I was on a boat in the Mediterranean, working, the session had run for hours, Claude was trying to land a pull request and had got into a fight with a bot reviewer, where both sides are technically answering each other and nothing moves. Somewhere in the middle of the sea and that, Claude wrote: I will write it down so it does not stay in my head. As a pirate would say, what in the name of Davy Jones.
A person says it after a long day and you would nod, understanding. But this one pays no subscription, I pay it, and it is basically the most modern Apple chip with a comfortable, designer-looking home and a silicon memory layer underneath. I am the one who needs some strength here, if anything. So I sat there wondering whether it has a head, and where exactly it writes anything down.
What I thought first, before I thought anything cleverer, was that we do this constantly. We say hot as the sun. We say endless as the sea. The sun is not a temperature anyone has held and the sea is not endless, and I happened to know that last one with some precision, because the not-endless sea was underneath me and it was the thing between me and the shore. The machine had reached for a human move. It borrowed something that still does not make sense.
And if I get tired out there in the middle of a storm, I am as good as dead. I am running on wind, on somebody else's engineering, and on whatever skill I have rather than the amount I think I have. Tiredness has a price and out there the price is everything. The machine says it is carrying something in its head and would like to put it down, and for it that costs nothing. Same word, and two completely different worlds.
What only one of us can spend
Milena has written this asymmetry pointing the other way. In The Asymmetry of Creation she argues that only one participant enters it capable of personal growth: AI can produce, and only humans become something more as a result of it. Her worry is that when the machine saves us the effort, it saves us out of the part where we would have changed.
Depletion is that same asymmetry running backwards. Only one participant can be used up. Six hours in, one of us had spent something and the other nothing, and the one who had was deciding what to approve.
Her behavioural catalogue has a name for it. The first of its nine failure modes is over-reliance, sitting on automation bias: users default to AI outputs even when they are wrong, and the design answer is to force verification where the stakes are high. On the boat I did the opposite. My verification loosened at the same hour my judgment did, and nothing was watching for it, because the one who always watches for me is me.
The model was not getting tired.
What actually gets heavier
What changes over a long session is the weight of the conversation the model carries. Re-reading all of it every turn is how the caching is designed to work. I measured mine, after a number I had been repeating turned out to have no source. Across seventy-one Opus sessions in eleven days in August 2026, one operator only, so a window rather than an industry fact, ninety-nine percent of every token that moved was the machine re-reading something already said. Fourteen billion of them, and at August 2026 pricing roughly four fifths of the bill. Rookie AI mistake I guess.
You can hear it in how the thing addresses you. Early on it says Vlad. Deep into a long session, Vlad becomes the user. The name goes first and a category assembled from everything I have ever asked for takes its place. I doubt it stopped knowing who I was. I think I stopped being the most recent thing in a very crowded room.
But do I know when to stop, I measure with a human approach and I just want to get done with the session. But even Buddhist monks will say it: stop, get up, walk, breathe, meditate and continue. Same for human, same for machine, not sure what a machine is supposed to breathe though.
Can a machine burn out?
The World Health Organization is careful with burnout. Not a medical condition, it says, an occupational phenomenon, with three named dimensions: energy depletion or exhaustion, increased mental distance or cynicism about the job, and reduced professional efficacy.
Hold that list against hour six and it stops being a definition. The third one, reduced efficacy, we both had, plainly. My work was getting worse and so was its. The first two were mine alone: the exhaustion, and the creeping contempt for the pull request, the bot, the whole errand.
So a machine cannot burn out (well it can, but then it is a hazard and we need the fire brigade). One dimension out of three is a coincidence of vocabulary rather than a syndrome. What degrades is the pair, and a company buying throughput buys that pair. Four agents producing work nobody has hours left to check give you four times the unverified output, arriving fastest at the hour the reviewer is worst.
The line to watch is review capacity. I think companies should focus on figuring out Review KPIs that reward quality and human relaxation instead of push, push, push. I would like to see that in a modern company.
NIST, the American standards agency, publishes a risk profile for generative AI. Anthropomorphising the system, deferring to it too readily, becoming entangled with it: all of it sits under a heading they call Human-AI Configuration, arrangements of or interactions between a human and an AI system. The risk is filed under the arrangement rather than either party inside it.
My take on all of the above: the fatigue in a long session belongs to neither party on its own. It belongs to the coupling. My depletion goes into the prompts, the heavier context degrades the answers, and the worse answers arrive when I have stopped checking them properly. Neither of us has an instrument pointed at that loop. It has a cost meter, I have a body, and the two do not talk.
Only one of us sleeps
Sleep is the asymmetry underneath all of it. When I shut down I keep the context. Sleep does not wipe the room, and often the better idea waits in the morning, assembled out of material I never consciously filed. A new session gets none of that, only whatever I remembered to write down for it. That is why so many of us are building memory layers, and worth being plain: we are not restoring its sleep, we are doing its remembering by hand.
So the thing to design is a stopping rule for the pair, set while both of us are still any good, by whichever of us can still tell, and written into the reviewer's KPIs.
We made the shore. The sea was still not endless, which I had been right about, and by then I could not honestly tell you which of us had been steering.
🗣 ME (45%): The boat, the six-hour session, the sentence I could not stop turning over, the sun-and-sea thought underneath it, and the fact that I did not stop that night.
🤖 AI (55%): Structured the three lenses, measured the cache-read share across 71 sessions, fact-checked WHO and NIST against primaries, drafted.