Intuition Is a Cache

I noticed it at night, after the phone was put aside. That is the hour I am most burnt out from work, and the hour my brain sets off on its own relaxed journey, asking why I do what I do and don’t do what I don’t. One night the question was about a machine-learning training setup that had taken me more than fifteen hours to learn. By then I was using it. The first few times, applying it took two or three hours. Now it takes fifteen minutes, sometimes ten. Same brain, same information. Why does it act so differently now?

The comfortable answer is that there are two kinds of minds. Sprinters and stayers. Flash thinkers and chain thinkers. The ones who see the answer and the ones who build it. Most people quietly decide early which kind they are, and the decision doubles as an alibi.

I no longer believe the pair exists. There is one runner with two states, and the fast state was bought slowly.

You don’t have a thinking style. You have a training history, and intuition is its cache.

In a computer, a cache is where you keep answers you have already paid for. When a computation is expensive, you run it once, store the result, and look it up ever after. The lookup feels like magic because the cost is invisible. It was paid earlier, off the books.

My fifteen hours were the computation: every step by hand, every wrong turn at full price. The two-to-three-hour phase was the cache warming, some steps stored, some still computed by hand, the unglamorous middle that never makes it into stories about talent. The fifteen minutes are the lookup. Call it sixty to one. Nothing about my brain got faster in between. It stopped paying for what it had already bought.

Machines just ran the same experiment on themselves, with the meter running. In late 2024, researchers asked a reasoning model for 2 plus 3. It answered 5. Then it kept going: thirteen separate solutions, every one of them arriving at 5, at roughly twenty times the tokens an ordinary model spends. The habit was worst on the easiest questions, and all the extra rounds added almost nothing to accuracy. The next year, another team built tasks where more thinking time made models measurably worse; given longer, they got distracted by irrelevant details and drifted from sensible starting points into false patterns. By 2026 the shape was mapped. Past a point, a model starts abandoning answers it already had right, and the useful length of a thought depends on how hard the problem actually is. The bill went up. The score went down. The papers now have a word for it: overthinking.

The fix the field reached for is the part worth stealing. When longer chains stall, the winning move is not to think harder. It is to run many short, independent attempts and keep the best one. Parallel bursts. The human version is many small actions. The cache fills one feedback event at a time, so it fills at the rate you do things that can be wrong. Ten cheap tries return ten signals. One long deliberation returns none. There is a condition: the action has to return a real signal, or you are writing noise. The pattern holds wherever paying has been measured. In the Berlin study that made deliberate practice famous, the violin students rated best practised in morning bursts of about eighty minutes, several a day, with real breaks between; the group that practised once a day, in the afternoon, was headed for teaching, not the stage.

And the inverse is invisible by design. An idea that was never acted on looks, from the outside, exactly like an idea never had. Whoever stays in the sequence forever leaves behind no lost knowledge. It never became knowledge at all.

I say this as a steady procrastinator, not a reformed one. I know exactly why the sequence zone holds people: it is fun. A thought experiment costs nothing to start, feels like work while it runs, and leaves a real tiredness behind when it ends. The tiredness gets treated with a screen, and the screen returns no signal. That is the whole trap in one evening. Heavy thinking, then no action, then the phone. Budget spent, nothing written to the cache. And the longer I sit in that mode, the higher the price of starting anything feels. I ran that loop for three months before this letter existed: thinking about writing, reading about writing, tired from neither. Then I published once. One cheap try that could go wrong. It returned more signal than the three months combined, and it turns out I enjoy this.

Here is the ledger I live on now. The fifteen minutes a burst costs me today were bought by fifteen hours, once. The hours those bursts free up go into the next fifteen-hour purchase. Bursts buy time. Time buys sequences. Sequences fill the cache. The cache makes better bursts. Read it backwards and you get the procrastinator’s ledger too: no small tries, so no time freed, so no sequence ever finishes, so the cache stays empty, so everything stays expensive forever.

These days my tries have a name. I am building an AI consultation and visual agency called Cows Can Fly, which is a long sequence of small bets that get better each time I create more cache. Working next to these models every day is what convinced me the cache matters more now, not less. A model’s cache is filled from text, from nearly everything everyone has ever written, and it is identical for anyone who can pay for tokens. Mine is filled from consequences I personally bore. That supply is small, slow, and not for sale, which is exactly why it is worth something.

Tonight the phone will go down again, and the questions will start again. One of them finally has an answer. The fifteen hours were not slow learning; they were the price. The fifteen minutes are not talent; they are the receipt. You don’t have a thinking style. You have a training history, and the only way to add to it is the next thing that might be wrong.

Researched and drafted with AI assistance; the ideas, positions and final edits are mine.