Call off the cults and the funerals. “AI”’s real advance is high‑dimensional prediction that generalizes without interpretable laws. That’s operational power, not Turing-Class cognition. Noah Smith says: Think of three magics: literate historical memory made knowledge accumulative; hypothesis-and-experiment science made it generalizable; AI makes it operational at scale via learning‑and‑search replacing inadequate low-dimensional cookie‑cutter models with extremely high-dimension extremely big-data extremely flexible-function prediction. But is that “Third Magic” really of the same scale as the first two? I would view things somewhat differently: I would add eyes-thumbs-brains-tools, language, and societal coördination via scaled-up gift exchange to writing and science as decisive magics. And I would say that at the moment “AI” is as likely to be a wishful mnemonic as a genuine Sixth Magic. Chatbots are useful, but they’re pass‑the‑story engines: blurry‑JPEG‑of‑the‑web plus rotoscoping, not minds. The hype machines—Downer, Boomer, Doomer—confuse cultural technology with cognition and policy with prophecy. The Downer critique underrates genuine capabilities; the Boomer gospel overstates sparks‑of‑AGI. The Doomer rapture is theology in tech drag. The economic story is a bubble build‑out: GPUs gush profit; most applications burn cash. Economically, chips win, most deployments don’t; fear of disruption fuels spending by platform monopolists remembering the fates of IBM and WIntel. Use AI where feedback is tight and stakes local; demand theory or rigorous trials where failure is catastrophic; measure, pre‑register, and watch the stages of the roll-out carefully to gauge what all this will really mean for us...
Noah Smith reposts <https://www.noahpinion.blog/p/the-third-magic-23f>
, with further thoughts, his “Third Magic” essay <https://www.noahpinion.blog/p/the-third-magic> from year-end 2022, which at the time I said was smarter than anything I had read on the internet in 2022.
Here’s why I said that, and what I said back then:
This by Noah Smith May Well Be Better than Anything I Read in All 2022:
But I would say not “three magics” but five—and maybe six:
Noah Smith: The Third Magic <https://www.noahpinion.blog/p/the-third-magic>: ‘A meditation on history, science, and AI: More profound and fundamental meta-innovations… these are ways of learning about the world: The first magic… history… knowledge… recorded in language…. Animals make tools, but they don’t collectively remember…. History… is what allows tinkering to stick…. Our second magic trick… was science… figuring out… principles about how the world works…. Controlled experiments…. It’s pretty incredible that the world actually works that way. If… in the year 1500… [you] told… [someome] one kooky hobbyist rolling little balls down ramps could be right about how the physical world works, when the cumulated experience of millions of human beings around him was wrong… they would have thought you were crazy. They did think that was crazy. And yet, it… worked…. But… complex phenomena have so far defied the approach…. Language, cognition, society, economics, complex ecologies these things so far don’t have any equivalent of Newton’s Laws, and it’s not clear they ever will….
The Third Magic…. “Statistical Modeling: The Two Cultures”… [a] split between… parsimonious models… [and] predictive accuracy…. Control…. Generalize… without finding any simple “law” to intermediate…. Halevy… Norvig, and… Pereira… “The Unreasonable Effectiveness of Data”…. “Represent all the data with a nonparametric model… with very large data sources, the data holds a lot of detail…. Trust… language…. See how far you can go by tying together the words that are already there…. Now go out and gather some data, and see what it can do”…. Underlying regularities that are difficult to summarize but which are still possible to generalize…. Black-box prediction…. We’re always in danger of overfitting and edge cases…. The “third magic” may be more like actual magic than the previous two…. But even wild, occasionally-uncontrollable power is real power….
[In] a number of subfields of economics… [circumstances] resist the natural experiment approach…. Might we apply AI tools?… Khachiyan et al. argues… “yes”…. Being able to predict the economic growth of a few city blocks 10 years into the future with even 30% or 40% accuracy…. And this is just a first-pass attempt…
What do I think? I think five or maybe six magics:
The first magic is at the individual level: eyes, hands, and brains that allow us, but just to use and make but to plan for tool-use, and so the East African Plains Ape gains fangs, claws, fur, levers, and much else.
The second magic is at the small group-collective level: ears and mouths that allow us to talk using language, and so the band of fifty of so East African Plains Apes becomes an anthology intelligence: what one knows, soon all know.
Then comes the third magic, which is Noah’s first: the artificial memory and communication device that is writing, and all of a sudden it is not just the band of fifty that is an anthology intelligence, but potentially all humanity in the present and the past.
The fourth magic is a magic of coordination: How do you get us all in harness, pulling in the same direction, but each doing our (relatively specialized) part? The fourth magic is gift-exchange, that in its hypertrophied form turns into the market economy. Doing favors for one another, in a social context in which there must br approximate balance, is what allows us to, collectively, not just know things but do things.
After that comes, as Noah correctly notes, the fifth magic of science: the experimental method, and the exaltation of ideas not because they solidify the band or help some élite run a force-and-fraud exploitation-and-domination game, but because the ideas are true and are a univeral remote control—they enable us to understand and hence control the universe.
Those five magics have brought us where we are today.
Now how likely is it that we are now at the cusp of a sixth magic? It would be: Predictive accuracy, generalization, and control without any simple intermediating laws, abstractions, or encapsulations that vastly exceeds any individual human grasp, or even the grasp of all humans working together.
Perhaps. But I need to explain what is going on. Predictive accuracy, generalization, and control based simply on the fact that we have a huge amount of data, plus enough computer power to allow us to conduct extremely high-dimensional analysis using extremely flexible functional forms. Thus we can classify situations very finely by looking at what the situations’ nearby neighbors—for the right meaning of “nearby” which we can now figure out—are?
On the one hand, I feel that this must be true—I find it very hard to imagine what our brains, or, say, a dogs’ brains, are doing in wetware other than this. On the other hand, our computers are still a lot less sophisticated than our brains. It is unclear if Moore’s Law will get us to computers of sufficient complexity. It is unclear if we can do as good a job of programming our computers as evolution has done at programming us. And, of course, we already have the real ASI—the Anthology Super-Intelligence that is all of us humans thinking together. Adding-in the new capabilities of silicon to our current ASI makes it vastly more capable. But is it enough to qualify as a genuine Sixth Magic? Perhaps. But, at least as I see it, magics one through five—eyes, hands, brains, and tools; ears, mouths, and language; writing; coördination at scale via gift-exchange amped up by money; and science—were all genuinely world-changing in a way that this latest is not yet proven to be.
However, it is now nearly three years later along this ride. What have we learned, and how do things now look different?
Today Noah adds on further comments:
Noah Smith: The Third Magic (2025) <https://www.noahpinion.blog/p/the-third-magic-23f>: ‘Back in 2023, ChatGPT was very new…. I’m not an AI engineer myself, but I could tell that this was a type of technology unlike any other ever created… actual magic… the actual kind of spells that wizards cast in storybooks…. AI isn’t interpretable; we can use it somewhat reliably to do amazing stuff, but there’s a ton of mystery meat in terms of how it actually did it…. Here, for example, is a cute video I made with the new Sora 2 AI video generation app from OpenAI:
Anyway, this is thrilling, but at the same time it’s slightly worrying as well. This technology is so powerful that we’re going to have no choice but to rely on it…. Every scientist must now be, to some degree, a spellcaster…. Technology will be more powerful, but less reliable…. On top of that, I worry that humanity will become infantilized by this new magic we’ve created…. We could find ourselves wandering, confused, in a world of ineffable mysteries and capricious gods.
That thought made me want to repost my New Year’s essay from 2023. In that essay, I argued that humanity had basically found two great tricks for gaining power over the world — history, which records the past, and science, which derives simple “laws” from controlled experiments. AI, I speculated, could be a third thing entirely…. As a coda, I thought I’d ask GPT-5 what it thinks of this blog post. Here’s its response….
“A compact rewrite you might consider (core claim): The first magic made knowledge accumulative (memory). The second made it generalizable (theory). The third makes it operational at scale (learning-and-search). We should… ask how to stack them: record more, explain what must be stable, and learn the rest fast enough to act…. Concrete additions that would level this up…. One macro use-case sketched end-to-end: e.g., place-based policy: (i) satellite/LLM features → growth forecast (AI), (ii) theory-guided constraints (agglomeration, congestion) (science), (iii) program memory: standardized interventions + postmortems (history). Close with how you’d validate (pre-registered policies, staggered rollouts, causal ML)….. You’ll convert skeptics while keeping the boldness that makes this piece fun…”
What do I see, looking back at Noah then, and sideways at what Noah adds now and what 2.5 years of the GPT LLM MAMLM—General-Purpose Transformer Large Language-Model Modern Advanced Machine-Learning Model—infotech ride has brought us? From my present standpoint, I see seven things going on that Noah mashes together, causing some confusion:
First, very big-data, very high-dimension, very flexible-function classification and prediction: The replacement of cookie-cutter by pattern-recognition and downstream algorithmic decision-making and societal-organizing systems. This is Noah’s real Third (and my perhaps-Sixth) magic. It will be powerful. It may be amazing. The jump in capabilities to conduct analysis has been remarkable—protein-folding, detailed geographic-impact estimation and -situation assessment (for, say, understanding global warming and disaster response more generally), materials modeling and discovery, literature surveys to keep discoveries and insights from being lost in the noise, multimodal screening, personalized education, logistics systems, and more. Here the build-out of capabilities has been more-or-less as I expected, demonstrating that while Moore’s Law as we knew it may be dead, that pace of progress in infotech is not.
Second, natural natural-language interfaces to structured and unstructured databases: This is really a subcategory of the first: VBDVHDVFF analysis applied to linguistic patterns. But we are so primed by cultural (and biological?) evolution to use natural language for information and communication that it looms as large in humanity’s near- (and far-?) term future as all the rest of the first put together. And it is wonderful! to be able to ask a question and get an answer! Yes, it is unreliable—bur having to interface through programming and query languages was unreliable from the end-utilizer standpoint as well. Here too the build-out of capabiliries has been more-or-less as I expected.
Third, ChatBots as a subcategory of the second, but worth distinguishing from the rest because we are so primed by cultural (and biological?) evolution to pay attention to things that talk to us that it looms as large in humanity’s near- (and far-?) term future as all the rest of the second put together. These are, I think, best seen as answering the question: “What would a human with near-infinite recall and ability to read and search typically say if they were in this linguistic situation?” That is excellent for summarization. That is very good for search—at least until the SEO vultures arrive in force. That is wonderful for editing. It is turning out to be more work than I thought it would be to get the quality up above that of the TIS—the Typical Internet S***poster. Unless you are happy with that quality (which you may well be if the end target is boilerplate, cliché, ritual, or less politely AI-slop) you have a lot of work to do. And the best use of the technology is to give you a starting point, after which you say “I can do so much better than that!” and set to work.
Fourth, the AI-Downer hype machine is a bizarre thing. Much as I love “Mystery AI Hype Theater 3000” <https://www.dair-institute.org/maiht3k/>, the AI-Downer reaction is vastly overblown. It is simply not the case that there is nothing there but non-conscious Markov-machine autocompleting stochastic parrots extruding synthetic texts that divert our attention away from language with meaning to meaningless linguistic strings, and in the process help the rich get richer by justifying data theft, fuel surveillance capitalism, devalue human creativity, replace meaningful work with machine-like jobs, impose high environmental and financial costs that doubly punish marginalized communities, propagate hegemonic viewpoints that encode biases that harm marginalized populations via synthetic text containing racist, sexist, ableist, extremist, and other derogatory ideologies, mislead the public and researchers, and divert attention from research directions that do not depend on LLMs-scaling.
My current belief is that those who see no thought—nothing, as I said, there but non-conscious machines that extrude synthetic texts—in these roiling boils of linear algebra are wrong.
Conside a prompt and then a ChatBot response up through token-n. Now the machine is about to output token-n+1. It looks back through its compressed training data at all of the cases that are sufficiently-close to the prompt-plus-response-through-token-n.
In each of those cases, there was a real human—a Turing-Class mind—thinking thoughts that led them to choose to write the next token. The ChatBot then does some kind of stochastic interpolation of the near-neighbor next tokens. In its decision to output that particular token, it is echoing some dimensions of the characteristics of the mind and thoughts in the training data underlying the choice in the training data to pick that next word. That is no more thoughtless than the human choices to make the links Google used for PageRank were thoughtless. It’s not a mere thoughtless synthetic text extrusion machine.
Fifth, the AI-Boomer hype machine is an equally bizarre thing. Extend our consideration to the subsequent tokens. Relatively soon—unless the ChatBot is plagiarizing—the set in the training data of near-neighbors to the prompt-plus-response shifts. Token choice is still drawing on and is still a shadow of human thought, but it is a different set of humans writing under different circumstances and thus having different thoughts. It is, at some level, the game pass-the-story where people sit in a circle, and each one gets their turn to continue the narrative. It is amusing. It is fun. It does not produce a narrative.
That makes those AI-Boomers who see thought approaching Turing-Class—“sparks of AGI” (or ASI)—in today’s GPT LLM, MAMLMs of today are just as wrong as the AI-Downers. The good metaphors are phrases like “blurry-JPEG-of-the-web” and “rotoscoping” and “pass-the-story”. Even as they are bolted on to immense search capabilities and the ability to rapidly conduct their very big-data, very high-dimension, very flexible-function classification of structured and unstructured databases, they are not going to be properly classified as Turing-Class minds, let alone Minds.
Sixth, the AI-Doomer hype machine is fully an order of magnitude more bizarre. I still have not found a better take on them than Cosma Shalizi’s:
Cosma Shalizi: O Ye Who Believe in the Resurrection & the Last Day (2023-5) <http://bactra.org/notebooks/nn-attention-and-transformers.html#language-models>: ‘Discussion… is polluted by maniacal cultists with obscure ties to decadent plutocrats… [who] go fromm… conditional probability, via Harry Potter fanfic, to prophesying… an AI god… judg[ing] the quick and the dead… condemn[ing}… to the everlasting simulated-but-still-painful fire…. Sydney/Bing is no more the Beast, or even the Whore of Babylon, than was Eliza.) This isn’t to deny that there are serious ethical and political issues about automated decision-making… large language models… [whether] the interface to… human knowledge [should] be a noisy sampler of the Web.… It is to deny that engaging with the cult is worthwhile…
Seventh, the econo-financial earthquake that is the bubble build-out of these technologies, driven as it is, I think, by one part reasonable expectation of providing and financially capturing true end-user value, two parts grifters seeking easy investor marks moving over from crypto, three parts millennarian religious hype on the part of those hoping for the Rapture of the Nerds, four parts perhaps reasonable expectation of improved advertising targeting for user good and (mostly) user ill, and five parts platform near-monopolists fearing the loss of their profit flows to a Christensenian disruption from the Next New Big Thing.
Now it is true that, of these, (5) and (6) are the same thing—or rather opposite sides of the same coin. And that (4) on the one hand and (5) and (6) on the other are also opposite sides of the same coin (bear with me: it’s a coin embedded in a four-dimensional space with complicated opposite sides). But (1), (2), (3), (7), and [(4), (5), and (6)] are all very close to being mutually orthogonal. (OK: it’s embedded in an eight-dimensional space; and if one sees (7) as itself five different things, twelve-dimensional.)
I think those who want to build on Noah’s (very good) points need to distinguish, need to work in that eight-dimensional space. But that things are really complicated does not mean that Noah’s insights are not really, really worth reading here.