You don’t walk up stairs in real-time. If you did, you’d be slow—each step requiring conscious processing of elevation, texture, muscle tension, and balance. Instead, your brain generates a high-resolution prediction of the next stair before your foot lands.1 1 You only become conscious of it when the simulation fails. When your foot hits floor sooner than expected. Or when you find empty air where a step should be. It’s a prediction error that jolts you back to consciousness until you resolve the error.2 2 When prediction error is low, you’re on autopilot. When prediction error is high, you jolt into consciousness and the model updates. This mechanism explains much about how you actually learn. And most people have the completely wrong model of learning. Most advice on “learning faster” feels right when you read it, then evaporates. Spaced repetition, active recall, deliberate practice—you nod along, try it for a week, then forget why it mattered. The advice doesn’t stick because it’s description without mechanism. It tells you what to do without explaining the machinery that makes it work. what This essay is the explanation. The machinery that makes techniques either compile into real capability or dissolve. What You Think Learning Is What You Think Learning Is Most people think learning is accumulation. Like filling a hard drive. You store information, you retrieve it later. More storage = more knowledge = more capable. accumulation When you learned to walk stairs, your brain carved a model: “Given this sensory input (stair edge), predict: elevation change, required muscle tension, expected sensory feedback.” This is a scheme; a compiled pattern of knowledge that runs automatically. This is a scheme Expert performance isn’t about knowing more. It’s about having schemes with lower prediction error. Chess experts use less neural energy, not more. Their brains expend less effort because they’ve learned what matters.3 3 The novice processes everything with brute-force calculation. When experts look at a chess position, they already know where to focus attention while ignoring irrelevant pieces. The difference between the expert and the novice is the scheme. You can’t shortcut this. This is the shortcut. is You either compile the schemes through iterations, or you don’t have them. The Only Two Ways To Learn The Only Two Ways To Learn People who want to learn “fast” are asking the wrong question. Once you understand how schemes compile, speed becomes irrelevant. You can’t literally fire neurons in your brain faster. There are only two legitimate paths to compile schemes—you’re either solving real problems (where speed emerges naturally from better systems) or you’re doing it for intrinsic reward (where you don’t care about speed at all). 1. You’re Solving A Real Problem 1. You’re Solving A Real Problem Not “learning React.” Shipping a feature to actual users. The errors have stakes. If you’re trying to cure your friend of a disease in medieval times, you don’t read to “finish the book”—you find what you need for a cure and go cure your friend. Deliberate practice isn’t repetition—it’s iterative adjustment toward goals with immediate feedback. Real stakes create this automatically. The developer who shipped 20 products has compiled schemes. The tutorial completionist who’s shipped zero has memes about schemes—they can explain the meaning of a long list of acronyms, but not the forward models that generate working code. memes about schemes Compiled schemes survive because they were built against reality’s feedback, not synthetic examples. If you’re stuck in a prison where only French is spoken, you’ll learn French faster than reading it in your house where no one speaks French. (This doesn’t mean you have travel to another country to learn a language. I’m only illustrating the learning process.) 2. Genuine Curiosity, The knowledge itself is the reward. 2. Genuine Curiosity, The knowledge itself is the reward. When you learned your first language: When you learned your first language: Your mom says “cup,” points at object. The word maps directly to your scheme of that cylindrical thing. When you learn a second language through apps: When you learn a second language through apps: You encounter “Bonjour.” It translates to “Good morning.” which maps to your scheme of morning sunlight, the taste in your mouth when you wake up and all the dimensions of the scheme of “morning”. You’re thinking in English and translating. The forward model runs in English, not French. This is proxy-chaining: using one language as an intermediary to access the schemes, rather than mapping words directly. The neural pathways are measurably different: Direct mapping: Language activates schemes directly (Meme => Scheme) Proxy-chaining: Language activates translation networks first, then schemes (Meme => Meme => Scheme)4 Direct mapping: Language activates schemes directly (Meme => Scheme) Direct mapping: Proxy-chaining: Language activates translation networks first, then schemes (Meme => Meme => Scheme)4 Proxy-chaining: 4 Of course, you can eventually switch from a proxy to the actual scheme if you continue learning. Play piano because music is enjoyable. Code because building is fun—not because you’re checking off “learned Python” or “I code cpp.” When the activity itself is rewarding, you run continuous high-variance schemes driven by curiosity. Ideally, you get both: solving real problems that matter while the process itself is rewarding. That’s when schemes compile fastest. How Schemes Become Memes (And Vice Versa) How Schemes Become Memes (And Vice Versa) A meme is a compressed description of a scheme. When you convert a scheme into a meme, you lose almost everything: Scheme => Meme (Compression with loss): Scheme => Meme (Compression with loss): You have a compiled scheme for playing guitar. Your fingers know where to go for a minor seventh chord—you predict the finger positions, pressure, and resulting sound automatically. Someone asks: “How do you play a minor seventh?” You produce a meme: “It’s the root, minor third, fifth, and minor seventh.” That meme describes the structure but contains none of the forward model. The person hearing it now has a description, not the entire system. They can’t play the chord from that meme alone. They have to compile their own scheme through repeated iterations—placing fingers, hearing wrong notes, adjusting until the model predicts correctly. Meme => Scheme (Decompression requires work): Meme => Scheme (Decompression requires work): You read: “Use spaced repetition for memory retention.” That’s a meme. It compresses someone else’s scheme (their system that knows when to review, how spacing affects consolidation, what constitutes retrieval practice) into a short phrase. when how what To convert that meme into a scheme, you must run iterations, explain why it works or fails, improve with a variation and repeat. Most people stop at the meme. They collect descriptions: “spaced repetition works,” “deliberate practice matters,” “sleep consolidates memory.” These are true memes—they point at real schemes—but without doing the decompression work, you’re just moving symbols around. Why experts struggle to teach: Why experts struggle to teach: Experts have schemes. Students need schemes. But transmission happens through memes. The expert’s forward model is so compressed and automatic they can’t fully articulate it. When they try, they produce memes that lost most of the information. “Just feel the rhythm.” “Make it pop.” “Keep your wrist loose.” These memes are correct but they’re uselessly compressed for someone without the scheme. The student needs thousands of iterations to decompress those memes into working forward models. correct This is why you can’t learn to play jazz by reading about jazz (although it helps). Why tutorial hell doesn’t produce developers. You’re collecting memes about schemes, not compiling schemes. In every domain, the scheme is a system calibrated through thousands of iterations. Collecting memes gives you the vocabulary of competence. Compiling schemes gives you actual competence. This is why tutorial hell fails. Why most “learning” is just meme collection. Most “Learning” Doesn’t Compile Anything Most “Learning” Doesn’t Compile Anything Most learning fails because it’s neither problem-driven nor immersion-driven. Tutorial hell: No real stakes, no genuine immersion (you’re following instructions, not exploring). You’re learning “how to follow tutorials.” Tutorial hell: Studying for exams: Optimising for a synthetic metric that doesn’t map to real-world iterations. You’re compiling “how to pass this exam,” not “how to think in this domain.” Studying for exams: Courses you “should” take: No intrinsic reward (it’s obligation, not curiosity), no real problems to solve. Your scheme never engages. You’re moving memes around without building forward models. Courses you “should” take: The Proxy-Chain Trap: The Proxy-Chain Trap: Some domains are impossible to learn without proxy-chaining first. You can’t directly map your way into quantum mechanics—you need to first understand the intermediary. But mastery is when the proxies dissolve. When you understand quantum mechanics, you’re not translating anymore. You think in the domain itself. If you never make that shift from proxy-chaining to direct mapping, you stay intermediate forever. You can code by translating through Stack Overflow patterns, but you’re not thinking in code. You can speak French by translating from English, but you’re not thinking in French. You can solve problems by translating through framework documentation, but you don’t understand the domain. When you remove the scaffold, does your performance collapse? If yes, you were proxy-chaining. The reason most people stay in “Meme” territory is that compiling is painful. The jolt on the stairs triggers a physiological stress response. compiling is painful To learn, you must intentionally seek out the feeling of missing a step. Tutorials are designed to make you feel like you are walking on flat ground. They eliminate the jolt, which feels good but ensures no model updates ever occur. To learn, you must intentionally seek out the feeling of missing a step. Tutorials are designed to make you feel like you are walking on flat ground. They eliminate the jolt, which feels good but ensures no model updates ever occur. Physiological Constraint Physiological Constraint You cannot compile schemes when your brain cannot execute the computation. People can actively hold roughly 4±1 items in working memory at once5. When your biology degrades through stress, sleep deprivation, or poor nutrition, that capacity shrinks noticeably—sometimes down to 1-2 chunks. 5 At that capacity you cannot: Hold complex problem structures Maintain multiple hypotheses Run counterfactuals Generate novel predictions Hold complex problem structures Maintain multiple hypotheses Run counterfactuals Generate novel predictions Chronic stress impairs working memory and cognitive flexibility, shifting cognition from high-level thinking to survival mode (reacting to immediate inputs with cached responses). No learning happens because no model updates occur. Sleep. Stress management. Nutrition. Environment design. These are prerequisites for your scheming to function. AI Is a Cognitive Prosthetic AI Is a Cognitive Prosthetic It is an extension of your thinking that can run thousands of scenarios on data trained on the internet’s collective intelligence. Your working memory can hold ~4 chunks. Complex problems have a larger number of variables, and we usually compress them. But some problems have variables that were previously incompressible. Now we can compress them because AI can hold more variables. AI can hold arbitrarily many intermediate states while you process them. This is genuine leverage—if you know how to integrate it. A writer uses Claude to hold 20 plot threads and character arcs while focusing on scene-level dialogue. A researcher maintains 15 paper citations and competing arguments while synthesising a novel framework. A developer offloads boilerplate so the entire system architecture stays in working memory. The schemes still compile in your brain. AI extends your working memory so you can handle more complexity. People anthropomorphise the calculation⁶. They imagine a dialogue where there is only a monologue. AI is not a conscious intelligence. It is a computational prosthetic. It is your own brain running on faster hardware. You normally generate lots of random genius and idiotic thoughts every minute and select the best ones. AI extends that so you can think even more—at an exponential rate because the collective intelligence on the internet is working for your goal. When you use the tool to “challenge” your ideas, you aren’t seeking a second opinion. You are using a high-velocity mirror to anticipate weak points your biology was too slow to map. It is your own suspicion, amplified. If the output is sycophantic, your intent is sycophantic. The collective history of human thought it holds is not knowledge until your intent selects it. A prosthetic limb doesn’t know how to walk—it executes the user’s balance. AI executes your search. If you proxy-chain through AI, no schemes compile. When AI is wrong or unavailable, you have nothing. This is identical to learning French through Google Translate. You’re not building direct mappings. You’re building dependence on an intermediary. This matters now because the leverage is huge. This matters now because the leverage is huge. What was impossible before is now hard (with AI assistance). What was hard is now easy. But only if you understand your own knowledge fundamentally and have real stakes that force verification. Otherwise, you just got infinite leverage on your initial thinking. You’ll confidently walk off cliffs at 10× speed. Most people can’t tell the difference because they have the wrong mental model of learning. They think it’s accumulation (store more facts) when it’s build better schemes. So they use AI as a ‘work companion’ instead of working memory extension. They optimise for tutorial completion instead of shipping real things. They confuse memes about domain knowledge with actual schemes that generate results. The gap is widening: People who understand the schemes of their knowledge just got 10× leverage. People who think learning is accumulation just got 10× faster at running towards a cliff. You cannot fake schemes. Reality tests them constantly. The person who learns is the person who cannot hide from being wrong. If you enjoyed this essay and want more, subscribe to my newsletter: https://crive.substack.com https://crive.substack.com References References 1 Predictive processing basics & stairs-like examples: Karl Friston on predictive coding (overview): https://en.wikipedia.org/wiki/Predictive_coding; Andy Clark’s accessible intro: https://www.mind-foundation.org/blog/predictive-coding 1 https://en.wikipedia.org/wiki/Predictive_coding https://www.mind-foundation.org/blog/predictive-coding 2 Prediction errors & consciousness: Friston’s foundational paper on predictive coding/free energy: https://pmc.ncbi.nlm.nih.gov/articles/PMC2666703/ 2 3 Chess expertise & neural efficiency: Systematic review on brain imaging in experts vs. novices: https://www.sciencedirect.com/science/article/pii/S3050642525000326 (experts show efficient, focused patterns) 3 https://www.sciencedirect.com/science/article/pii/S3050642525000326 4 Immersion vs. explicit/classroom: Morgan-Short et al. (2012) – implicit/immersion yields native-like ERP patterns: https://pubmed.ncbi.nlm.nih.gov/21861686/ 4 https://pubmed.ncbi.nlm.nih.gov/21861686/ 5 Working memory ~4 chunks: Cowan (2001) – The magical number 4: https://pubmed.ncbi.nlm.nih.gov/11515286/ 5 https://pubmed.ncbi.nlm.nih.gov/11515286/
