing a different, more real game than the one school forces on them, and recognizing school as temporary and artificial is itself a partial cure.

The essay's central claim is that "what do you want to do with your life" is the wrong question to ask a teenager — instead of committing to a goal and working backward, you should work forward from promising situations toward whatever increases your future options, then let curiosity about hard problems, not a fixed plan, determine the path.

Graham begins by dismissing the standard graduation-speech advice ("don't give up on your dreams") as dangerous, since it implies being bound by an early plan — the "premature optimization" of a life. He also attacks the myth of genius: biographers streamline lives into destiny, but the sixteen-year-old Shakespeare or Einstein would have seemed merely impressive, not a different species; believing in innate genius is mostly an excuse for laziness. Still, some variation in ability is real (a four-foot-tall kid won't play in the NBA), so the speech should really say "what someone else with your abilities can do, you can do" — but that phrasing gives no actionable guidance.

His replacement is "stay upwind," borrowed from gliding: since a glider can't fly into the wind without losing altitude, you must keep your options open by choosing, at each point, whichever available path preserves the widest range of future possibilities rather than committing early — e.g., a college freshman should major in math over economics, since math keeps the economics-grad-school door open but not vice versa. Concretely, this means finding smart people and hard problems. Spotting real intellectual rigor is tricky because much of academia fakes difficulty; he cites the *Social Text* affair, in which a physicist submitted deliberately nonsensical, jargon-laden prose to a literary-theory journal and got it published, proving some fields reward obscurity over substance. The test for real difficulty is worry: if you're not anxious your work might come out badly, it isn't hard enough — and overcoming that worry, not avoiding it, is what makes achievement exhilarating.

On ambition, Graham argues most people need to feel good at what they do, but high schoolers are fed a fake job (being a "student") that misdirects this need. The real adult/child divide isn't earning a living but taking intellectual responsibility for oneself; he'd treat high school as a "day job" — done adequately but not identity-defining — while pursuing real work on the side. The regret nearly everyone reports about high school is wasted time and boredom, which he says isn't inevitable, since the same people weren't bored as eight-year-olds; extracurriculars like charity drives don't count as "getting something done" because they aren't hard.

The "Corruption" section indicts the college admissions process: admissions officers, unlike professors, can't actually judge intelligence, which is why prep schools can "hack" a kid's apparent quality without changing his actual ability. This traps students designing their lives around a process as arbitrary as reality TV. Graham warns against rebelling (his own high-school mistake) as much as against obedience — both let external rules define you. The fix is an "orthogonal vector": treat school as a day job, neither fighting it nor being consumed by it.

Under "Curiosity," he redefines "passion" and "aptitude" as really meaning curiosity — adult curiosity is narrow and deep rather than broad and shallow like a child's. He argues, using Wittgenstein's admitted lack of self-discipline and stories of chronically procrastinating high achievers, that genuine interest replaces discipline once work begins; Einstein's breakthrough came from genuine bafflement at Maxwell's equations, and mathematician G.H. Hardy only grew to like math once he started asking questions rather than answering them.

Practically ("Now"), Graham advises picking small, sub-month projects chosen purely for interest, iterating until the process becomes self-sustaining, involving only serious friends, and hunting actively for the rare good book, since most textbooks are bad. He closes by redefining adulthood as simply deciding to take responsibility for your life at any age — hoping his readers' biggest high-school regret won't be wasted time.

Business's real lesson from open source and blogging isn't about Linux or Firefox, but about the forces that produced them — forces that will reshape far more than software. A Forrester survey (Business Week, Jan 31, 2005) found 52% of companies replacing Windows servers with Linux; more telling than the number is which companies — anyone still deploying Windows on servers should explain what they know that Google, Yahoo, and Amazon don't. Open source and blogging share key traits: both are made for free by people who enjoy the work, both beat paid professionals, both police quality Darwinically (the audience spreads good work and ignores bad, instead of relying on rules to stop employees screwing up), and both were unlocked by the Web lowering the cost of reaching an audience.

The first lesson is that people work harder on what they love — "amateurs," a word whose etymology (love) got buried under twentieth-century "professionalism," itself largely an artifact of narrow "channels": with only a few journalism jobs, competition made the average professional good, while anyone's bar-room opinion sounded amateurish by comparison. Online, that constraint disappears — nobody reads the average blog, so the real comparison is best online writer versus professional, and traditional media are losing that fight. Graham reports reading far more individual writers' sites than newspaper sites, and finding New York Times stories through aggregators like Google News, Slashdot, and Delicious rather than the Times' own front page — evidence of how little overlap exists between what editors choose and what's actually interesting. Most "news" (a presidential speech restating known opinion, a shark attack, a plane crash) isn't new at all; when professionals produce that, amateurs can beat them.

The second lesson is that the standard office is a bad place to work. Startups, often begun in apartments with used furniture, odd hours, and casual dress, are typically at their most productive stage. Fixed office hours exist mainly because companies can't measure output, so they use presence as a proxy — producing "facetime," meetings ("an opiate with a network effect"), and workdays fragmented into unproductive pieces. Graham proposes a "Work Day" experiment: one day per company with meetings banned. His own company had no fixed hours (he never arrived before 11am), and the startups he funds work out of apartments not to save money but because separating "work" from "life," a tenet of professionalism, actually hurts the work.

The third lesson is that ideas and quality control can flow bottom-up rather than top-down, resembling a market economy even though companies that praise markets often run themselves like communist states internally. Open source software is more reliable precisely because anyone can find bugs; Graham felt safer publishing essays for Hackers & Painters that had already had thousands of page views online than ones that hadn't been "tested." Steve Wozniak's employer HP turned down his pitch to build microcomputers, forcing him to leave and found Apple — top-down management blocking a good bottom-up idea.

No single manager benefits tomorrow from these forces — companies don't get smarter, but dumb ones die, like a gene pool. The deeper implication is that the employer-employee relationship, carrying "master-servant DNA" (the word "boss" comes from Dutch baas, "master") plus legal cruft around hiring and firing, should give way to an investor-founder relationship: paying people as investment rather than salary so they work on projects that are both creatively and economically their own. Google partly recognizes this with stock grants and 20%-time. Graham's own experience — feeling infantilized working for bosses after Yahoo acquired his startup, versus energized as a founder — illustrates the difference; he doesn't argue everyone should start a startup, only that more ambitious people should take investment over employment.

It may not matter much where you go to college. Graham grew up treating a good college as the bottleneck for his entire future, and only recently, prompted by parents obsessing over kindergarten admissions, did he realize he no longer believed that.

His evidence is Y Combinator, the seed-stage firm he runs with three partners, funding about 40 startups a year chosen from roughly 900 applications covering about 2,000 people, most founders around three years out of college. Unlike a hiring manager, YC gets a fast, unambiguous verdict: a startup succeeds or fails within a year, decided purely by whether users like the product, and users don't care about pedigree. Judging so many people against that clean test, Graham and his partners caught themselves assuming MIT, Harvard, and Stanford graduates "must be smarter than they seem," but learned school prestige barely predicts founder quality — the variation between individuals swamps the variation between schools.

He explains why big organizations still lean on brand names anyway: like buying IBM, hiring elite graduates is a "safe" choice for a recruiter no one holds accountable for rejected talent. He cites Mitch Kapor's wife Freada, who ran HR at Lotus and once sent recruiters the anonymized resumes of the company's first 40 (already-proven) employees — not one got an interview. Elite graduates also tend to be obedient and confident, traits suited to old corporate hierarchies where individual performance is hard to measure, but worthless against a startup market's judgment.

How much you learn, Graham argues, depends far more on the student than the school — you can't reliably distinguish graduates of colleges three times apart on the US News list — because the real asset at elite schools is smart peers, reproducible elsewhere, and faculty quality varies far less across schools than student quality does. The real cost of not attending a prestigious college is one's own insecurity about it; what actually matters is what people make of themselves.

Adults lie to kids constantly and systematically, and the "conspiracy" is broad enough that everyone agrees which questions get "ask your parents" — proof the lying is coordinated. Kids usually discover the deception only by tripping over contradictions; Einstein described the "crushing impression" of realizing the Bible wasn't true and the state was deceiving youth. Graham felt the same by 15.

The most common justification is protection, reasonable for a newborn who needs a quiet, safe world. Stretched to 18, though, that shielding becomes something like abuse — visible in suburban teenage malaise, since suburbia suits 10-year-olds but turns suffocating by 15. A friend who left Manhattan said her 3-year-old had "seen too much": drugs, poverty, madness, gruesome medical conditions, sex, violent anger — anger, Graham says, would worry him most. Parents rarely audit this, leaving 18-year-olds with unearned confidence about running the world — a distortion only ~100 years old, begun with royal children and spread by suburbia.

Parents' discomfort with teenage sex exceeds the practical risks of pregnancy and disease, pointing to an inborn taboo against child sex that persists once kids are biologically capable — many societies accept teen motherhood, so the instinct isn't fully universal. Adults also hide that sex and drugs bring real pleasure, which combined with poor teenage judgment makes both dangerous; telling the truth wouldn't work, since bad judgment includes believing you have good judgment.

Kids stay "innocent" partly because helplessness is endearing (cuteness is helplessness), partly because learning the world is brutal too early forecloses curiosity; smart adults, Graham says, often stay innocent for this reason. Swearing is banned for no semantic reason ("shit" vs. "poopoo") but purely to mark adult status.

Death is concealed because small children find it terrifying. Graham's parents invented an anesthesia-death story for their cat, revealing decades later his sister, then three, had broken its back; his grandmother recast his grandfather's hours-long fatal heart attack as simply going still and "gone." Graham didn't grasp his own mortality until 19.

Ethnic and religious identity requires two lies: that a child belongs to group X, and the arbitrary false beliefs that differentiate X from outsiders. That arbitrariness is what makes identity stick, and it can carry a "payload" of useful values (e.g., "Xes are honest") alongside bizarre customs.

The least excusable lies protect adult authority — from a molester silencing victims to a father hiding an affair out of vanity, or "issues" books like Peter Mayle's Why Are We Getting a Divorce?, insisting divorce is never one parent's fault. This double standard leaves kids feeling disproportionately guilty. Teachers rarely admit "I don't know" — Graham learned teachers were fallible only in sixth grade, when his father dismissed one as "just an elementary school teacher."

Curricula compromise among interest groups via omission and emphasis: Carver was ranked with Einstein and Curie because he was black, not because his work compared — the honest lesson, Graham thinks, would have taught more about the obstacles blacks faced. Kennedy and King were sainted despite womanizing, Kennedy's amphetamine use, and King's plagiarism; the biggest school lie is that success comes from "following the rules," when rules are mostly hacks for managing groups.

The deepest reason for lying to kids mirrors why kids lie to adults: avoiding a bad reaction. When a friend's 5-year-old asked if the Thanksgiving turkey wanted to die, his parents improvised that it did. Such calming lies train people not to worry — Graham's mother soothed away his childhood panic over a pollution documentary — and this, he argues, is why real problems persist into adulthood.

Graham closes with an oxygen-debt analogy: adults carry a "truth debt" nobody formally corrects, since parents forget most of what they told. Clearing it is solitary work — feeling like a skeptic leaving high school doesn't make you neutral, since a mind isn't a blank slate by default; it must be consciously erased.

Working for a large organization is intellectually unhealthy the way processed food is physically unhealthy: a normal job may be as bad for a programmer as white flour and corn syrup are for a body, because humans weren't designed to work in groups of hundreds. Having worked with over 200 startup founders, Paul Graham noticed they aren't necessarily happier than employees, but happier the way a wild lion is more alive than a zoo lion — working in a more natural, if statistically abnormal, way.

The root problem is group size. Impala herds run to about 100, baboon troops to 20, lion prides rarely exceed 10; humans similarly work well in groups of about 8, strain by 20, and become unwieldy at 50. Large companies, forced to divide into smaller units, arrange them in tree structures with bosses as the connecting points. A group of 10 managers isn't just 10 people — each represents an entire subgroup, so the whole must act as one person, meaning individual freedom shrinks in inverse proportion to the size of the entire tree. A team of 10 inside a big company thus feels like a "fake tribe": the right size, but missing the initiative real tribes have — the job equivalent of high-fructose corn syrup, appealing yet lacking something essential, much as junk food beats vegetables by being cheaper and more immediately gratifying.

This hits programmers hardest, since programming means building new things and tree structures resist novelty. A founder who joined Google expecting to learn more than at a startup instead found legacy code, overhead, and other teams' interfaces blocking most of what he wanted to try; blocked ideas stop occurring, while startup-level freedom generates more of them, like a low-restriction exhaust boosting an engine.

Large organizations therefore inevitably slow down as they grow; avoiding this would require structureless, market-like organization (unexplored so far), so companies should instead stay small and hire only the best, since mediocre hires get less done and force further hiring. Individually, Graham advises working only for small companies or starting one's own: a failed startup leaves net worth at zero rather than negative, and Y Combinator founders — often conservative from big-company backgrounds — visibly transform within months into people who "come to life," happier and more alive, like lions returned to the wild.

Every check an organization adds against past mistakes carries a hidden cost that almost never gets weighed against its benefit. Requiring suppliers to prove solvency, for instance, screens out the best bidders who won't bother qualifying or fall just short of an inflated threshold. Joel Spolsky told Y Combinator that once corporate software purchases exceed about $1000, they need committee approval — so costly to navigate that vendors must charge $50,000 for something otherwise worth $5000, meaning the buyer pays ten times more. Government checks scale the damage: China's restriction on long trading voyages after 1400 let Europe overtake a once-richer, more advanced China; Sarbanes-Oxley's added checks on public companies crippled the US IPO market by burdening young firms that General Electric could shrug off. The cost is rising because software matters more, and checks hit programmers hardest — unlike most workers, good ones prefer working hard and shipping fast, which is why startups impose none. Three programmers whose acquired startup went from instant releases to two-week approval cycles said they'd trade up to half the acquisition price for instant shipping back. Steve Jobs's "artists ship" cuts both ways: artists insist on shipping, so blocking releases loses your best people.

American life is increasingly governed by performance rather than credentials, a reversal from a generation ago — contrary to South Korea, where a 2008 New York Times piece quoted a parent saying college entrance exams determine "70 to 80 percent" of a person's future, a claim that would have fit the US 25 years earlier too.

Credentials themselves were originally progress: China's imperial civil service exams, introduced in 587 AD (per Miyazaki's *China's Examination Hell*), replaced pure bribery and family influence with tested performance. But credentialing inevitably spawns cram schools — "leaks in a seal" that convert one generation's wealth into the next's credentials, seen in Ming China, nineteenth-century England's Sandhurst prep, and modern SAT tutoring. Since parents will always push to transmit advantage directly, any society aiming to reward merit must fight this, much as a prison fights smuggled heroin.

One fix is tightening credentials — pushing transparency at bottlenecks like college admissions, where practices like legacy admissions let institutions covertly weight family status by adjusting cutoff "buckets." But this is slow, especially since credentialing institutions don't really want airtight tests.

The better fix is making credentials matter less by measuring performance directly. Credentials exist because large organizations can't measure individual output; markets of many small organizations approximate doing so by keeping only the good. This explains the US shift: an economy of big companies (where pedigree becomes self-fulfilling) is giving way to one of startups, where a company that ignores real performance simply goes bankrupt. Graham, as a VC, describes himself as pushing people from the credentials world into the performance one.

A parallel driver: mid-century firms paid by seniority, not value (Graham cites 1950s law associates and his father's underlings at Westinghouse earning more via tenure). Employees stopped trusting deferred rewards, and law, consulting, and finance began paying young people market rates — producing "yuppies," a term novel enough in 1985 that a 25-year-old affording a BMW seemed remarkable. Measurement, once established, "spreads like heat" through an economy, even reshaping government and credentialing bodies themselves — market forces, Graham argues, achieve what credentials only approximate.

The most damaging thing school teaches isn't any subject — it's the habit of optimizing for grades instead of learning. Graham recalls that even as a diligent, genuinely interested student, nearly all his college work aimed at a good grade on something, and studying for tests, not the classes themselves, was when real work happened. This isn't laziness: nearly all tests imposed by an authority are badly hackable, because — like software nobody bothered to secure — their creators never made an effort to prevent hacking. Studying for a medieval history final, the smart move isn't reading the best books on the subject but mining lecture notes, assigned readings, and old exams for the sharply-defined chunks likely to be tested, including matching whatever political slant a professor holds. Getting a good grade and actually learning diverge so much that students must choose one, and rational ones choose grades, since graduate programs, employers, and even parents judge by grades alone.

College admissions is the extreme case: nominally a test of being "really smart," it actually measures whether an applicant suits the taste of admissions officers who simply "accept who they like" — an inherently hackable, subjective target. That's why whole industries (test-prep firms, admissions counselors, private schools) exist to hack it, and why ambitious teenagers' lives get warped into artificial memorization, forced extracurriculars, and essays aimed at an unstated target.

The deeper damage surfaced only decades later, at Y Combinator. Young founders kept overcomplicating everything — asking what "trick" would make VCs invest, how to engineer a launch for "exposure," why Tuesday is the best launch day — rather than grasping that the way to get funded is to actually be a good investment, and the way to do that is to build something good enough that users recommend it to friends, driving exponential growth. It took Graham years to see why: these founders had been trained that winning means hacking the test, and, never having faced a non-artificial test before, assumed fundraising numbers were simply the new test to game. He realizes he'd unconsciously undone the same training in himself — his lifelong aversion to big companies and pull toward startups, which he'd attributed to bureaucracy or "yuck" factor, was really an avoidance of hackable-test environments.

He proposes a rule for spotting bad tests: those not imposed by any authority (a football match) are inherently unhackable, since no one claims they measure more than the result itself; those imposed by authorities (grades, admissions) are proxies for something else and must be deliberately engineered to be unhackable, which rarely happens — so bad tests are roughly synonymous with authority-imposed ones. This used to be the price of getting rich, back when the mid-twentieth-century economy ran on oligopolies where climbing meant playing their internal game; today you can get rich directly by doing good work, part of why people are newly excited about wealth. As the link between work and authority erodes generally — in startups, and in writing, where authors now reach readers without publishers or editors — hacking bad tests should matter less, starving those fields of talent while fields rewarding good work draw the most ambitious people, eventually forcing education to stop training the reflex.

Footnotes add: Lambda School already runs pass/fail with no grades, using tests only to gate progression; YC's founder-selection is unhackable-by-authority because a tight feedback loop (bad picks show within a year) corrects bias, unlike admissions officers who are, Graham says, "looking in a mirror" at applicants' manufactured personalities; and a note to "tiger parents" that training kids to win by hacking bad tests trains them to fight the last war.

Variation in income is the same pattern found in every specialized skill—chess, painting, novel-writing—and in a modern democracy it signals economic health, not injustice. Three beliefs make people treat income variation as different: the childhood "Daddy Model" of wealth (mistaken), the disreputable way most fortunes were made (outdated), and the fear that it harms society (empirically false). Children see wealth as parent-issued and fixed, so believe it should be divided equally—a mindset some adults never outgrow. Wealth is really the goods and services people make; money merely trades it, so income tracks the wealth people generate for others. A US CEO earns roughly 100 times the average worker (S&P 500 median CEO pay $3.65 million in 2002 vs. a $35,560 mean US wage); NBA players earn about 128 times average ($4.54 million), baseball players 72 times ($2.56 million)—unremarkable next to ancient Rome, where slave prices varied 50-fold by skill. A basketball team wouldn't trade a star for 100 random people, nor could 100 average people have designed Apple's next product as well as Steve Jobs: skill doesn't scale linearly. Calling underpaid work "unjust" just means people want the wrong things—lamentable, not unjust.

Great fortunes historically came from theft: cattle raiding, William the Conqueror's 1060s confiscation of Anglo-Saxon estates, Henry VIII's 1530s seizure of monastic property, or routine taxation. That changed as medieval Europe's independent middle class rose—townsmen in Genoa and Pisa who, unlike serfs, kept what they created and so had reason to create more. Wealth creation didn't overtake corruption until the Industrial Revolution: seventeenth-century England still resembled a corrupt developing economy; by the nineteenth century the archetypal rich man was an industrialist, not a courtier. Balzac's line "behind every great fortune there is a crime" is a misquotation—he meant only unexplained fortunes were probably criminal, true of much of today's world but not of his own novelist's income. Most countries still run on the theft model, so rich countries with widening gaps are wrongly seen as sliding toward it, when they're actually a step further away.

Technology amplifies productivity differences—a farmer with a tractor plows six times as much land as one with horses—and its leverage grows exponentially, so income variation should keep widening. Graham traces this in his own life: mowing lawns in high school, then buying a computer in 1985 and immediately earning as a freelance programmer—a job that hadn't existed years before. Yet technology narrows most other gaps: mass production means rich and poor now drive, dress, and furnish homes alike, since custom goods are inconvenient while mass-produced ones are cheaper, often better—a quartz Timex, accurate to half a second a day, beats a $220,000 Patek Philippe mechanical watch rated at -1.5 to +2 seconds. Brand is the one thing technology can't cheapen, hence its outsized attention. Rich and poor converge materially and socially even as incomes diverge—Lenin touring Yahoo or Intel would think communism won, until checking the bank accounts.

If income doesn't keep pace with technology's growing variance in productivity, only three explanations exist: innovation has stopped, the most capable people aren't creating wealth, or they're creating it unpaid—the first two are bad, and the third works only for fun, not for the unglamorous 90% of real work. Wealth creation switches on and off like a fan with the ability to keep what's earned: on in Northern Italy by 1100, off under feudal France, on in 1800s England, off under Britain's 98% investment tax in 1974, on again in the US, and on in West vs. off in East Germany. Suppressing income variation, by confiscation or taxation, leaves society poorer. Absolute poverty, not relative poverty, is what to avoid—better poorest in a richer unequal society than richest in a poorer equal one. Rich people matter not through trickle-down spending but through what earning their fortunes forces them to build: Henry Ford's wealth came from making you a tractor, not hiring you as a waiter.

The best way to get rich is to start or join a startup: it compresses forty years of low-intensity work into three or four years at maximum intensity. Graham's math: a good hacker in his twenties might earn $80,000 at a big company. Doubling the hours, tripling output-per-hour by focusing, doubling that again by escaping a middle manager's drag, and tripling once more by being smarter than the job requires yields a 36x multiplier — about $3 million of work a year. He doesn't defend the exact figure, only the structure: probably above 10, rarely above 100. This isn't a path to Bill-Gates-level billions — Microsoft's fortune rests on IBM's blunder of granting a non-exclusive DOS license, i.e. luck — but a way to earn millions through wealth creation, not speculation, marriage, or fraud.

Wealth is not money — it's stuff people want (food, cars, houses); money is the medium invented for trading wealth once specialization made barter impractical. This kills the "pie fallacy," the belief that wealth is fixed so one person's gain is another's loss: restoring an old car makes you richer without making anyone poorer. Programmers are the modern craftsmen who see this most clearly, creating finished, valuable products directly. At Viaweb, one exceptional programmer added several hundred thousand dollars of market value in a single day, while a mediocre one generated zero or negative value via bugs — supporting his claim that the top 5% of programmers write 99% of good software. Wealth can be given away too, as with penicillin or FreeBSD.

Getting rich requires measurement (legible performance) and leverage (decisions with outsized effect) — piecework has measurement without leverage; actors and CEOs have both, with real downside if they fail. Startups supply measurement through smallness: a ten-person company puts you within a factor of ten of measuring individual contribution, unlike a thousand-rower "giant galley" where no one's effort shows. Steve Jobs said a startup's fate depends on its first ten employees (Graham says five): small groups can be select, an all-star team rather than a village. Leverage comes from technology — technique whose value multiplies across every user, the way woven cloth made Florence rich in 1200 and shipbuilding made the Dutch rich in 1600. Big companies develop technology too slowly and cede fast-moving fields to startups; Viaweb's rule was "run upstairs" — choose the harder feature, since difficulty is the barrier to entry VCs look for. Patents offer weak protection — Philo Farnsworth invented television, but RCA made the money after years of litigation — so the real defense is being too hard to duplicate.

Two catches temper the deal: competitors, not you, set the effort level, forcing everyone to work maximally hard; and the payoff is only proportionate on average — the mean return might be 30x but the median is near zero, since most startups fail outright (Viaweb nearly died several times before selling near the top of its cycle). Startups, like mosquitos, are all-or-nothing. Graham favors getting acquired once cruising altitude is reached, to diversify and because acquirers judge you by user count, not independent technology assessment — sound, since users are the real proof of created wealth. Treat a startup as an optimization problem scored by users, shipping v1.0 fast rather than guessing what's valuable.

Historically wealth mostly came from conquest, inheritance, or theft, until the rule of law let medieval merchants keep fortunes safe from feudal lords — enabling the Industrial Revolution. Societies that later suppressed wealth accumulation — the Soviet Union, or Britain under 1960s-70s Labour governments — saw innovation stall, illustrated by a Soviet mathematician's stealth-plane theory left unrealized without a computer industry to run the calculations. The same recipe — measurement via small groups, leverage via new technique — has produced wealth since Florence in 1200; Graham argues it also explains national power: let entrepreneurs keep their earnings, and you rule the world.

Reducing economic inequality is mathematically identical to taking money from the rich, and doing so necessarily destroys the incentive to take entrepreneurial risk — which means it necessarily destroys startups.

Graham builds this chain step by step. Giving money to the poor and taking it from the rich are the same act, since the money has to come from somewhere. Raising the poor through education (turning checkout clerks into engineers) is a great way to raise living standards, but the last 200 years show it doesn't shrink the gap, because it makes the rich richer too — Henry Ford couldn't have gotten rich building cars for subsistence farmers. So genuinely compressing the gap requires pushing down on the top, via taxation or price limits on what top earners can charge. But risk and reward must stay proportionate: a bet with a 10% chance of paying off has to pay more than one with 50% odds, or no one takes it. Cap the rewards, and you cap people's willingness to take risks — including the risk of starting a company. Startups fail constantly (Graham estimates roughly 1 in 10 succeeds, defined as an IPO or an above-valuation acquisition); his own startup paid its first investors a 36x return, meaning it was rational to fund them at only 1-in-24 odds. Take away that payoff and venture investing stops making sense. Government can't substitute for it, because bureaucrats are rewarded for choices defensible after failure, not for risky bets. And founders — who invest irreplaceable time and ideas, not just money — will simply choose safer paths (Graham says he'd have sought a tenured research job) if the upside is capped. Since startups drive most new technology and jobs, and possibly the riskiest ones are the most valuable, throttling risk appetite proportionately throttles growth and could kill the best startups first. A country that falls behind economically becomes dependent on others for technology — a fate escapable only through isolation, which requires a police state.

Graham concludes that critics of inequality aren't really objecting to wealth but to wealth translating into power — construction firms buying government contracts, rich parents buying college admission (36% of Princeton's class of 2007 came from prep schools, versus 1.7% of American kids who attend them). He cites Vanderbilt's grandson Reggie killing two pedestrians in the 1920s versus Ted Kennedy's one at Chappaquiddick in 1969 as evidence that what changed over time was accountability, not wealth variation. His remedy: attack corruption directly through transparency and logging, rather than attacking wealth and destroying risk-taking as collateral damage.

Economic inequality is not one phenomenon but the sum of many causes, some bad and some good, and conflating them prevents fixing the bad ones. Graham raises the puzzle personally: as a Y Combinator co-founder, helping startup founders get rich technically increases inequality, yet that doesn't seem wrong — which shows inequality itself isn't the real target.

The most common error is the "pie fallacy": assuming the rich get rich by taking from the poor, a zero-sum view Graham traces partly to childhood experience and partly to writers like Joseph Stiglitz (The Price of Inequality, 2012), who describe the rich "grabbing" a larger share. In reality a woodworker who sells a chair creates wealth through voluntary exchange, unlike a high-frequency trader who profits only when a counterparty loses. Describing inequality as a ratio between income quantiles reinforces the fallacy by implying money literally moves between groups; instead, understanding it requires tracing individual people and asking what they would have done in an earlier era. Applying this "would-have" method to startup founders shows that in 1960 they would have joined big companies or become professors — Zuckerberg's fallback was Microsoft — so today's greater wealth reflects technology making fast-growing startups easier to build, not a political shift since the Reagan era.

Variation in productivity, driven by accelerating technology, is the "irreducible core" of inequality: even after eliminating fraud and rent-seeking, it remains and grows, surrounded by a "Baumol penumbra" of people paid enough to keep them from striking out on their own. Suppressing it domestically without eliminating startups is impossible, and driven-to-get-rich people would simply pursue wealth through startups or emigrate — Richard Florida recounts Europeans wanting Silicon-Valley-style entrepreneurship while denying it implies more inequality. The mid-20th-century compression of inequality that Louis Brandeis's warning ("We may have democracy, or we may have wealth concentrated in the hands of a few, but we can't have both") reflects was, per Graham's earlier essay "The Refragmentation," a historical anomaly, not the norm; Graham bets on the exponential curve of technological progress over Brandeis. He argues concentrated Silicon Valley wealth isn't destroying democracy, partly because getting rich today doesn't require buying politicians as Gilded Age tycoons did.

Because "you make what you measure," aiming policy at inequality itself will miss the actual problems it symptomizes — poverty and blocked social mobility — since inequality and poverty are distinct (someone facing a water shutoff over unpaid bills, as in Detroit, is hurt regardless of Larry Page's net worth, and Page's own path from a modest, not poor, background shows inequality per se isn't what blocks mobility). Graham's conclusion: attack poverty directly, and stop people getting rich through deception or lobbied loopholes because it's theft — not because it worsens inequality.

Fragmentation -- political polarization, cultural divergence, the creative class clustering in a few cities, rising economic inequality -- is one phenomenon, caused not by a new divisive force but by erosion of two mid-20th-century forces that had pushed Americans together: total war and the rise of giant corporations, a one-time combination unlikely to recur. World War II flattened incomes: the National War Labor Board fixed wages from 1942-45, military pay scaled by rank, wartime profit increases above prewar levels were taxed at 85%, and top individual income at 93% (FDR vowed "not a single war millionaire"). Socially, over 16 million Americans served -- roughly 80% of men born in the early 1920s -- under uniform conditions; effects outlasted the war -- federal power, taxation, and conscription stayed high through the Cold War -- and the GI Bill sent 2.2 million veterans to college.

The second force was the national corporation. Beginning with J.P. Morgan-era consolidation, thousands of founder-run firms merged into a few hundred giant, professionally managed ones; Rockefeller declared in 1880 that "the day of combination is here to stay," and by 1945 most sectors were oligopolies or government-backed cartels. This produced a low-resolution "Duplo economy": consumers had only two or three choices of everything -- three TV networks broadcasting identical programs nationwide, interchangeable cars and "red delicious" apples -- and firms like IBM enforced one model of dress and behavior the middle class imitated. Oligopoly compressed incomes from both directions: unions, protected as labor monopolies since the 1914 Clayton Act, extracted above-market wages companies passed to captive customers, while top executives were paid below market -- a 1952 study found three-quarters of the 800 highest-paid executives had over 20 years' tenure -- because pay was illiquid and there was no real market to price against. Starting one's own company wasn't seen as ambitious; climbing an existing hierarchy was the prestigious path. College enrollment rose from about 2% of the population in 1900 to 25% by 2000, making "work for Henry Ford" acceptable and "be Henry Ford" not.

Starting in the 1970s the Duplo economy fragmented on every front. Vertically integrated firms like Ford's 100,000-person River Rouge plant gave way to supply chains, because computers cut the coordination costs that had justified doing everything in-house. Technology beat economies of scale, markets globalized and changed faster, and government shifted from protecting oligopolies to dismantling them: Carter-era "deregulation" was really de-oligopolization, which cut prices in air travel and phone service. Apple pioneered microcomputers; IBM entered without crushing it; Microsoft supplanted IBM via a non-exclusive DOS license, owning the PC standard. A 1980s wave of hostile takeovers -- enabled by court rulings, Reagan-era sympathy, and Michael Milken's junk bonds -- broke conglomerates into more valuable pieces. Average S&P 500 tenure fell from 61 years in 1958 to 18 by 2012.

As companies shrank and job-switching became normal, salaries moved toward market price and, since productivity varies enormously, diverged sharply -- visible in the early-1980s coinage "yuppie" for young lawyers, financiers, and consultants demanding market pay now instead of paying career dues, a norm that spread economy-wide and eventually lifted CEO pay too, partly from prestige competition with athletes and startup founders. The same logic cut the other way for unionized labor, whose wages fell toward lower market levels as oligopoly protection eroded, worse where automation cut demand. Social fragmentation tracked the economic kind: diverging diets, cars, dress, religious practice. Graham argues this cannot be reversed by policy tweaks, since war and a one-time consolidation phase can't be reproduced, and technology's ever-lengthening "lever" on individual output predates and outlasts policy; mid-century conformity only masked rising variation in wealth-creating capacity. Combined with the Baumol Effect -- wealth-creators drag up peers' pay, as when Google overpays to keep talent from startups -- this makes inequality a durable reversion to the mean, so the better response is to mitigate fragmentation's consequences rather than chase its elimination.

Becoming a billionaire by starting a company and exploiting people are nearly opposite skills, and the proof lies in what Y Combinator actually screens for in its interviews. As a professional "billionaire scout," Graham argues that if exploitation were the key trait, YC partners would recruit for it the way NFL scouts recruit for speed — instead YC's whole motto is "Make something people want."

A startup can't force unwanted products on customers the way a big company can, so it must delight users to survive. Since a market economy already satisfies known needs, a founder must point to something new — a new need or a new way to meet one — and, critically, something uncertain (certainty would already show up as revenue). In a ten-minute interview, partners act as professional guessers trying to determine whether there's a path to a huge market. Often the path runs through what Graham calls a "larval market": small now, but growable, as Apple's 1976 market for home computers was. The ideal founders are "living in the future" and building what they and their peers already want, as Wozniak, Zuckerberg, and Larry and Sergey did. The initial market just has to exist — some users willing to use a buggy product now — from which growth techniques follow.

The most convincing interview answer is "because we and our friends want it, and it's already spreading." Airbnb didn't meet that bar — they'd built something they wanted but it wasn't yet growing — but they showed such deep first-hand knowledge of hosts and guests that partners funded them anyway; growth arrived about three weeks into the batch. Because interviews are short, partners need "random access" via direct questions, not a rehearsed pitch — Graham calls turning an interview into a pitch the worst advice he's heard. Candor matters more than polish: claiming an idea has no flaws, or that competitors don't exist, reads as clueless or dishonest, since seed-stage ideas are inherently risky bets.

Once a market looks plausible, the second question is whether these particular founders can find it — their determination, resilience, and relationship. Airbnb's founders won over YC largely through their story of funding themselves with Obama- and McCain-themed cereal, evidence of resourcefulness rather than idea quality. Genuine interest in the problem — not money or coolness — is what makes founders keep working long after they could have sold out; that's what actually produces billionaires. YC does see exploitative founders, but rejects them: exploitation starts with cofounders, then users, and such companies never get big. Graham traces politicians' "billionaires exploit people" narrative to a legitimate discomfort with inequality distorted into a false claim, noting no correlation (if anything, an inverse one) between bad behavior and wealth — and warns this myth especially harms poor kids' sense of how success actually works: by making what users want.

People get rich today mainly by starting companies rather than inheriting wealth, a return to the historical norm. Comparing Forbes's 100 richest Americans in 1982 and 2020: in 1982, 60 had inherited (including 10 du Pont heirs); by 2020 only 27 had, even though inheritance taxes fell over the period. The real change is that more people are making new fortunes: of 2020's 73 new fortunes, 56 came from founder or early-employee equity (52 founders, 2 early employees, 2 founders' wives) and 17 from managing investment funds, a category absent from the 1982 top 100.

The businesses also changed. In 1982, 24 of 40 new fortunes came from oil or real estate; by 2020 that had shrunk to 2 and 4 respectively, while about 30 of 73 came from "tech" companies (8 of the top 10 fortunes). Tech companies win through better technology rather than dealmaking, unlike the 1982 oil and real estate magnates or the courtiers of 16th/17th-century European courts.

In fact, 1982, not 2020, was the historical anomaly: 84% of its top 100 got rich via inheritance, resource extraction, or real estate deals. An 1892 New York Herald Tribune list of America's 4,047 millionaires found only about 20% had inherited, and economist Hugh Rockoff traced many of the rest to the new technology of mass production — much like today. The anomaly arose because J. P. Morgan-era financiers consolidated thousands of firms into oligopolies; by World War II, per Michael Lind's Land of Promise, most sectors were cartels or oligopolies, making startups nonviable in 1960 — the corporate ladder was the only route up.

That order broke down from the 1970s via internal decay, Carter-era deregulation, and new technology (especially microelectronics) — pictured as ice cracking on a pond, letting startups punch through the middle instead of only around the edges. Since then, starting a company has kept getting cheaper and companies have kept growing faster: IBM took 45 years to reach a billion 2020 dollars in revenue, Hewlett-Packard 25, Microsoft 13, versus 7-8 years now. Cheaper starts, better terms from investors who increasingly need founders more than founders need them, and faster growth compound to make founders' equity worth more, sooner — explaining the rising Gini coefficient without invoking any political "right turn" under Reagan.

Taste is not mere personal preference but a real, improvable skill: designers get better at their craft over time, which means earlier taste was objectively worse, not just different. Children are taught otherwise ("you like your way, he likes his") to stop sibling squabbles, then contradicted when the same adults insist Leonardo is a great artist. If quality is real, it can be studied, and across fields — mathematicians; Thomas Kuhn on Copernicus's aesthetic objections to Ptolemaic equants; Kelly Johnson's Skunk Works belief that a beautiful airplane flies well; G.H. Hardy's dictum that ugly mathematics has no permanent place — the same design principles recur.

Good design is simple: shorter math proofs are better, architecture should rest on a few structural elements rather than ornament, writing should say what it means briefly. It's timeless: appealing to people in 1500 is a way of also appealing to people in 2500, since both stand outside present fashion. It solves the right problem: stove dials should match the burners, not line up for manufacturing convenience; mid-twentieth-century sans-serif type optimized for pure letterforms rather than the real goal, legibility (a Times Roman lowercase g is easier to tell from y). It's suggestive: Jane Austen's spare description lets readers build scenes themselves, and software should offer a few combinable primitives, like Lego, rather than dictate use.

It is often slightly funny — Durer's engravings, Saarinen's womb chair, the Pantheon, the original Porsche 911, and Godel's incompleteness theorem all carry an uncanny humor, tied to the confidence of not taking oneself too seriously. It's hard: constraints — a difficult site, a tiny budget — strip away the inessential, which is why "form follows function" (misquoting Sullivan) works only when function is demanding enough to leave no slack for error. It looks easy: an expert pianist plays faster than conscious thought allows, and Leonardo's spare line drawings look like a few strokes anyone could place — except each must be exactly right.

Good design uses symmetry and recursion, from sentence structure to the Eiffel Tower's tower-on-a-tower, though architects self-consciously abandoned it from the Victorian era through 1920s modernism. It resembles nature, which has had eons to solve the same problems — ribbed boat hulls echo ribcages, while early aircraft copying bird form failed for want of light engines (the Wright brothers' engine weighed 152 lbs, 12 hp). It is redesign: Leonardo's sketches show five or six attempts at a single line, the Porsche 911's signature rear only emerged from reworking an awkward prototype, and Wright inverted a ziggurat half of the Guggenheim's original plan.

Taste in copying matures in three stages — unconscious imitation, a conscious drive for originality, then selfless willingness to use whatever answer is already right; Raphael's mid-nineteenth-century dominance produced imitators of imitators, provoking the Pre-Raphaelites. Good design is often strange — Euler's Formula, Bruegel's *Hunters in the Snow*, the SR-71, and Lisp are uncanny, not just beautiful — and strangeness can't be aimed at directly; Einstein didn't try to make relativity strange, only true. It happens in chunks: fifteenth-century Florence simultaneously produced Brunelleschi, Ghiberti, Donatello, Masaccio, Filippo Lippi, Fra Angelico, Verrocchio, Botticelli, Leonardo, and Michelangelo, while comparably sized Milan produced no comparable names — community matters as much as innate ability (see also the Bauhaus, the Manhattan Project, the *New Yorker*, Skunk Works, Xerox PARC). Finally, it is often daring: Renaissance art was considered shockingly secular (Vasari reports Botticelli renouncing painting, Fra Bartolommeo and Lorenzo di Credi burning their own work), and Einstein's relativity wasn't fully accepted in France until the 1950s.

Great work generally starts from intolerance of ugliness rather than a positive vision of beauty — Giotto rejected formulaic Byzantine madonnas as wooden, Copernicus rejected the equant as an intolerable hack — but that instinct only works once deep field knowledge gives it a reliable nose for what needs fixing. The recipe: very exacting taste, plus the ability to gratify it.

Americans excel at software and movies while producing ugly cars and cities, and both outcomes stem from the same trait: impatience. The "just do it" ethos treats making something as a messy, fast process rather than a careful plan — and this works when the medium is malleable. Good software craftsmanship means working fast: a slow, meticulous build just yields a polished version of your initial, mistaken idea, so it's better to prototype quickly and let new ideas emerge. Movies work the same way, assembled fast despite imperfections, with hacks concealing seams.

Physical products don't forgive this. Cars and cities have real constraints, so winning requires taste and attention to detail rather than dramatic reinvention — and "taste" sounds pretentious or effeminate to American ears, so nobody defends it. American cars, like the AMC Matador with its excess sheet metal, get designed by marketing departments, not designers, then "improved" with tail fins or bigger bodies. McMansions are just larger flimsy boxes, not better-built ones. Japan, by contrast, has centuries of craftsmanship (1200-era swords still look implausibly fine), and Japanese executives would be horrified to ship a bad car. Software and film succeed in America because programmers and directors, not marketers, hold power — same logic, different mechanism than Japan's culture-wide design obsession. Detroit instead chases focus groups, losing share as buyers migrate to Lexus.

Apple is the exception: Steve Jobs's design obsession made the iPod beat Sony's players. Exurban sprawl is the worst case, since developers build piecemeal, so no market pressure forces good towns. Graham predicts national traits will give way to occupational ones — increasingly willful Japanese hackers, increasingly tasteful American products.

The way to avoid imitating the wrong things is to copy only what you genuinely like, not what merely impresses you. Graham traces three cases where he confused the two: in high school he imitated fashionable short stories about people suffering in complicated ways, though he didn't enjoy them; in college he imitated dense, jargon-laden philosophy journal papers that turned out to say nothing definite enough to ever be refuted; in grad school he was awed by "expert systems" built on inference engines, which professors wrote books about and startups sold for a year's salary, though the ideas were trivial and led nowhere.

He offers two tests for telling liking apart from being impressed. First, ignore presentation: ask what you'd pay for a painting found dirty and frameless at a garage sale, with no name attached. Second, notice guilty pleasures — what you read when you don't feel like being virtuous, rather than books like Ulysses that people read mainly to feel they're reading Ulysses.

Even genuine models carry a trap: copying their flaws instead of their merits, as when painters imitated the brown tone of Renaissance paintings, not realizing it was centuries of dirt later removed by cleaning.

Painting cured him: the art world's corruption forced him to see he had to judge quality himself — authorities can't be trusted on it.

Paul Graham argues that taste is not merely personal preference — there really is such a thing as good art, because art has an audience, and audiences share things in common. If nothing were better than anything else, an artist deciding whether to improve a canvas could just as well leave it blank; since that conclusion is absurd, some art must genuinely be better than other art.

The case for shared standards rests on human commonalities. Nearly all people find human faces engaging — babies recognize faces from birth — so a painting with faces will interest more people than one without. Preferences aren't random: humans (and possibly mathematicians debating proofs, per Paul Erdos's "God's book" of maximally elegant proofs) share reactions to primary colors, 3D representation, and edge-finding. Graham pictures tastes as concentric ripples — some appeal only to you and your friends, others to your age group, others to nearly all humans — and argues "good art" implicitly means art that would engage that widest ring, the set of all possible humans.

But you can't just take a vote to find the best art, because man-made objects, unlike apples or beaches, are subject to deliberate trickery and self-deception. Illustrators fake a look of speed to seem more skillful than they are. Brand overwhelms honest judgment: the Mona Lisa, a small dark painting mobbed behind glass in the Louvre, would draw a shrug from viewers seeing it unlabeled among other works, yet most people, primed by its fame, can't see it as a painting at all. Adults also fool themselves, forcing themselves to like what they think they're supposed to. A vote just measures these errors, the way a compass next to a magnet just measures the magnet.

Instead, Graham says, use yourself as a guinea pig and strip out the error sources. Tricks are the easier problem: cataloging them — the shiny airbrushed lettering that impressed him at ten, the SUV "make it look tough" school of design, avant-garde art aimed at intellectuals rather than made with real interest — makes you immune to them, the way professional magicians see through others' illusions. Circumstantial bias is harder to escape, but traveling widely across cultures and eras helps; a work that would appeal equally to your friends, to people in Nepal, and to the ancient Greeks is probably onto something.

Graham's real motive is not connoisseurship but permission for makers: art schools have retired the idea of "good," teaching students only to explore personal vision, and curators hide behind euphemisms like "significant" rather than judge work outright. He contrasts this with fifteenth-century Florence, where painters believed great work was achievable and competed fiercely to outdo each other — and were right, since they produced things like the Sistine Chapel ceiling. Since good art exists, so does good taste — the ability to recognize it, achieved by being hard to trick and not simply liking whatever you grew up with — and ambitious artists today should reject the claim that trying to make great work is naive.

Paul Graham's rule for making new things is to find (a) simple solutions (b) to overlooked problems (c) that genuinely need solving, (d) deliver them as informally as possible, (e) starting with a crude version 1, then (f) iterate rapidly -- and this formula reliably provokes initial contempt. He traces the pattern across four projects: Viaweb (1995), dismissed by VCs because it ran on the server rather than as a "real" Windows app and didn't even process credit card transactions its first year, yet it crushed its competitors; Y Combinator, seen as inconsequential next to million-dollar series-A rounds, now widely imitated; his essays, initially met with "who is this guy?" skepticism; and Arc, criticized as flimsy after years of work. Simple solutions look less impressive than complex ones, overlooked problems are by definition dismissed by others, informality forces people to judge substance over presentation, and a crude v1 always looks incomplete -- but these same traits are the advantage: less competition, no wasted effort or self-deception from window dressing, and, per Feynman, a released v1 benefits from "the imagination of nature." He cites Cezanne and Klee as painters using this method, and Reddit, whose minimal design looked like no design but solved the real problem of surfacing what's new.

Brand is what's left once the substantive differences between products disappear — and technology naturally erases those differences, as Swiss watchmaking shows. In the early 1970s a triple disaster hit it: Japanese makers, having swept the 1968 Geneva Observatory trials, could now build watches both cheaper and better; the 1973 collapse of Bretton Woods sent the Franc from .228 to .625 USD by 1978, making Swiss watches 2.7 times pricier for Americans; and quartz movements made accuracy, once expensive, a commodity. Unit sales fell nearly two-thirds over the decade and most watchmakers went insolvent. Survivors became luxury brands rather than precision-instrument makers — revenue flattened through the "quartz crisis," then rocketed upward from the late 1980s.

In the 1945–1970 "golden age," makers competed on thinness and accuracy, a genuine engineering tradeoff; complications (moonphase, chimes) were a sideshow. The "holy trinity" — Patek Philippe, Vacheron Constantin, Audemars Piguet — stood on both prestige and performance; once quartz beat everyone on both, only prestige remained. Omega kept fighting on performance, releasing a 45%-higher-frequency movement in 1968 that proved too fragile, wrecking its reputation for reliability; it was insolvent by 1981. Patek instead began designing its own cases (previously outsourced), launching the Golden Ellipse in 1968 and expanding the maker's name from roughly 8 to 800 square millimeters of visible surface. In general, branding is centrifugal, design centripetal; the two coexist only in vast design spaces (painting) or unclaimed territory — not in a mature field like watches, so branding could only advance at design's expense.

Patek's 1968 Ellipse ads dropped any mention of accuracy and sold scarcity instead ("only 43 watches are signed out each day"). Audemars Piguet's 1972 Royal Oak — designed by Gérald Genta, in steel — went further: "steel at the price of gold," priced "from $35,000," its case integrated into the bracelet so the whole wrist proclaimed brand. Patek's 1976 Nautilus (also by Genta, 42mm versus the golden age's 32–33mm) was too extreme for its time but is now Patek's most sought-after model. The 1984 hobnail Calatrava, engineered by ad man René Bittel, became the 3919, the "banker's watch" beloved by 1980s New York bankers; by 1987 sales were climbing for good. Mechanical watches suited wealth display uniquely well — visible on the wrist, more legitimate than gold chains — and accurate enough (about 5 seconds a day, versus quartz's fraction of a second) that quality became a threshold protecting brand reputation, not a selling point. Women mostly skipped the shift, since they could wear actual jewelry instead.

Since 1985, the "brand age," most storied names survive only as tiers within six holding companies; just Patek, Audemars Piguet, and Rolex remain independent. Watches have generally swollen and sprouted brand oddities. Rolex needed no reinvention: it had abandoned serious watchmaking research around 1960 (patents fell from 16.6/year in the 1950s to 1.7/year in the 1960s) for a recognizable, oversized waterproof case — a "luxury Jeep" logic later echoed by SUVs like the Porsche Cayenne. The Nautilus epitomizes the age's extremes: years-long loyalty purchases and waitlists, plus secondary-market policing — Patek buys back hundreds of its own watches yearly to trace resellers and cuts off offending retailers — sustaining a managed asset bubble via the "comb-over effect," where small compounding choices add up to something freakishly wrong.

Brand-age oddities exist because there's no function for form to follow: brand answers to some of the worst features of human psychology, guaranteeing something strange and ugly. The prescription: avoid brand, buying or selling it — chasing it is a bad problem to work on. More broadly, fields cycle through golden and lesser ages beyond anyone's control; a golden age is recognized only in hindsight, felt then merely as smart people working on interesting problems. The way in isn't to chase the label but to follow interesting problems — historically how participants found themselves in what later got called a golden age.

Startups form in geographic clusters — thick in Silicon Valley and Boston, thin in Chicago or Miami — and though the recipe (a great university near a town smart people like) can work anywhere, America is an unusually humid climate for that condensation: ten distinct advantages, most fixable by a rival, explain why.

First, the US allows immigration; half of Silicon Valley speaks with an accent, and Japan's aversion to immigrants likely dooms any Japanese equivalent. Second, America is rich: poor countries lack infrastructure founders take for granted (a friend of Graham's broke her ankle on uneven railway steps in India), and there may be a generational speed limit on how fast an economy's attitudes can change. Third, the US is not yet a police state — China's censorship and Singapore's chewing-gum bans suppress the "odd ideas about politics" that tend to accompany odd ideas about technology, though Graham notes the US itself has lost civil liberties recently.

Fourth, American universities are better: Germany's policy of keeping all universities equal produced none that stand out, and Germany's expulsion of Jews in the 1930s may explain why it never recovered a top-tier university. Fifth, you can fire people in America — rigid European labor law removes the strong inverse correlation, visible across actors, professors, and athletes, between job security and performance. Sixth, American work is less identified with employment; even Steve Wozniak initially refused to quit HP after Apple's founding, showing how strong the old "job for life" model still is. Seventh, America is not too fussy about regulation — HP, Apple, and Google all began in garages, while a friend of Graham's needed $20,000 in capital just to incorporate a company in Germany. Eighth, America's 300-million-person domestic market lets startups launch locally before internationalizing, unlike Sweden or the multilingual EU (Skype, notably, tackled an intrinsically international problem). Ninth, America has venture funding, especially angel money from founders who cashed out earlier — Google's seed came from Sun co-founder Andy Bechtolsheim — a self-reinforcing but slow-to-bootstrap cycle. Tenth, America has "dynamic typing" for careers: students choose majors and PhDs later and more loosely than in Europe, so grad students, and even graduates of "lousy" public high schools, stay open to founding companies instead of committing early to a fixed occupation.

Notably absent from the list is national character: Graham argues Americans aren't intrinsically more entrepreneurial — Indians and Chinese seem equally so — and that Europe's problem is a lack of visible examples (Stanford students out-found Yale students for this reason) rather than a lack of nerve. European ambition was likely discredited by the disasters ambitious people caused in the early twentieth century, and should reassert itself over time.

Graham then sketches how a rival could beat Silicon Valley outright. Geographically, Palo Alto sits thirty-plus miles from San Francisco, forcing a choice between boring sprawl and a long commute, and public transit is decent by American standards but "third world" by Japanese or European ones — so a deliberately dense, walkable, well-connected town could out-compete it. Nationally, a rival could set capital gains taxes near zero, as Belgium already does, which matters more than income tax since assets, not people, can relocate to chase a better rate. Most of all, a rival could fix immigration: the H1B system requires a college degree (excluding the likes of Steve Jobs and Bill Gates) and ties a visa to employment by someone else rather than to founding one's own company, so a country that simply let in the smart people America turns away could capture over half the world's top talent overnight. None of this requires sacrifice, Graham concludes — better universities, livable cities, civil liberties, flexible labor law, open immigration, and low capital-gains taxes are good in themselves — and a country that gets there first stands to gain enormously as startups become the default career path for the ambitious.

Silicon Valley could be reproduced anywhere, because unlike old trading cities that depended on waterways, a tech hub today depends only on getting the right people to move there — Graham estimates ten thousand would suffice to turn Buffalo into a silicon valley (a footnote suggests as few as thirty, hand-picked).

Only two types of people are the limiting reagents: rich people and nerds — the only ones present at a startup's founding, since everyone else will follow. Miami has rich people but few nerds; Pittsburgh and Ithaca (home to Carnegie Mellon and Cornell, both top CS departments) have nerds but no rich people willing to live there, thanks to bad weather and no compensating old-city charm. Government can't substitute for the rich people: bureaucrats lack the experience, personal stake, and competitive pressure that make startup investors effective — even corporate VC arms usually aren't allowed to lead their own deals. Nor is it about buildings: "technology parks" like Sophia Antipolis (home to Cisco, Compaq, IBM, NCR, Nortel — not startups) miss that a company stays wherever it was sitting around a kitchen table when it got funded, long before it needs office space.

What you actually need is a university good enough to be a magnet — one of the world's best. Graham argues this could be bootstrapped almost overnight: pay 200 top researchers $3 million hiring bonuses (roughly half a billion dollars) and the resulting faculty quality becomes self-sustaining, since professors choose jobs based on their colleagues.

The university must sit in a town with "personality" — dense, varied, built one building at a time, with local shops rather than chains — which argues for banning large development projects, including government-led ones. Nerds specifically want a town where people "walk around smiling": not LA (nobody walks) or New York (walks, but joylessly, chasing glamour); their model is Berkeley or Boulder — cafes, bookshops, hiking. The town must also feel young, ruling out reviving declining industrial cities like Detroit or Philadelphia, and needs an intact center — no startup hub has a dead downtown. This tolerance for oddity is why tech cities are reliably liberal: liberal places tolerate the odd ideas smart people have. The 2004 election county map is offered as confirming evidence that "traditional values" counties never become startup hubs. Boulder and Portland are named as closest to qualifying, needing only a great university.

Growth after that is organic and slow, illustrated by lineage: William Shockley moved to Palo Alto in 1956; his "traitorous eight" left in 1957 to found Fairchild Semiconductor, spawning Intel (Gordon Moore, Robert Noyce) and Kleiner Perkins (Eugene Kleiner), which in 1999 funded Google via John Doerr. Because startups beget startups, hubs can't be partial — Chicago, America's third-largest metro, produces negligibly fewer startups than 15th-ranked Seattle. Competing against Silicon Valley now is harder because venture firms cluster and pull startups in via acquisitions, but founders hold ultimate power, and Silicon Valley's own weakness — its "soul-crushing suburban sprawl" — leaves an opening for a rival that avoids becoming one giant parking lot.

Great cities send an implicit message about what kind of ambition matters, and that message shapes what the people living there actually achieve. New York's message is "you should be richer"; Cambridge (Massachusetts) says "you should be smarter"; Silicon Valley says "you should be more powerful," meaning: how much effect do you have on the world (why Larry and Sergey are admired for controlling Google, not merely for wealth).

This matters more than intuition suggests. Graham cites the "Milanese Leonardo" problem: nearly every renowned fifteenth-century Italian painter came from Florence, though Milan was equally large, so someone in Milan must have had Leonardo-level talent and never developed it — proof that environment can overpower innate ability. This is why Graham has deliberately chosen where to live. He'd expected Berkeley to be "Cambridge with good weather," but its actual message is "live better" — pleasant, civilized, but not humming with ambition, because comfort filters for people who prioritize comfort. Cambridge, by contrast, is expensive, grubby, and bad-weathered, which selects for people who'll tolerate that to be near smart people. He argues Cambridge is currently the world's intellectual capital because Harvard and MIT sit almost adjacent, surrounded by ~20 other colleges, unlike the Bay Area's two great but geographically separated universities, or New York's smart people "diluted by a much larger number of neanderthals in suits."

Cities transmit their message through low-level, hard-to-block signals — overheard conversations, glimpses through windows at dusk (bookshelves in Cambridge versus the blue glow of TVs in Palo Alto). Because encouragement and discouragement are asymmetric (like loss aversion with money), and because admiration is zero-sum, each city tends to concentrate on one type of ambition, making rivals feel second-class — which is why Graham doubts New York can become a real startup hub. As partial evidence of New York losing ground to Silicon Valley on its own terms, he notes the ratio of New York to California residents in the Forbes 400 fell from 1.45 (81:56) in 1982 to 0.83 (73:88) in 2007. LA's message is fame (an "A List"); DC's is who-you-know insider status; San Francisco currently shares Berkeley's "live better" message but could shift if enough good startups relocate there; Paris's is style and art appreciation; London still faintly transmits aristocracy.

Graham's full inventory of city messages: wealth, style, hipness, physical attractiveness, fame, political power, economic power, intelligence, social class, and quality of life — noting attractiveness-for-men and hipness are recent additions, while social class is fading as economic power (increasingly about directing technology, not just controlling resources) supplants it upstream.

Not everyone needs a great city: fields like math and physics, where peer judgment is reliable and audience doesn't matter, just need a good department (even in Los Alamos). Chaotic fields — art, writing, technology — need a city's larger funnel of peers and audience. The critical years are early-to-mid career, not childhood or college; the Impressionists, born and died all over France, were defined by their years together in Paris. Graham's advice: since most ambitious people don't yet know what they're ambitious about, try living in several cities when young to find where the message resonates.

Paul Graham argues that startups may be a new economic phase, since founders and early employees are far more productive than at established companies (think Larry Page and Sergey Brin), and also a business type—like movies—that flourishes only in specialized hubs such as Silicon Valley. If both are true, this revolution will be unusually localized: unlike agriculture, cities, and industrialization, which all spread widely, startup culture is spreading slower than the Industrial Revolution despite faster communication.

Industrialization spread because steam engines were portable: once Boulton and Watt built one, any manufacturer in a stable economy could import the technique, and geography fixed factory locations. Startups don't transfer that way—they're a social phenomenon requiring a community of expertise, as in film, and no market forces a country to build "a Microsoft of France."

Existing hubs arose from accidents, not policy: Shockley's return to Palo Alto seeded Silicon Valley, and Gates and Allen's homecoming built Seattle. Deliberately growing one—pairing a great research university with a place the rich want to live—worked once, but a new attempt now would lose its best startups to existing hubs. Graham's proposal to pay startups to relocate, sparking a chain reaction, he admits is impractical, offered only as a thought experiment. His prediction: startups will keep spreading, slowly, via the same random factors, increasingly overwhelmed by the pull of established hubs.

Startups don't get killed by most towns; failure is the default everywhere, and only a few places, Silicon Valley chief among them, supply an antidote that saves some. That antidote has two components, both driven by the sheer density of startup people nearby: an environment where founding a company feels normal, and chance meetings with people who can help.

The environment matters most in the transition from wanting to start a company to actually doing it. Elsewhere founders get treated as unemployed; in the Valley people pay attention rather than defaulting to skepticism, having seen too many unpromising founders become billionaires. Graham notes he himself once nearly reconsidered starting another company after a VC's eager reaction. This social pull also draws in people unsuited to running startups, but since suitability is hard to predict beforehand, that's an acceptable cost; trying is the best test.

Chance meetings matter across both the starting and succeeding transitions, compensating for the disasters that hit startups everywhere; in the Valley "lightning has a sign bit." Example: a college-startup founder who moved to Palo Alto for the summer ran into Sean Parker on the street; Parker understood the market, knew investors, and championed founder control, reshaping Facebook in 2004. Like ideas surfacing after sleep, chance meetings work through acquaintance "drifting" just the right amount, as when Larry Page met Sergey Brin. It's the people, not the Valley's physical infrastructure or weather, that now sustain the reaction; people there help each other with no expectation of return.

Density matters for three reasons: it's needed for chance meetings to happen at all, it raises the odds a hub contains whoever a startup needs most, and enough people redirect social norms toward ambition rather than the mean.

Wisdom and intelligence are not different domains or different sources (experience versus innate talent) but two different shapes of the same performance curve: if you graphed situations on the x-axis and how well someone handles each on the y-axis, a wise person's graph would be uniformly high — a high average — while a smart person's graph would have occasional high peaks. You judge intelligence at its best and wisdom by its average, the same way talent is judged at its best and character at its worst. This single distinction, Graham argues, explains both conventional theories people reach for: human problems are the most common type, so a high average depends heavily on handling people well (hence "wisdom = people skills"), and a high average is built mostly through experience while genuine peaks require rare innate qualities (hence "wisdom = experience, intelligence = innate"). It also implies there's no single thing called "wisdom" — the word just names a grab-bag of qualities (self-discipline, experience, empathy) that raise the average, the way "intelligence" names whatever raises the peaks.

Historically, the number of "points" on that curve — the range of situations we're tested in — keeps growing as knowledge specializes, so wisdom and intelligence, once nearly identical in Confucius's and Socrates's day, have been diverging like a digital image gaining pixels. Society has effectively voted for intelligence: we admire the genius, not the sage. The converse is unflattering — you can be wise without being smart, which gets you James Bond, competent in countless situations but dependent on Q for the math.

For Confucius and Socrates, wisdom, virtue, and happiness were inseparable: the wise man always chooses rightly and is therefore always content ("The superior man is always happy; the small man sad"). But a mathematician Graham once read about (possibly Andrew Wiles) said he usually went to bed discontented, feeling he hadn't progressed enough — and Graham finds the same restlessness in himself when writing essays, though not when merely advising people. The difference is that advising, like ancient administrative work (a peasant deciding whether to mend a garment, a king whether to invade), involves choosing the best of existing alternatives — bounded work with a knowable ceiling, guided by prudence. Inventing or writing has no ceiling; performance is unbounded, guided by inspiration rather than duty, so makers stay anxious the way Confucius's "small man" lived at the mercy of circumstance. A runner can be content winning gold even if she could've run faster; a novelist has no equivalent finish line.

The two also have opposite recipes: wisdom comes from cutting away childish bias — self-control and experience strip out idiosyncrasy, which is why the wise resemble each other while smart people are smart in distinctive ways. Intelligence comes from cultivating idiosyncrasy, following curiosity, sometimes aided by an inflated sense of one's own ability to keep working. Since most education uses wisdom-style recipes (submission, absorbing prescribed material), it's poorly suited to building intelligence, which needs a teacher who plays appreciative-but-hard-to-impress audience to a student's own inventions. Graham's closing reassurance: discontentment in unbounded, inventive work isn't a sign of failure — like a good runner, you get tired because you're running fast.

Philosophy went wrong because it lets words get pushed past the point where they hold precise meaning, and it can be fixed by redefining its goal: not the most general truths, but the most general *useful* truths. Everyday concepts are fuzzy and break under enough pressure — even "I." Graham traces this to a class taught by Sydney Shoemaker: since you could split a brain and transplant each half into a different body, there is no indivisible self, only a "collection of cells that lurches around" and calls itself I. Outside math, where terms are defined precisely, words work well enough in daily life but snap when stretched too far — Wittgenstein is popularly credited with noticing that most philosophical controversies are confusions over language, and Graham calls this "the central fact of philosophy": debates over free will or whether abstractions "exist" collapse once you ask what the words mean.

Historically, Socrates, Plato, and especially Aristotle turned philosophy toward analysis, likely spurred by progress in math, and Aristotle essentially invented logic and zoology. But they raced through genuinely new intellectual territory without realizing words break down at high resolution — "arguing about artifacts induced by sampling at too low a resolution." Aristotle's central error, in Book A of the *Metaphysics*, was conflating motive and result: because people who seek deep understanding are often driven by curiosity rather than practical need, he concluded the most noble knowledge must be the most useless, and deliberately aimed at uselessness — with no alarm bells to stop him getting lost in abstraction. The *Metaphysics* is consequently "a failed experiment," among the least-read famous books, valuable mainly for a few salvageable ideas. Because it became the map for later explorers, it sent them the wrong way too: instead of treating Plato's and Aristotle's works as superseded "version 1s," subsequent centuries revered them as fixed texts, and only around 1600 did anyone confidently treat Aristotle as a catalog of mistakes.

Unclear writing about big ideas creates what Graham calls a "singularity": it looks tantalizingly like real depth to ambitious but inexperienced readers, and because there is no way to prove a text meaningless (the closest anyone got was Alan Sokal's hoax paper in *Social Text*), critics of nonsense usually just quit the field rather than fight it, letting the nonsense self-perpetuate. Bertrand Russell complained in an 1912 letter that philosophy attracted people who "loved the big generalizations," repelling those with "exact minds," and responded by pointing Wittgenstein at the problem. Graham credits Wittgenstein not with discovering that prior philosophy was largely wasted effort — every sharp person who sampled philosophy suspected as much — but with staying inside the field to force a reckoning, "like Gorbachev." Since then, philosophers mostly analyze how language works, while literary critics have colonized the vacated metaphysical territory under labels like "critical theory," producing comparable word-salad (Graham quotes a Duns Scotus passage, changed only by substituting "gender" for "number," to show the genre is old).

Graham's proposed fix keeps Aristotle's goal — discovering the most general truths — but reaches it from the opposite direction: ask not "what are the most general truths" but "of the useful things we can say, which are most general," using the test of whether an idea makes a reader act differently. Model examples: the controlled experiment, evolution's application to genetic algorithms and product design, and Harry Frankfurt's lying-versus-bullshitting distinction. This won't win tenure, since it can't hedge into vagueness or trivial narrowness the way humanities scholarship does, but it's open to anyone — you can start from something as specific as "Joe's has good burritos" and gradually generalize. Graham closes by arguing philosophy is younger than its 2,500-year history suggests, having spent most of that time as commentary on Plato and Aristotle or entangled with religion, so it is "as young now as math was in 1500," with correspondingly more left to discover.

Life is genuinely short — not just a complaint — and having children proved it by turning time into countable discrete quantities: you get only 52 weekends with a two-year-old, and if Christmas-as-magic runs from ages 3 to 10, you watch it happen just 8 times. That scarcity gives real force to the sentence "life is too short for X," and for Graham the X is bullshit — unnecessary meetings, pointless disputes, bureaucracy, posturing, other people's mistakes, traffic, addictive but unrewarding pastimes. Bullshit enters life two ways: it's forced on you, since making money mostly means errands (the law of supply and demand ensures the more rewarding work is, the cheaper people will do it), or it tricks you, as with arguing online, where instinct says defend yourself even though the fight is stealing your life; technology keeps making distractions more addictive. Some people opt out of the conventional grind for more authenticity; freelancers and small companies can cut forced bullshit by firing toxic customers even at a cost to income. Beyond avoiding bullshit, seek what matters: a good test is whether you'll still care about something later, since fake importance spikes sharply but has little area under the curve — coffee with a friend matters, chasing status in middle school didn't. Because life's shortness ambushes you — Graham regrets not spending more time with his mother before she died — cultivate impatience about what you want to do rather than waiting, as James Salter's title Burning the Days captures. With something scarce, get more of it and savor what you have; how you live also affects how long you live.

Having kids, something Graham dreaded, turned out to be wonderful — and much of what he feared beforehand rested on bad data. Before fatherhood he saw parents as uncool and joyless, based on two skewed samples: he only noticed kids in bottlenecks like airplanes, where a toddler is at its worst, and he judged parenthood by his own unruly childhood, concluding it was mostly law enforcement. His mother had told him, around when he was 30, that she'd genuinely enjoyed raising him and his sister; he only understood this after having his own. Birth itself triggered a rapid chemical shift — driving his newborn son home, he found himself thinking, at a crosswalk, "every one of these people is someone's child." Kids also become real companions: interesting to talk to, and, surprisingly, fun to play with again after being unbearably repetitive (a footnote notes 6-year-olds see 2-year-olds as merely defective 6-year-olds, while adults appreciate their complexity).

Some fears were valid: kids make you work to their schedule and reduce ambition, since attention is zero-sum and kids displace "the top idea in your mind." Graham counters this by writing essays for his kids and using promises to them (like a trip to Africa after finishing his book Bel) as deadlines. He misses pre-kids freedoms like spontaneous travel, but admits he rarely used them anyway, paying for that freedom mainly in loneliness. Counting actual happy moments, not just potential ones, he finds far more after kids than before.

The clearest pattern across Paul Graham's working life is that his best moves came from chasing whatever wasn't yet prestigious, on the theory that low status signals unclaimed ground rather than a dead end. Before college he wrote bad plotless short stories and, at 13-14, programmed an IBM 1401 in Fortran via punch cards. A TRS-80 bought in 1980 let him write games, a model-rocket predictor, and a word processor his father used for a book. Planning to study philosophy, he found other fields had claimed all the interesting ground and switched to AI, drawn by Heinlein's novel "The Moon is a Harsh Mistress" and Terry Winograd's SHRDLU. At Cornell, with no AI classes, he taught himself Lisp and reverse-engineered SHRDLU for his thesis. Only Harvard admitted him to grad school, where he concluded that AI-as-practiced -- encoding sentences into formal symbolic structures -- was a hoax, capturing only a narrow subset of language, never real understanding. He pivoted to Lisp itself, drafting "On Lisp" (published 1993).

Dissatisfied that software inevitably becomes obsolete, he wanted to build things that last, and a 1988 visit to the Carnegie Institute convinced him paintings do. He audited Harvard art classes while still nominally pursuing a CS PhD, until in 1990 professor Cheatham asked if he could graduate that June; he wrote a dissertation on continuations in five weeks. He then studied at RISD and the Accademia di Belli Arti in Florence, where faculty and students had tacitly agreed to demand nothing of each other; he painted still lifes at night, arguing that emphasizing perceptual cues (edges, color shifts) can make a painting more "real" than a photograph in an information-theoretic sense. Broke, he took a job at Interleaf, a document-software company with a Lisp-based scripting layer, where he learned that "low end eats high end" -- cheap, unprestigious tools displace expensive ones -- plus lessons about bad meetings and dangerous bureaucratic customers. Back at RISD full-time in 1992, he found painting students chasing a marketable "signature style" (his example: Roy Lichtenstein) and dropped out in 1993 for New York, funding himself by writing "ANSI Common Lisp."

In 1995, inspired by the nascent web and a failed attempt with Robert Morris to put art galleries online, Graham realized ecommerce sites needed the same tech; the two built Viaweb, discovering that running the store-builder entirely on the server (a "web app") let merchants need just a browser. Funded by $10,000 from Idelle Weber's husband Julian for 10% -- the template later reused at Y Combinator -- and joined by Trevor Blackwell, Viaweb launched in January 1996 with 6 stores, grew to 500 by 1997 (7x/year), and was bought by Yahoo in 1998. Graham left in 1999, unproductive and burned out, with options then worth about $2 million a month.

A failed return to painting gave way, in 2000, to an abortive company (Aspra) for browser-based app hosting, which he shrank into the open-source Lisp dialect Arc. In 2001, an essay he posted online got 30,000 views via Slashdot, revealing that publishing no longer required editors' permission -- essays became a permanent third pursuit, later collected as "Hackers & Painters." Meeting Jessica Livingston in 2003 led, on March 11, 2005, to founding Y Combinator with Robert Morris and Trevor Blackwell: a self-funded angel firm that stumbled onto the batch model (the Summer Founders Program, 8 startups from 225 applicants, including reddit, Twitch's founders, Aaron Swartz, and Sam Altman), funding pairs of founders $6k each for 6%. Hacker News, built in 2006, became YC's largest stress source.

In 2010 Robert Morris urged him not to let YC be his last cool thing; his mother's 2012 illness and 2014 death pushed him to hand the presidency to Sam Altman and return to painting, which he abandoned again that November. From 2015-2019 he built Bel, a self-interpreting Lisp spec, largely in England, before resuming essays in 2020.

VCs aren't jerks by nature; the way their funds are paid makes them act that way. VCs earn about 2% annual management fees plus a cut of gains, so they want huge funds—hundreds of millions—which forces each partner to place multi-million-dollar bets. That single fact explains everything founders hate: paranoid, glacial due diligence (too much is at stake); stealing ideas and leaking secrets to competitors (deviousness under pressure); installing board members or even new CEOs as "political commissars" to watch founders; and investments too large for most startups to absorb usefully—Google could legitimately spend huge sums on servers and bandwidth, but most companies just hire armies to sit in meetings. Big rounds also force inflated valuations, narrowing exit options: a founder might gladly sell for $15 million, but VCs who invested at an $8 million pre-money valuation won't allow it. Graham's own startup spent only $2 million total and sold to Yahoo for $50 million, while a 1997 competitor that raised $20 million was trapped riding its inflated valuation down. This same dynamic pushed Bubble-era companies into IPOs as their only exit. Only elite funds like Mike Moritz's or John Doerr's escape this, since top deal flow lets them avoid needing giant, forced investments.

Venture funding works like gears: a startup should take just enough money at each round to reach the speed needed to shift into the next one. Most get this wrong by being underfunded; a few are overfunded, like starting to drive in third gear. Understanding what investors are thinking matters as much as the mechanics of funding, because conflicts with investors — not competitors — caused the worst problems in the author's own startup, Viaweb: "Competitors punch you in the jaw, but investors have you by the balls."

There are five funding sources. (1) Friends and family — easy to reach (Excite's founders borrowed $15,000 from their parents, stretched over 18 months) but risks mixing business with personal life, and non-"accredited investors" (net worth over $1 million, or income over $200,000) can complicate a future IPO. (2) Consulting — building the product for paying clients funds development without revenue risk, but client calls distract from building a real startup, defined as selling one product to many people rather than doing custom work. (3) Angel investors — wealthy individuals (e.g., Tim O'Reilly in del.icio.us) who bring contacts as much as cash. Angels introduce exit strategy (investors need liquidity via acquisition or IPO) and valuation, which is "voodoo" at the early stage. Terms vary wildly since angels often lack standard agreements; lesser-known angels have less reputation to protect and can behave badly once a startup is desperate for cash, as happened to the author. Syndicates like Boston's Common Angels or the Bay Area's Band of Angels exist, but most angels are independent. (4) Seed firms — companies like Y Combinator (which the author resists calling an "incubator," despite ~800 such incubators per the National Association of Business Incubators) that invest small, standardized amounts at the idea stage and evaluate people over ideas. (5) Venture capital funds — organized as funds charging ~2% management fees plus ~20% of gains; only about 50 of roughly 1,000 US VC funds are consistently profitable, since success is self-perpetuating (Google's returns for Kleiner and Sequoia drew them more deals). Lower-tier VCs have worse brands but offer better terms and higher desperation; some fake interest just to lock founders up, with breakage rates reportedly as high as 50%. VC money brings vesting (4-5 years), board seats, and liquidation preferences (sometimes an abusive 4x). Founders should never mail unsolicited plans — VCs treat this as laziness — and should pursue warm introductions and never believe a deal until the check arrives.

The essay then traces a hypothetical startup. Three founders start with $15,000 from a rich uncle for 5% of the company, keep a 20% option pool, and split the rest evenly (25% each), planning five months of runway — and must start fundraising immediately since deals take time to close. After ten weeks they have a prototype and skeletal business plan. In the angel round, an investor puts in $200,000 at a $1 million pre-money valuation, taking 16.7% of a now-1,200-share cap table, closing after two weeks (three months into the company's life). They hire employee #1 for a modest salary plus 3% restricted stock vesting over four years, shrinking the option pool to 13.7%. By month six they have real users and press attention, and start meeting VCs. A VC firm ultimately invests $2 million at a $4 million pre-money valuation (33.3% ownership), also expanding the option pool, leaving the angel at 10.3%, the uncle at 2.6%, and each founder at 12.8% — with mandatory vesting and a reconstituted board (two VCs, two founders, one neutral member).

The piece closes by warning that this "ideal" narrative omits the norm: deals fall through constantly, far more than real-estate deals, because investors feel buyer's remorse as the risk becomes real. Founders should expect disaster at every stage — but the uncertainty that scares most people away is also what makes starting a startup possible for those willing to try.

Investors are a foreign culture hackers must learn to read, since misjudging them can cost a startup as much as a bad product. Drawing on Y Combinator experience, Graham lists what distinguishes investors.

Investors, not technology, make a startup hub, since capital is the least mobile ingredient — rich people won't relocate for talent, but hackers will move for money. Angels (investing their own money) matter more than VCs: half to three-quarters of companies raising VC series-A rounds already took angel money, and angels take bigger risks and give better advice. Google was effectively angel-funded: its Kleiner/Sequoia series B came at a $75 million premoney valuation, after it already looked like a winner. Angels skip publicity, needing no limited partners to court; VCs build brand to win competitive deals, not cold pitches — deals come via introductions.

Most VCs are dealmakers, not hackers, better at reading people than technology, so they invest like momentum traders, noticing a company already taking off rather than predicting it. They weight traffic first, other investors' opinions second, team third, producing wild unearned swings in a startup's perceived heat. VCs chase potential Googles since fund returns hinge on a few huge wins, and VC terms block early acquisitions; angels are fine with a likely $20 million exit. A $400 million fund split among 10 partners needs $40 million-plus checks, so cheapness to fund can be unattractive; angels write checks as small as $20,000.

Valuations are fictions: desired investment divided by desired ownership stake, not a measure of worth — the same company gets different valuations depending on how much changes hands. A high valuation can backfire, forcing founders to hold out for a huge exit (roughly $100 million after a $10 million premoney round) rather than maximizing odds of a good outcome; less dilution is better achieved by raising less. Investors now chase "the next Larry and Sergey" as they once chased "the next Bill Gates" — flawed, since Microsoft's rise hinged on an IBM contract — and increasingly favor technical founders over MBAs, though only top funds grasp how unimpressive great founders look at the start.

Investors' contributions are underestimated by a founder-obsessed press; founders' contributions are overestimated. VCs are surprisingly timid, avoiding choices — like funding two 18-year-olds — that could look bad to limited partners, even when a "safer" bet (ex-bankers outsourcing development) is objectively riskier. Rejection carries little signal: investors reject for superficial reasons (one killed a deal purely over paperwork), and even Google was turned down. Investors are more emotional than expected, issuing "exploding termsheets" from fear of losing a deal to a rival; negotiation continues to closing, since termsheets aren't binding and "minor" details get renegotiated against founders.

Investors syndicate deals — angels to limit exposure, VCs because multi-VC interest signals safety, and because they collude and trade favors rather than bid competitively (investing is exempt from antitrust law). Large investors optimize their whole portfolio, not one company, sorting startups into successes, failures, and the "living dead," pushing struggling ones toward a win-or-die gamble. Being rich, investors rationally prefer a 20% shot at $10 million over a certain $1 million, so they push founders to reject acquisitions; partial cash-outs spread only slowly. Quality varies enormously among VCs, with the top ~20 firms forming a self-reinforcing tier worth targeting, trusting founders to run their own companies.

Raising money costs enormous time at the worst moment — five to six months typically, 44% of one company's life — yet investors rarely say a clean no, hiding rejection behind lines like "we want to stay in close touch." Nearly every successful startup takes outside money, since even a small lead (Yahoo's, before Google overtook it) can compound into decisive advantage. Investors favor founders who don't seem to need them, so startups should stay cheap to start, keep a backup plan, and behave like a cockroach: small, ugly, hard to kill.

Umair Haque's claim that startups resist acquisition out of a deep sense of purpose is wrong: Google's founders were willing to sell, and Facebook would have sold too, but Yahoo and Microsoft simply offered too little. Startups that reject offers usually do better -- via a bigger offer or an IPO -- mainly because founders bold enough to refuse tend to be exceptional, which is why acquirers who get turned down should raise their offer rather than walk away.

The real reason there aren't more Googles is that VCs won't fund the most innovative startups. After three years running Y Combinator, Graham found VCs surprisingly conservative -- more bureaucrat than pirate -- terrified of truly novel ideas and driven by consensus among themselves (echoing Howard Aiken's line about ramming ideas down people's throats), which guarantees they miss outliers.

A funding gap separates YC's ~$20k seed checks from VCs' ~$2 million later rounds; only scarce angels like Andy Bechtolsheim, who gave Google $100k, cover the ~$200k middle, even as startups get cheaper and the median YC company now wants just $250-500k. Graham's fix: VCs should make five $400k bets instead of one $2 million bet, doing less diligence and skipping board seats. The gap will get filled, by VCs adapting or new investors emerging -- yielding more Googles, as long as acquirers stay stupid.

Raising money is startups' second-hardest problem after building something people want, and it feels harder than expected because the investor market is fundamentally broken. Investors number fewer than ten realistically reachable per deal, so one investor's randomness matters disproportionately; forced to make large decisions about things they don't understand, they turn skittish — enthusiastic one day, unreachable the next, from indecision rather than malice. Worse, investors watch each other: the biggest factor in any investor's opinion of a startup is other investors' opinion, producing an unstable herd rather than a self-correcting market. Y Combinator's fix is to grow both startups and investors, hoping to approach an efficient market — "as t approaches infinity, Demo Day approaches an auction" — but founders must survive today's imperfect system.

An alternative is bootstrapping via consulting, since pure self-funding rarely works: Viaweb, charging $140/user/month, still needed a year to cover costs — more than Graham could have lived on savings. Most "bootstrapped" companies either got lucky or began as consultancies that gradually became product companies, a slow path viable mainly for non-obvious ideas (Joshua Schachter built Delicious on the side on Wall Street because no competitor saw the opportunity). For startups building something obviously necessary, that delay can be fatal. Given a fixed amount of pain, raising money beats consulting, since technology is worth more sooner — but the fundraising process itself, not just failure to raise, can kill a company, requiring survival techniques distinct from persuasion.

Nine techniques follow. (1) Low expectations: assume any deal will fall through — YC's mantra — since deals collapse at the last moment rather than solidifying predictably. (2) Keep working: fundraising silently consumes all attention, so partition roles (one founder handles investors) and prioritize product over meetings, since a stalled company looks undynamic, and investors' own indecision causes the stagnation that turns them off. (3) Be conservative: take any reasonable offer from a reputable investor rather than gambling on a better one, and never let an interested investor "sit" — close immediately or write them off. (4) Be flexible: refuse a fixed funding target; offer a range ($50k for living costs, a few hundred thousand for hires, millions to scale fast), and for angel rounds let the raise expand on the fly or use a "rolling close." (5) Be independent: reaching "ramen profitable" (as little as $2,000/month) transforms bargaining power, since investors favor founders who'll succeed with or without them — Sam Altman exemplifies a founder whose toughness substitutes for profitability. (6) Don't take rejection personally: VC David Hornik's numbers — 500-800 plans read, 50-100 first meetings, ~20 that interest him, 5 pursued seriously, 1-2 deals closed a year — show rejection is the norm; since good ideas must look wrong to be novel, investors judge poorly, so sort rejections for specific, fixable objections rather than blanket condemnation. (7) Downshift into consulting when the product overlaps with services (Viaweb built free sites for early merchants without billing, to avoid becoming a consultancy). (8) Avoid inexperienced investors, whose overcautious lawyers wreck small deals — one YC startup lost a deal when a novice angel's lawyer sent a 70-page agreement he couldn't retract; drive the paperwork yourself or fold novice money into a round led by someone experienced. (9) Know where you stand: the worst outcome is "the long no," so probe investor intentions without pestering, use competing offers to force decisions, and weight prospects by likelihood of yes over deal size, since converting one investor makes the rest more likely to follow.

Graham closes hopeful that falling startup costs and a growing investor pool will make fundraising faster — YC itself now decides in about 20 minutes — and argues current investors' inefficiency is an opportunity for faster-moving competitors. The biggest danger remains surprise: founders expecting fundraising to be easy get demoralized when it isn't, so his final advice is to go in knowing it will be hard.

Picking the right startups to fund matters so much more than anything else in angel investing that dwelling on deal mechanics is almost a distraction — nobody remembers an angel for negotiating a 4x liquidation preference, only for having invested in Google.

The mechanics themselves are simple. You give a startup money and get either preferred stock or convertible debt (which converts to stock at the next big round, sometimes at a discount to protect you when nobody can yet say what the company is worth). Angels often "syndicate," joining a lead investor on shared terms, or invest solo using the standard Series AA documents Wilson Sonsini and Y Combinator published. Two numbers matter: investment size and valuation. Put $50,000 in at a $1 million pre-money valuation and you get 4.76% (post-money $1.05M). Later rounds dilute that stake, which is normal and fine as long as the agreement lets you invest in future rounds to hold your percentage — founders won't dilute you without diluting themselves equally. Angel checks typically run $10,000 to low millions, with a $150,000 round from five people being typical; valuations usually run $0.5–5 million (above that is VC territory). There's no rational way to set valuation — it just reflects bargaining leverage — so guess, since it rarely determines outcomes anyway; success does.

Picking winners is the hard, decisive part. Angels, unlike VCs (who mostly react once something is already winning — and most VCs actually lose money), must predict potential before it's obvious, spotting a great product before users notice (Google) or a company an iteration away from its real hit (PayPal, originally PDA-to-PDA money transfer). This favors people who can judge founders over people who know termsheets — Paul Buchheit was as good as Graham at picking startups almost immediately, because empathizing with founders matters more than experience. The trait to look for is the opposite of "hapless": relentless resourcefulness. On the people-versus-market debate, Ron Conway backs people while Marc Andreessen (along with Jawed Karim and Joe Kraus) prefers a hot market, having each ridden a huge "thermal" in their own startups; Graham sides with Conway, since thermals are unpredictable but good people can ride them when they hit.

Deal flow comes overwhelmingly through referrals rather than events like Y Combinator's twice-yearly Demo Day — and the way to get referred deals is to already be a good, proven investor, since insiders won't send deals to someone unproven. With only a couple hundred serious angels in Silicon Valley, they're the field's limiting reagent, so becoming one measurably widens the startup pipeline.

Being a good angel investor requires decisiveness ("he writes checks") — guess early, since VCs will already have funded the obviously good bets — which a nervous beginner can manage by capping each check at a painless amount (e.g., $15,000 on $5 million in assets, for 3–4 educational investments). Strategic indecision that strings founders along is the worst failure mode. Finally, be genuinely good to founders: startups create wealth rather than compete for a fixed pie, mistreated founders demoralize and underperform, and the most successful angels simply help everyone and trust that good things flow back.

Power in startup funding is shifting from investors to founders, and every feature of the funding process founders dislike is going to get eliminated.

Two years earlier Graham had flagged a gap between VCs, whose model requires large checks, and startups wanting a couple hundred thousand dollars rather than millions. That gap is closing: VCs now make angel-sized bets, and a new class of "super-angels" has emerged — investing like angels but with other people's money, like VCs. There's still room for more such investors, since the distribution of investors should mirror the power-law distribution of startups. Angel rounds may gain enough prestige to compete with series A rounds, though Graham still advises founders to take a series A from a good VC over an angel round, since VCs still deliver more attention. He argues VCs should fear super-angels more than the reverse, following the pattern where market "invaders" (YouTube vs. TV networks, PayPal) usually beat incumbents who treat the new territory as a side business rather than their whole business — angel investing is super-angels' entire game, but just deal flow for VCs. The counter-risk: since startup returns cluster in a few huge winners, super-angels who miss those winners fail regardless of volume.

Why don't VCs simply do more, smaller series A deals? The constraint is board seats: assuming a 6-year startup lifespan and 12 boards per partner, a fund can manage only 2 series A deals per partner per year. Graham argues VCs could take fewer board seats without losing effectiveness, since the useful help doesn't require a board seat. Some VCs will streamline and could plausibly do 2-3x as many series A deals; others will make superficial changes, drifting into what are effectively series B rounds — taking less than the current 25-40% equity, investing later, and likely getting fewer losers in exchange for smaller wins, netting similar risk-adjusted returns.

Deal mechanics will loosen too: rounds will use rolling closes instead of a fixed size and single negotiating lead, standardized paperwork, and increasingly convertible notes with valuation caps rather than a single fixed valuation — letting startups stack multiple notes at different caps simultaneously. The old lead-investor model persisted because it let investors hide behind "I'll invest if others will," producing deadlock; going forward, investors who insist on contingent commitments will simply get shut out of hot deals and end up with worse returns, since herd-following doesn't actually improve investor judgment.

Two further predictions follow the "founders get what they want" heuristic: investors will stop waiting for "traction" before committing serious money (tranched deals — a small initial check with an option on more — are a related abuse that's disappearing), and rounds will close far faster, since YC itself decides in about 20 minutes of review with only a 10% regret rate. Institutionalized delays and option pools mainly served investors' interests and will erode.

Graham closes by arguing this benefits investors too: like novice pilots overcontrolling a plane, investors have been overcontrolling portfolio companies, and investor stress is founders' biggest complaint, more than competitors. Since jobs, not other investors, are startups' real competitor, making the process less painful will expand the total pool of startups — a non-zero-sum gain for investors as a class.

Startup funding is being reshaped by a new investor type, the "super-angel," that sits between angels and VCs and is pushing valuations up, deal speed up, and dilution down for founders. Traditionally angels invested $20k–$50k of their own money with quick decisions and no board seats, while VCs invested other people's money in $1–5 million series A rounds, took board seats, and demanded about a third of the company. This left a "no man's land" around the $400k most Demo Day startups wanted — too big to stitch together from angels, too small to interest VCs — which super-angels now fill, typically investing about $100k, deciding in hours, and (like angels) skipping board seats, while making up to ten times as many investments per partner as VCs.

Super-angels threaten VCs on two fronts: they compete for startups and for the investors who fund them. If top startups get 10x higher valuations before raising a series A, VC returns from winners could fall roughly proportionally (though not exactly tenfold, since later investing also means smaller losses on failures and smaller stakes in winners). VCs' one defense is that startup returns are concentrated in a handful of huge successes (the "chance it's Google"); they could lose most individual deals to super-angels and still win overall if they land those rare giants, aided by superior brand and ability to help portfolio companies — though how much that help is actually worth is, Graham says, undecided and now market-priced for the first time.

Because VC board seats last ~5 years and each partner can handle only ~10, a VC fund can do about two series A deals per partner per year, forcing them to demand a large stake each time. Angels and super-angels, taking no board seats, are happy with a few percent. This is enabling a new middle option: angel-round financing roughly half the size of a series A — e.g., $600k on a convertible note with a $4 million cap, yielding investors 13% versus the 30–40% a series A extracts. Angel rounds also preserve founder control (versus the typical two-founder/two-VC/one-neutral board with VC veto rights), close faster (series A can take weeks to months, ending in a partner-meeting vote with ~25–50% rejection odds, versus hours for the fastest super-angels), and fail gracefully rather than all-or-nothing, letting founders even raise the price mid-round as demand appears.

VCs are fighting back by making angel-sized, valuation-insensitive investments themselves — since they view these as recruiting bait for later series A rounds — which can inflate valuations against super-angels. Some super-angels resist high valuations because, unlike VCs chasing IPO-scale winners, they may be optimizing for quick acquisitions (e.g., a $30M sale), where rate of return, not multiple, matters: a fast 10x beats a VC's eventual big multiple achieved over six years.

Graham predicts convergence — super-angels writing bigger checks, VCs moving faster — but expects several years of intensified competition producing the best of both worlds for founders: fast, high-valuation rounds. He downplays "signalling risk" (VCs in an angel round declining to follow on), arguing strong traction erases it, and notes some YC startups hedge by limiting how much any single VC contributes.

Startup investing is uniquely counterintuitive because two facts govern it: nearly all returns concentrate in a handful of huge winners, and the best ideas initially look like bad ones. Y Combinator's funded companies total roughly $10 billion, but Dropbox and Airbnb alone account for about three-quarters of it — a 1000x variation in outcomes human intuition isn't built to handle. Only about one startup per YC batch meaningfully moves returns; the rest are a cost of doing business. Because a startup's chance of succeeding really big isn't a constant fraction of its chance of merely succeeding, investors must ignore the obvious signal — does this look likely to work? — and chase the harder question of whether it could become huge, much as pilots in clouds must trust instruments over their body, since without a visible horizon the inner ear can't distinguish gravity from acceleration.

Compounding this, great ideas look bad at first — if obviously good, someone would already be doing them. Peter Thiel's Venn diagram of "seems like a bad idea" overlapping "is a good idea" captures the sweet spot; Facebook, dismissed by Graham as "a site for college students to waste time," fit this pattern, as arguably did early Microsoft and Apple (Google didn't — it just looked crowded, not bad).

Worse, there's no way to check a pick for two years, and the one measurable proxy — fundraising success after Demo Day — is not just useless but inversely predictive: since winners can return 10,000x, funding 1,000 failures per winner still nets 10x. Graham calculates YC's true risk-adjusted fundraising rate should be near 30%, yet actual rates run far higher (94% in 2010) — a gap he admits YC will likely never close, since a 30% Demo Day would feel, wrongly, like failure, dilute the brand, and be demoralizing. The real reason, he concludes, is that YC still hasn't fully internalized the 1000x variation. The best defense is to keep asking, when a founder's idea sounds crazy but the founders seem sharp, "who cares what investors think?" — the question YC asked itself about Airbnb.

Founders should stop trying to convince investors with a pitch and instead first convince themselves their startup is worth investing in, then simply explain that clearly. Because startup outcomes follow a steep power law, investors treat "big success" as nearly binary — conventional wisdom holds there are about 15 huge winners a year, and most investors care only whether you seem like a candidate for that group. (A few angels will take moderate success, but even they prefer big wins.) Seeming like a candidate requires three things: formidable founders, a promising market, and usually some evidence of traction.

Formidable founders matter most: investors typically decide within minutes whether you're a winner or loser, and that verdict then colors how they read everything else — the same slow sales cycle becomes a reason to invest or not to, depending on their initial impression of you. Ideas are "fuel for the fire" that starts with liking the founders. Formidable means seeming like you'll get what you want despite obstacles — "justifiably confident." Few people pull this off naturally (some genuinely are formidable, others are skilled con artists), and inexperienced founders shouldn't fake swagger, which just lands them in an uncanny valley.

The real route to seeming formidable is sticking to the truth: confidence comes easily when you know you're right, so genuinely convince yourself the startup is worth investing in — which requires real domain expertise, since without it your conviction is just Dunning-Kruger, detectable by how you answer questions. Graham recalls a professor demanding of a student, "Which one of these conclusions do you actually believe?" — schooling trains people to fill pages even absent real content, and founders carry that habit into fundraising, presenting before they're actually convinced. Since investors are far better at detecting bullshit than founders are at producing it, you should raise money only when you can convince investors, not when you need cash or hit a deadline like Demo Day. Convincing yourself first also forces you to organize your thinking into a real roadmap.

On market: you don't need to prove you'll succeed (unknowable), only that you're a good bet — formidable founders plus a plausible path to a big total addressable market (customers × price). The market needn't be big yet; Microsoft wasn't going to get huge selling Basic interpreters, but was positioned to ride the microcomputer wave up the stack — though at three months in, Microsoft looked merely good, not obviously great, no better a bet than many YC companies. Standards scale with age: a three-month-old just needs a promising experiment; a two-year-old raising a Series A must show it worked.

Investors often ask "who else is investing?" Rather than bluffing about imminent commitments (the single most common lie told to investors), founders should explain candidly why other investors passed and why they were wrong — a tactic that works best on top investors, who know great ideas often look bad initially. Dropbox was passed over by every Boston investor in its YC batch before Sequoia funded its Series A weeks later. Investors also use concise, jargon-free explanation as a proxy for real understanding. The recipe: build something worth investing in, understand why, then explain it plainly.

Raising money in phase 2 -- the round after a small phase-1 seed and before later growth rounds -- is a puzzle as much as a hard sell: investors are pinched between fear of missing a winner and fear of funding a flop, so they habitually stall, mislead, and lead founders on, and inexperienced founders' wishful thinking combines badly with that. External rules, not intuition, get founders through it.

Only raise money if you want to grow faster and outside money would help, and only once you can convince investors -- trying too early burns your reputation. Once you start, be either fully in fundraising mode or fully out, because fundraising becomes the "top idea in your mind" and stalls real growth; outside fundraising mode, accept only money needing zero convincing and zero negotiation (e.g. a standard, well-priced convertible note), and decline "just one meeting."

You need an introduction; the best comes from an investor who just committed, the next-best from a founder they've backed. Treat sites like AngelList, FundersClub, and WeFunder as supplementary, not primary. Treat every investor as saying no until an unequivocal yes -- many never explicitly refuse, they just stop replying, keeping a free option on you. Talk to all prospects in parallel, never serially, weighting attention by expected value (probability of yes times how good a yes would be), so you drift from flakes automatically. After every meeting, ask what happens next; investors' actions, not enthusiasm, show where you stand. The first substantial commitment -- roughly $50k from a known VC or angel -- is the hardest step and unlocks the rest via herd behavior; until money is in the bank, treat it as not raised, since buyer's remorse and shocks (a competitor, a lawsuit, a cofounder quitting) can kill a "done" deal overnight. Investors who won't "lead" -- who invest only once others have -- have zero expected value early on; deprioritize them.

Prepare multiple funding plans (rough ceiling: hires times $15k times 18 months) and quote a lower target than you want -- say $250k when hoping for $500k -- since being over half-raised draws more interest than a third of the way to a bigger number ("angle of attack"). Reaching profitability without any money -- "type A" fundraising, versus need-driven "type B" -- is the strongest position, echoing YC's push for ramen profitability before Demo Day. Don't chase valuation: it's at best third priority after money and good investors, and the real test is revenue -- Dropbox and Airbnb raised post-YC at premoney valuations of just $4 million and $2.6 million. Avoid naming a valuation before you must, and approach "valuation sensitive" investors last. Accept acceptable offers immediately rather than gamble on a better one later (a "greedy algorithm"); a three-day deadline is fine, but shorter "exploding" offers signal a sketchy investor -- top investors like Fred Wilson rarely explode offers at all.

Don't sell more than 25% of the company in phase 2, on top of under 15% in phase 1, or you'll struggle to leave enough stock for a series A. Have one founder, usually the CEO, own the process so others keep building. You'll want a one-page executive summary and, increasingly optionally, a deck. Stop fundraising once you're clearly "getting air in the straw." Don't get addicted to the process -- it's a means, not success itself -- and don't raise too much: an inflated valuation sets an unreachable bar next round, and extra cash breeds rigidity. Stay gracious in refusal, since investors who reject you often return in phase 3. Assume phase 2 money is your last: by phase 3 you'll need real profitability, since startups hose themselves either by being slow to reach it or by hiring too fast on new cash. It reduces to one sentence: avoid investors until ready, pursue them in parallel by expected value, and accept good offers as they come.

The dot-com Bubble was absurd, but the trends it popularized were mostly right, and over the long run what it got right will matter more than what it got wrong. Graham grounds this in his time at Yahoo: when the stock traded near $200, he calculated a fair value of $12 and told his friend Trevor, who couldn't muster real indignation because he knew the valuation was crazy. Yahoo's earnings were partly artificial, too: startups that received VC funding spent much of it advertising on Yahoo, so a capital investment this quarter became Yahoo earnings next quarter, fueling further investment — an unintentional Ponzi dynamic. Starting January 2000 the stock crashed 95%, yet even at its post-crash 2001 valuation, Yahoo was still an $8 billion company built in six years — proof there was real substance underneath the hype, the way even Isaac Newton and Jonathan Swift got burned by real value inside the 1720 South Sea Bubble.

Graham then lists ten things the Bubble got right. (1) "Retail VC": taking earnings-less companies public isn't inherently foolish — it's just VC investing opened to public markets, and markets will eventually learn to price such stock as well as VCs do. (2) The Internet is genuinely a huge deal, but investors erred by backing the most "Internettish" companies (e.g., Pets.com) rather than indirect beneficiaries — as with the railroad boom, the big money went to Carnegie's steelworks and Standard Oil, not the railroads themselves; expect ten JetBlues for every Google. (3) New communication technologies always cause big shifts because they multiply choices: the Internet lets anyone find you at near-zero cost and speeds word-of-mouth, replacing bottleneck control (the old "channel") with "build it and they will come" — Google, with over 82 million monthly users and roughly $3 billion in annual revenue, barely advertises at all. (4) Young founders (some 26-year-olds) can out-execute 50-year-olds because vision matters most and management/SEC compliance can be delegated. (5) Informality (open-necked shirts replacing suits) reflects nerds' instinct that dressing up substitutes for good ideas. (6) Nerd culture is rising because nerds prize substance over self-marketing, and technology increasingly rewards that. (7) Stock options are fair and effective — people work harder when they own equity — though options can perversely reward pumping the stock price rather than building real value, so the mechanism may need tweaking. (8) Startups increasingly exist to be sold, developing technology "on spec" for acquisition, as Graham's own Viaweb was built to be; this lets big companies get innovation with less bureaucracy and more accountability. (9) Silicon Valley's culture, not just its bubble-era money, remains genuinely superior soil for building the future, unlike Boston's entrenched institutions. (10) Technology will multiply productivity variation, not just add to it, invoking Fred Brooks's point that small teams are disproportionately efficient — meaning future companies may stay surprisingly small.

Graham's overall theme: in the coming century good ideas will increasingly beat formality, connections, and marketing — a meaningful acceleration of a process that historically took decades (as with relativity or the failure of central planning), even if it doesn't fully justify the "new economy" label.

PR firms function as a submarine beneath the news: more than half of non-political, non-crime, non-disaster stories in traditional media originate with them, not because PR is dishonest but because it exploits reporter laziness through selective truth-telling, like a flatterer who never lies but chooses which true things to say. Graham learned this running a startup that paid $16,000 a month for PR while assembling its own computers to save money, and landed press hits in 60-plus publications in 18 months. Trade publications, dependent on ad revenue, sometimes print PR pitches nearly verbatim; top papers like the New York Times and Wall Street Journal are more skeptical, but their weak point is vanity — stories must appear to be the reporter's own idea. Graham's firm engineered a two-step coup: get an estimate of "5000 stores on the Web" printed as fact, then cite that figure to claim "20% of the online store market" with just 1000 customers. Reporters crave punchy, definitive numbers over hedged ones — Spamhaus's rough ROKSO-list guess became the unquestioned claim that convicted spammer Jeremy Jaynes was "one of the 10 worst," and a guessed infection count for the 1988 Internet worm ("6000 computers") is still cited as fact today.

PR's more deliberately misleading tactic is manufacturing "buzz" by feeding one story to many outlets at once, creating the illusion of a trend — as with the Windows 95 launch, or the recurring "the suit is back" story that reappeared in February, September, June, and March 2004 and several times before that. Searching shared phrases reveals the pattern: pieces in the Boston Globe, US News & World Report, Sexbuzz.com, and the Detroit News all cited the same GQ creative director as an "independent" expert, tracing back to a Men's Wearhouse ad campaign literally titled "The Suit Is Back."

Online writing, by contrast, is honest because people publish what they actually want to say, without PR's fingerprints — which is why newspapers face a structural, not cyclic, decline, and why PR firms will inevitably adapt to target bloggers next.

Software patents are not a distinct evil: since software is just another substrate for control systems once built from levers and gears, anyone who opposes software patents logically opposes patents in general. Patent law's exclusion of "algorithms" is an 1800s holdover, so lawyers now disguise algorithm patents as "system and method" claims. The real problem is the USPTO, overwhelmed by the volume and novelty of software applications, routinely granting patents that fail the non-obviousness test — Slashdot's default patent-story icon is now a knife and fork stamped "patent pending." Amazon's one-click patent is the classic case: obvious, not novel, and examiners' fault for not narrowing it, not Amazon's for applying (applying is a negotiation where you ask for more than you expect to get). Amazon's real sin was enforcing it against Barnes & Noble, "the equivalent of a nuclear first strike" that cost more reputation than the suit was worth, given B&N was weak anyway. Big companies like Microsoft mostly hold patents defensively, as deterrents.

Graham separates two conflated questions: is it wrong to apply for patents under current rules, versus is it wrong that the system allows patents at all. Like medieval self-defense before police existed, or checking in hockey, applying is just playing by the game's current rules — refusing is like refusing to use TCP/IP. Startups almost never get sued for infringement: too poor to sue for money, too young for competitors' patents to have issued, and established players (Microsoft never sued a startup) prefer buying or locking rivals out of sales channels. Big companies suing smaller ones (Unisys over LZW compression) signals decline — "when a company starts fighting over IP, it's a sign they've lost the real battle, for users." Startups get sued only once large or public, chiefly by trolls. Graham still advises patenting anyway, not to sue but because patents aid the "mating dance" with acquirers, who prefer buying to building and want an excuse to admit they can't replicate the technology.

Three reasons patents matter little in software: unlike processes such as ore smelting, software is too subtle for "qualified experts" to replicate from a patent alone — design, not implementation, is the scarce skill, so a real product outweighs the sum of its patents (contrast the baggage-scanner industry, where startup Reveal was sued by incumbents InVision and L-3 before shipping; software's only parallel is Yahoo's minor 2005 suit against gaming startup Xfire). Second, startups beat incumbents obliquely (Writely versus Word) rather than head-on, and big companies' denial keeps them from noticing threats in time to sue (as IBM ignored microcomputers). Third, hacker opinion constrains big companies — Google's "don't be evil" stance and the boycott risk that kept Steve Ballmer from actually attacking Linux over patents.

Patent trolls — lawyer-firms holding patents solely to extract settlements — are the exception and unambiguously "evil," comparable to pre-industrial courtiers extracting crown-granted monopolies, and to the mafia as an obsolete business model. They've extracted hundreds of millions of dollars and resist counter-suits since they make nothing, though Graham predicts the loophole closes eventually and argues trolls don't suppress innovation since they strike only after a startup has already succeeded.

On whether patents help or hinder innovation overall, Graham declines to answer, noting most opinions are "religious conviction" rather than research. Patents trade exclusivity for public disclosure, replacing the inefficiency of a secretive "need to know" world (echoing medieval guild secrecy, like Venice executing glassblowers who left the city). His conclusion for software: patents don't meaningfully affect innovation either way; startups should ignore competitors' patents, focus on building something great with lots of users, and treat their own patents as mostly acquisition currency.

Microsoft, which loomed over the software industry for two decades starting in the late 1980s (succeeding IBM as the feared monopolist), had quietly stopped being dangerous by around 2005 — evidenced by Y Combinator never inviting Microsoft to its startup demo days. Four simultaneous forces killed it. First, Google took the lead in 2005, propelled by Gmail, which proved the company could do more than search. Second, Gmail also showed the power of "Ajax" web apps, making the desktop obsolete — ironically built on XMLHttpRequest, which Microsoft itself created in the late 1990s for Outlook, plus Javascript, which Microsoft tried to keep broken until open-source libraries overcame it. Third, broadband internet access removed the need for local desktop software. Fourth, Apple's OS X drove a comeback so complete that Windows became "for grandmas" among people who cared about computers, with Apple also beating Microsoft in music and gaining ground in phones. Graham argues Microsoft could theoretically fight back by using its cash pile to buy up "Web 2.0" startups and isolate them from Redmond, echoing how cheaply Google itself was once for sale, but predicts they won't, because they still don't grasp how much they now suck.

Community sites don't have to be ruined by growth—dilution is a hard but solvable design problem, and two years of running Hacker News reveal what actually works. Weekday traffic grew from about 1,600 daily uniques at launch (February 2007) to around 22,000, a 14x increase Graham calls faster than he'd like; he wants growth but not Digg- or Reddit-scale growth, since that would dilute the site's character. Performance has stayed "consistently mediocre" despite the growth, guided by McIlroy and Bentley's maxim that elegance, not special-casing, is the key to performance.

Dilution, Graham argues, is measured more in behavior than in users: bad behavior is what to exclude, and behavior is surprisingly malleable—people tend to meet the standard expected of them. The broken-windows theory applies to communities: Giuliani's reforms transformed New York, while Reddit's opposite policy (censoring only spam, prioritizing growth over thoughtfulness) transformed it the other way, even though Reddit's traffic far exceeds HN's. A failed experiment—coloring high-scoring users' names orange—split a previously unified culture into haves and have-nots and was reversed.

On submissions, bad stories have proven less dangerous than bad comments; the frontpage has held up well. The real threat is "the Fluff Principle": on a voted site, the easiest-to-judge links win out unless deliberately prevented. HN counters this by banning off-topic fluff (kittens, political rants), having editors kill or rephrase linkbait titles, watching for "linkjacking" (paraphrased reposts that out-perform originals), and exposing killed submissions via a "showdead" toggle so editorial honesty stays checkable.

Comments are harder: meanness is controllable by rule, but stupidity is harder to police because it's less recognizable to the person committing it. The most dangerous stupid comments are short dumb jokes, not long wrong arguments—comment length correlates with quality, and bad comments spread like kudzu, setting the tone for replies around them. Graham floats delaying replies in proportion to predicted comment quality as a fix.

Ultimately, good communities come from people, not technology; HN's plain design deliberately repels casual users to attract only those interested in ideas. Graham closes on addictiveness as an unsolved problem—an online "town square" more dangerous than a physical one because visiting costs just a click—comparing today's addictive social software to crack in the 1980s: invented before we've learned to protect ourselves from it.

Consumers never really paid for content, only the medium: publishers sold paper, not words, and price tracked format, not quality -- Time ran 8.6 cents a page ($5 for 58 pages) versus a cheaper 8.1 cents for the better Economist ($7 for 86). Print media's real business was marking up paper; as paper becomes unnecessary, so does the revenue, leaving publishers nothing to sell.

Selling information directly has always been marginal -- stock-tip newsletters, Bloomberg terminals, Economist Intelligence Unit reports -- because people pay only for information they expect to profit from, not content generally. iTunes isn't a counterexample: it's a tollbooth exploiting Apple's control of the default path onto the iPod, charging amounts small enough to ignore; once a toll grows painful, digital content is too easy to route around. Digital books face the same downward pressure as writers realize they don't need publishers. Software looks like a counterexample too, but businesses buy it partly to avoid piracy liability and treat it as a tool, not content -- hence "content" for information that isn't software. Movies have fared better because broadcasting isn't publishing (no copy changes hands), though that shelter has limits, and streaming subscriptions won't save labels if they're just the same files as mp3s.

The likely path: give content away and profit indirectly (music via concerts and merchandise, writing via ads), or embody it in physical objects worth buying -- some magazines via lush print, some books since publishers already found, by the 1960s, the cheapest format buyers would still accept. Movies might become ads, or make theaters a treat again, or the business could shrink toward game development. Winners use new technology to give people something new; losers merely defend old revenue.

Addictive things are concentrated versions of milder predecessors — hard liquor from wine, heroin from opium, crack from cocaine — and the process that concentrates them, technological progress, is accelerating, so more and more things we like are becoming things we like too much. English lacks a word for this beyond colloquial "addictive." Factory farming and food processing did it to food; World of Warcraft and FarmVille did it to games; Facebook out-competed even TV. The world is already more addictive than 40 years ago and will get more so in the next 40, pulling apart the two meanings of "normal" — statistically typical versus optimal — so that living well increasingly looks eccentric. Societies eventually build customs as "antibodies," as with cigarettes, which took about 100 years to go from universal to seedy before legislation followed; but antibody formation can't keep pace with accelerating addiction, so people must individually learn to suspect everything new, including old things (like the Internet) turning newly addictive. Graham's own response: no iPhone, and long hikes instead of running, for uninterrupted thought. He predicts people will increasingly be defined by what they say no to.

Yahoo failed for two reasons Google didn't share: easy money and ambivalence about being a technology company. In 1998 Paul Graham showed Jerry Yang "Revenue Loop," a shopping-search ranking algorithm sorting results by bid times conversion rate—essentially the model Google later used for ads. Yang was unmoved: advertisers already overpaid for banner ads, with sales chief Anil Singh's team landing million-dollar orders from companies like Procter & Gamble. Worse, Yahoo rode a de facto Ponzi scheme—investors excited by Yahoo's revenue growth funded new startups, which spent the money buying Yahoo ads, inflating Yahoo's revenue further. Both kinds of advertiser wanted brand exposure, not targeting, so raw traffic mattered more than its quality, and search seemed unimportant: when Graham told David Filo that Yahoo should buy Google, Filo replied that search was only 6% of traffic while the company grew 10% a month. Money was "the most opaque obstacle" blocking Yahoo from seeing search's value.

The second problem was cultural. Yahoo insisted it was a "media company"—partly because it earned money via ads like a media company, partly from fear of being crushed by Microsoft as Netscape had been. It adopted media jargon ("producers," "properties") and treated programmers as commodity labor executing product managers' specs, rather than building the hacker-centric culture of Microsoft, Google, and Facebook. This produced weak hiring, and since good programmers want to work with good programmers, Yahoo entered an irreversible "death spiral." Graham found 500-person Google still felt like a startup, with engineers asking unprompted how to fight SEO gaming; Zuckerberg later said Facebook hired programmers even for HR and marketing roles. Graham's conclusion: any company needing good software can't afford a "suit-centric" culture—the opposite of what Yahoo executives meant by "adult supervision." As he puts it, there are worse things than seeming irresponsible: losing.