// writing

2026-06-01 // 8 min // ai, cognition, learning

the reps you stop taking

I caught myself reaching for the LLM before I’d thought about the problem. Not after getting stuck. Before. The cursor was sitting in the chat box and I hadn’t even finished loading the question into my own head. Every time I do that I hand over a rep I used to take myself, and those reps are where the thinking happens. Hand enough of them over and there’s less of me doing the thinking at all.

What the LLM takes

I wrote a while back that the brain drives the tool, that the LLM recombines instead of originating and the work stays yours. Here’s the part I left out. The brain doing the driving isn’t a fixed thing. You can lose it, slowly, without noticing, and the LLM is the most comfortable way I know to do it.

Start with what’s at risk, because this is easy to get wrong. It isn’t your judgment. I argued before that the LLM is bad at deciding, at knowing what it doesn’t know, at originating, and none of that has changed, so you keep those, and you’d feel it fast if you tried to hand them over. What’s at risk is the skill underneath the judgment. Writing the code. Writing the argument. Doing the math. Reading something hard enough that you have to fight it. Taking a problem from a cold start with nothing on the screen yet. The LLM does a passable version of every one of those on request, and that is exactly what makes them easy to give away.

The move most people miss is that offloading isn’t one thing. Stay in the loop, steering and checking and correcting, and you can come out sharper, because you’re spending more of your reps on the judgment and fewer on the rote part. The damage comes from leaving the loop, prompting and accepting and moving on.¹ The catch is that the good version and the bad version feel identical in the moment, and the slide from one to the other is invisible. Reviewing is built on top of being able to do the thing. The day you can’t write the code is the day your review turns into nodding along, and nothing tells you it’s happened.

So no, you don’t lose decision-making by delegating decisions. You lose it by delegating everything that fed it. The skill is the soil the judgment grows in, and it’s the only thing that lets you catch the LLM when it’s confidently wrong. Hand it all over and the atrophy doesn’t stop at the skill. It eats the supervisor too.

Run the second-order on it, since that’s the move that turns the trap inside out. You let the LLM write the function. First order: twenty minutes saved, real, banked, yours. Second order is the consequence of that consequence. Do it a thousand times and you can’t write that function anymore, and you can’t reliably review it either, so all of it routes to the LLM, which wears you down faster, and the judgment that was the whole point has nothing left feeding it. The twenty minutes were real. The compounding is the only thing that matters over a career, and you’ve set it running against you.

None of this means do everything the hard way. Hand over the genuinely disposable, the syntax you’d look up anyway, the boilerplate, the thing you’ve done ten thousand times. That clears room. The discipline is knowing which reps are load-bearing for you in particular, and refusing to give those away no matter how badly the LLM wants them.

Not a race

The reflex here is to read all of that as a reason to run faster. Learn quicker, stay ahead, don’t get left behind. That frame is the wrong one, and it’s the anxious one.

The LLM is a compression of a compression, and what Dennett calls competence without comprehension: a system that gets the answer right while understanding none of it.² You’re the only thing in the loop that originates, that decides, that answers for the result when it’s wrong. The LLM can’t become that. So the move isn’t to outrun it. It’s to keep the one thing it can’t be.

That’s why the word is perpetual. Not because the target keeps moving. Because the organ decays without use, which is biology, and was always going to be true. The LLM doesn’t set the pace of your learning. It raises the price of letting yourself go, because now there’s something standing right there, glad to fill whatever space you leave. You keep the brain because it’s yours, and it’s the only one you get.

Keeping the mind

So the upkeep is the real subject. The floor is doing the work you’d be tempted to skip. Write the code. Write the argument. Do the math by hand often enough that the muscle stays. The reps you protect are the load-bearing ones from a few paragraphs back.

On top of that, two jobs, and they aren’t the same one. The first is taking on hard new things: a language, an instrument, a field you know nothing about, a book that’s over your head. Not because you need the Portuguese or the piano. Because being bad at something and slowly getting better is the exact state the LLM offers to spare you, and it’s the state that keeps the whole machine plastic. The second is the meta layer, the skills that make you good at pointing the LLM: judgment, and second-order thinking, the mental model I reach for most. New things keep you able to learn at all. The meta skills keep you fit to drive. They fail in different ways when you neglect them, so they’re worth keeping up on purpose, and separately.

The brain is the body

There aren’t two things. The brain is an organ, thinking is something a body does, and the mind isn’t a layer hovering over the meat that you can keep up on its own. Sleep isn’t good for your thinking. Sleep is part of how the thinking runs. Exercise is maintenance on the actual tissue doing the cognizing. Neglect the substrate and you degrade the brain, because the substrate is the brain.

If that lands as odd in a piece about AI, the oddness is the dualism, not the claim. You can’t protect the irreplaceable organ while starving it, because the organ is physical. And it stays irreplaceable while being physical: this meat originates and decides and answers for itself in a way the silicon doesn’t, so being made of meat doesn’t drag you down to the LLM’s level. It just puts the irreplaceable thing somewhere you can wreck.

The facts back it quietly. Your brain spends every night flushing out the day’s metabolic waste, the gaps between its cells widening by around 60% to let the fluid through.³ Short the sleep and the garbage stays in. A year of regular walking grew the hippocampus by about 2% in older adults and turned back one to two years of its aging, while the ones who sat still kept losing ground.⁴ The rest runs the same way: lift, eat like the food becomes part of the brain (it does), keep your clock regular, each one upkeep of the organ rather than a favor to it.

And it runs both ways, which is the part that should keep you off the ledge. The hippocampus grew back. The substrate answers to maintenance. This isn’t a countdown you’re losing. It’s a maintenance schedule, and you’re holding the wrench.

What I keep

So I keep a list, not written down, of the work I don’t hand over. The first real pass at a hard problem is mine. The architecture call is mine. The sentence that carries the argument is mine. The LLM comes in after, to check the work and draft the parts I already understand cold. The body list is shorter and less negotiable: train, sleep, eat like it counts, because none of the rest survives on a wrecked substrate.

And it isn’t only my problem. The LLM drifts to the average by design. Catching that drift takes someone who still has judgment the average can’t reach, and if everyone hands the judgment over, there’s no one left who can. The fix is the same at both scales, and the only one I control is this: keep the reps.