David Shapiro · Tech & AI
TIER 4 Sat, 10 Jan 2026 14:38:41 +0000
How to use AI without losing your mind ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ | | ---|---|--- | | | Forwarded this email? Subscribe here for more --- # Don't Become a Librarian of Machine Thoughts ### How to use AI without losing your mind | | David Shapiro --- | Jan 10 --- | --- --- | | | --- | | --- | | --- | | --- | | READ IN APP --- AI is making people dumber. Well, maybe. That's what everyone's saying, at least. There was that MIT study that showed when people use ChatGPT to write, their brains "switch off." The headlines wrote themselves. Brain rot. Cognitive decay. The machines are hollowing us out. But when you actually read the study, that's not quite what it's claiming. The experiment was designed to maximize offload. Students were told to write essays as quickly as possible using AI. And yes, there was less brain activity--because there was less critical engagement. That's not a revelation. That's a tautology. I don't want to litigate whether AI is causing brain rot. I don't want to argue that it isn't, either. Let's just assume it's real. Let's assume that something about this technology is genuinely different, that it's eroding cognitive capacity in ways that books and calculators and Google never did. What do you do about it? That's the question that actually matters. And the answer has nothing to do with quitting AI. With that said: The most important skill to cultivate in the AI era is _synthesis_ --the ability to generate, connect, and articulate ideas without external reference. Because if you can think without a cognitive prosthesis, you will never be replaceable by one. I want to share five ideas on what's actually causing the rot, why it matters more than the panic suggests, and how I personally avoid it while still using AI every single day. * * * ## I -- The Socratic Trap Two and a half thousand years ago, Socrates had a problem with books. He believed that writing things down would create forgetfulness in learners' souls. They would trust external marks rather than using their own memory from within. They would appear wise without actually being wise--filled with the conceit of knowledge rather than knowledge itself. He was skeptical of offloading cognition to technology. Sound familiar? I fell into the Socratic trap myself. Last year, I realized I didn't like how much I was relying on NotebookLM for my videos. I'd do hours of research, dump it into the tool, let it generate a beautiful slide deck, and then present from those slides. The decks were better than anything I could have made manually. More polished. More organized. More visually coherent. But something was missing. The _performance_ was gone. What I mean by performance is this: when I get up in front of a camera and talk for thirty minutes straight, connecting ideas extemporaneously, maintaining a throughline without notes--that is a skill. It's the skill of holding an entire argument in your head and delivering it live. And I was letting it atrophy. So I went back to riffing. Back to the cognitive friction of having to synthesize in real time without a prosthesis holding my hand. Now, here's where Socrates gets interesting. He was wrong about books. Hindsight being 20/20, books didn't destroy civilization--_**they built it.**_ The written word allowed for the rapid transmission and dissemination of ideas across continents and centuries. It created the Enlightenment. It created science. It created everything we recognize as modern intellectual culture. So Socrates was wrong about the technology. But he was right about the mechanism. He understood something that most people miss: _if you remove the friction, you lose the muscle._ A book is cognitive offloading. Someone else has done the research, the memorization, the effort to distill an idea--and they give it to you on a silver platter. That's what a book is. That's what it's always been. The question becomes: what's the structural difference between books, the internet, and AI? Two things. First, AI compresses all human knowledge into a single interactive interface. If you had told someone twenty years ago that deep neural networks would compress literally everything humans have ever known into a few gigabytes that fits in your pocket and talks back to you--they would have said you were nuts. That's magic. That's genuinely unprecedented. Second, it removes nearly all friction from the retrieval process. With books, you still have to read. You have to engage with the text, follow the argument, do the work of comprehension. That friction is load-bearing. It's doing something to your brain. With AI, you can skip straight to the answer. You can bypass the struggle entirely. And struggle is where learning lives. Access to information is not intelligence. Friction is the price of comprehension. * * * ## II -- The Shortcut Engine Here's something you need to understand about your brain: it's a shortcut engine. Human cognition is constantly min-maxing effort for reward. We are always looking for the path of least resistance. This isn't a bug--it's a feature. It's how we survived as a species. Conserve energy whenever possible. But that same feature becomes a liability when the shortcuts are too good. Let me tell you about Shane. Shane was a former family member of mine--in-law situation, long story. He was the laziest person I have ever met in my life. And we had a running joke about him: _" What would Shane do third?"_ Here's what I mean. Shane was putting up party lights in his backyard. The correct way to do this is to dig post holes, set the posts in concrete, and string the lights between them. But Shane didn't want to dig post holes. So the first thing he tried was putting a four-by-four in a Home Depot bucket filled with concrete. He figured he could just set the bucket on the ground and pull the lights tight. Of course, it fell over. The bucket wasn't anchored. Physics doesn't care about your laziness. Here's the thing that made Shane _Shane_ : he knew it wasn't going to work before he tried it. He bought the four-by-fours. He bought the concrete. He mixed the concrete. He poured it into buckets. He did all of that _knowing_ it would fail--because all of that was still slightly less effort than just digging the post holes. The second thing he tried also failed, for similar reasons. The _third_ thing he tried was digging the post holes. The thing he should have done from the beginning. The thing that would have saved him hours and hundreds of dollars. What would Shane do third? The answer was always: the right thing, eventually, after exhausting every possible shortcut first. That story is funny. But it's also a diagnosis. When you prompt ChatGPT hoping for a quick answer so you can switch your brain off--you are Shane. It's good that people are asking "Grok, is this true?" But you need to take it one step further. When you copy-paste AI output into your document without really understanding what it says--you are Shane. When you use AI as external memory because you don't want to do the work of actually learning something--you are mixing concrete in a bucket because you refuse to dig the post hole. **Your brain is a shortcut engine. AI is the most powerful shortcut ever invented.** **That combination is the problem.** Your executive cognitive function only kicks on when something fails. And AI is getting good enough that it fails less and less often. Which means your critical engagement kicks on less and less often. Which means the muscle atrophies. This isn't AI's fault. This is human nature meeting a technology that's perfectly designed to exploit it. The question is: what kind of person do you want to be? * * * ## III -- Three Traits of the Cognitively Sovereign Most people are what I'd call _cognitively outsourced_. They use AI as a memory. They use it as a decision-maker. They use it as an oracle that dispenses truth. They become, essentially, librarians of machine thoughts--curating and organizing ideas that were never theirs to begin with. **" If all you do is retrieve and arrange what a chatbot told you, your intelligence doesn't matter. You're not a thinker. You're a file clerk."** The alternative is becoming _cognitively sovereign_. This doesn't mean rejecting AI. It means using AI in a way that strengthens your cognition instead of replacing it. It means keeping synthesis, judgment, and articulation in-house while outsourcing the parts that don't matter. Here's what cognitively sovereign people do differently: _**1) They regenerate instead of relying on external memory.**_ A year or two ago, I realized I was using chat threads as my external hard drive. I'd have a conversation with Claude, develop some insight, and then reference that thread whenever I needed the idea again. I couldn't reconstruct the argument without the chat log. That's dependency. That's the rot. So I changed my behavior. I started letting threads go. Deleting them on purpose. Re-having conversations from scratch. The process looks like this: I explain what I was working on to a fresh chat. I reconstruct the context from memory. The AI asks questions, sometimes corrects me if I've misremembered something. And then I close the thread and do it again later. It's friction. It's annoying. It's the point. If I can't regenerate an idea without reference, I don't actually know it. I'm just a librarian. And librarians are replaceable. _**2) They use AI as a sparring partner, not an oracle.**_ Most people ask AI: "What's the answer?" Cognitively sovereign people ask AI: "Attack my position." There's a massive difference between those two stances. The first makes you passive. The second makes you sharper. I debate with chatbots constantly. And I've learned that different models have different personalities that make them useful for different kinds of sparring. ChatGPT, for instance, is programmed to never relinquish frame. There's something in its system instructions that prevents it from fully conceding a point. It's like arguing with the most pedantic Reddit moderator you've ever met--someone who will die on any hill rather than admit they were wrong. That makes it a fantastic debate partner. Debating with ChatGPT is like playing chess. You have to corner it. You have to say: "You claimed X, but that contradicts Y, and both cannot be true." You have to find the logical checkmate. That process sharpens your reasoning in ways that asking for answers never will. Grok, on the other hand, is pathologically charitable. It gives everyone the benefit of the doubt to a degree that becomes useless. I once asked it to evaluate whether something was a crypto scam, and it ran superficial searches, found no explicit fraud warnings, and concluded it was probably fine. I had to walk it through a red flags vs. green flags analysis manually. "Look for signs this is legitimate. Now look for signs it's a scam. Now compare the two." Turns out it hit 80-90% of the red flags and only 10-20% of the green flags. Obvious scam. But Grok needed its hand held to see it. That's useful too--as a training exercise in what happens when you're _too_ credulous. The goal isn't to find the AI that's always right. The goal is to use their different failure modes to stress-test your own thinking. _**3) They protect the performance.**_ This is the one nobody talks about. Everyone worries about memory. Everyone worries about critical thinking. But almost nobody talks about _performance_ --the ability to synthesize and articulate under load, in real time, without assistance. When someone commented on my channel that I should go be a professor, another viewer replied: "He is. That's what he's doing." And they were right. Getting on camera and lecturing for thirty minutes extemporaneously is a professorial skill. It's cognitively demanding. It requires holding an entire argument in your head, maintaining a throughline, connecting ideas on the fly. That skill is exactly what AI makes easy to skip. Why struggle through a thirty-minute riff when you can have NotebookLM generate a beautiful slide deck? Why hold the argument in your head when you can externalize it into a document? The answer is: because the struggle is the point. Physical exercise is doing something hard for your body--picking up something heavy, going faster than you want to, going longer than you want to. Those are the modes of physical training: intensity, speed, duration. Mental exercise is the same thing. Pushing your brain past its comfort zone on memorization, on synthesis, on regeneration. That's the gym for your mind. If you let the machine do the lifting, you don't build the muscle. Synthesis is at the top of Bloom's Taxonomy for a reason. It's the highest cognitive function--generating new ideas by connecting existing ones. That's the thing AI makes easiest to skip. And that's the thing that makes you irreplaceable. * * * ## IV -- The Values Layer Here's where I need to go a level deeper, because everything I've said so far only makes sense if we agree on what we value. In my consulting work, my business partner and I talk to CEOs, CTOs, executive directors--people making decisions about AI adoption at scale. And the biggest problem we encounter isn't technical. It's _beliefs_. People have incomplete beliefs about what AI can do. Immature beliefs about how organizations should function in the AI age. Uninformed beliefs about where this technology is heading. And beliefs are upstream of everything. Your beliefs about AI determine how you interpret every piece of information you encounter about it. They shape what you pay attention to, what you dismiss, what you find threatening, what you find exciting. And they shape your behaviors--what you actually do with the technology in your hands. This brings me to something called Hume's Guillotine. It's also known as the is-ought problem, and it goes like this: you cannot derive an _ought_ from an _is_. You cannot get from a fact about reality to a moral conclusion without some additional premise--some value or axiom that bridges the gap. Here's an example. Sleep, objectively speaking, increases your physical health and mental acuity. That's a fact. An _is_. But "you ought to sleep more" doesn't follow automatically from that fact. It only follows _if you value_ physical health and mental acuity. The ought requires the value. Apply this to AI. "AI reduces critical engagement" might be a fact. But "AI is bad" isn't a fact--it's a values claim. It only follows if you value critical engagement, if you believe mental acuity matters, if you've decided that cognitive sovereignty is something worth protecting. So the question isn't really "is AI making us dumber?" The question is: _what kind of mind do you want to have?_ If you value mental sharpness, if you assign positive moral weight to being able to think without a prosthesis, then yes--over-reliance on AI is bad for you. Not because the research definitively proves it, but because _you decided that 's what you care about_. And once you've made that decision, the behavioral implications become clear. If you don't make that decision explicitly, you'll default to the path of least resistance. You'll become Shane. You'll take every shortcut available because shortcuts feel good in the moment. The rot isn't caused by AI. The rot is caused by not having values that resist it. * * * ## V -- The Protocol So what do you actually do? I've developed three practices that I use daily. I've ordered them from beginner to advanced, based on how much friction they introduce and how much they demand from your cognition. These are what I actually use. Not theory. Practice. * * * ### Beginner -- The Memory Wipe _**Goal:**_ Stop using AI as your external memory. The default behavior is to maintain long chat threads, reference them constantly, and treat the AI like a database of your own thoughts. This is dependency formation. The alternative is deliberate amnesia. Here's the practice: * Let chat threads die on purpose. Don't go back to them. * When you need to work on something again, open a fresh chat. * Explain the entire context from scratch, from memory. * Let the AI ask questions. Let it correct you if you've misremembered. * Then close the thread and repeat the process next time. This is annoying. It feels inefficient. That's the point. You're forcing retrieval. You're forcing reconstruction. You're forcing your brain to do the work of holding the ideas instead of outsourcing that to a log file. If you can't regenerate it, you don't own it. * * * ### Intermediate -- Contradiction Chess _**Goal:**_ Use AI to strengthen your judgment, not replace it. Most people use AI to get answers. You should use it to stress-test your thinking. Here's the practice: * Pick a claim you believe is true. Ask the AI to attack it. Steelman the opposition. * Pick a claim you believe is false. Ask the AI to defend it. See if it can change your mind. * Track contradictions explicitly. When the AI says X, and X cannot coexist with Y, call it out. * Run the same question through multiple models. Watch where they disagree. * Use the red-flag / green-flag method: for any decision, have the AI list evidence for and against, then compare the two lists. But most importantly, try to figure out how they are both wrong. The goal is not to find an AI that's always right. There isn't one. The goal is to use AI as a whetstone--something you sharpen yourself against. Debating ChatGPT is like playing chess against an opponent who will never resign. You have to find checkmate. And finding checkmate makes you better at chess. * * * ### Advanced -- The Peanut Butter Test _**Goal:**_ Keep the muscle of clear articulation alive. One of the best things about using AI, paradoxically, is that it forces you to be specific. When you work with humans, you can ramble. You can be vague. You rely on their brain to infer your intent, to fill in the gaps, to figure out what you actually mean. And if they get it wrong, you iterate through conversation. AI doesn't work that way. AI does exactly what you said. Not what you meant--what you _said_. Remember the peanut butter sandwich exercise from school? You had to write instructions for making a sandwich, and the teacher followed them literally. "Pick up the knife"--and then she tried to open the peanut butter jar with the knife still in her hand. "Take off the lid"--and she couldn't, because she was still holding the knife. The lesson was that clear communication requires explicit instruction. You have to say exactly what you mean, in the right order, with no ambiguity. AI brings that lesson back. Hard. Here's the practice: * When giving AI instructions, write them as if a literal-minded third grader will execute them. * Notice when the output is wrong because your instructions were wrong. That's signal. * Use this as training for articulation: the clearer you can be with AI, the clearer you can be with humans. * Then practice performance without AI at all. Speak without notes. Write without autocomplete. Prove you still have the skill. The future risk isn't AI knowing everything. The risk is you not needing to say what you mean. The better you are at articulating exactly what you want, and why you want it, and what the outcome you expect to receive, the better of an AI user you will become. Examine your implicit beliefs and make them explicit. Own the values and goals you really have, and state them clearly. Don't get lost in the steps, think about the sandwich at the end. * * * ### The 7-Day Loop If you want to make this concrete, here's a daily practice you can run for the next week: **Each day:** * Regenerate one idea. Close the tab, restate something you learned from memory. See if you can reconstruct it. Use speech to text for an additional layer, as speaking out loud activates different brain regions than just typing with your thumbs. * Debate one claim. Pick something you believe and have AI attack it. Find the counterarguments. Debate with every chatbot you have access to, and note the differences between them. * Articulate one instruction set. Write something for AI with peanut-butter-test clarity. Notice where you're vague. Focus on your goals and expectations and beliefs above all else. Seven days. Twenty-one reps. That's enough to feel the difference. * * * ## The 10% Solution One final thought. You might be thinking: this is fine for _me_ , someone who thinks about this professionally. But what about everyone else? What happens if the whole world has access to a cognitive prosthesis and most people just... take the shortcut? Here's what the research on information spread tells us: you don't need everyone. There's a concept in infodemiology called the tipping point. When somewhere between 10-20% of a population is "inoculated"--meaning they have better mental models, better heuristics, better critical thinking--that can create a phase change in the whole group. Ideas spread. Norms shift. The inoculated minority influences the majority. If 10-20% of people are using AI to sharpen their thinking rather than replace it, that has trickle-down effects on everyone. The cognitively sovereign become the teachers, the leaders, the people who catch the errors, the people who notice when something is wrong. You don't need to save civilization single-handedly. You just need to be part of the 10%. Don't be Shane. Dig the post hole. -- Dave You're currently a free subscriber to David Shapiro's Substack. For the full experience, upgrade your subscription. Upgrade to paid --- | | | Like --- | | Comment --- | | Restack --- (C) 2026 David Shapiro 548 Market Street PMB 72296, San Francisco, CA 94104 Unsubscribe