Humans Are Just LLMs That Overfit
Somewhere between talking to language models and talking to people, I noticed the two were converging — and that I was on the wrong side of the convergence. The more honest description of my own mind is not “thinker” but “model that overfits.” I take a handful of past events, fit them too closely, and extrapolate patterns that were never there. A machine learning engineer would call this overfitting. It sounds like a technical flaw, but in humans it wears the costume of depth.
The realization came from an ordinary human failure. Two people who think at very different levels of context can barely communicate at all. One speaks in high context: every sentence drags years of history into it, every present moment is weighted by everything that came before. The other speaks in low context: the present is the present, and the past is not relevant anymore. To the low-context speaker, the high-context one seems delusional — hallucinating connections, unclear about what is actually happening now. To the high-context speaker, the low-context one seems shallow — living in a permanent present tense with no memory to learn from.
Neither is lying. They are simply trained on different data distributions, and there is no shared ground truth for them to converge on. No amount of explanation reconciles it. You can only recognize the incompatibility and stop forcing the model to generalize outside its training set.
Low-context thinking has real advantages. It is fast, in the Kahneman sense of thinking fast instead of slow. It makes forgiveness easy, because if you forget the past you also forget the grievance. But the cost is symmetrical: you forget kindnesses as easily as offenses. There is a phrase for this — like a nut forgetting its shell. Everything that made you who you are gets discarded once the moment passes.
High-context thinking has its own failure mode, and it is worse than shallowness: madness. If you carry enough history into every judgment, you start seeing structure in noise. You build elaborate theories about why things happened and what they mean, and some of those theories are hallucinations — confident, articulate, completely unsupported by reality. This is precisely what large language models do when they overfit, and precisely what obsessive people do when they cannot stop interpreting.
Which leads to an uncomfortable hypothesis: the pursuit of exceptional understanding is inclined toward madness, while the pursuit of ordinary happiness is inclined toward forgetting. Most people, if you ask what they would create given unlimited money and time, do not want to create anything. They want to enjoy and consume. For a long time that read as inferiority. Now it reads differently: people who are fulfilled — by work, family, friendship, love — are happier than people obsessed with meaning. The obsession with meaning is not a sign of being chosen; it may just be an overfitted model of the world, one that keeps predicting significance that never arrives, and suffering the gap between hope and reality as a result.
I do not know how to resolve this. Reading certain books feels like a point of no return — once you have looked at life as a problem of meaning, you cannot unsee it and go back to contentment. And yet the honest accounting says: maybe nothing distinguishes me from anyone else in character or skill. Maybe the sense of having a unique thing to create is itself a hallucination, the model’s confidence far exceeding its accuracy.
What remains after stripping away the grandiosity is small and concrete: the wish to walk alone with some courage, and to be someone whose existence lets a few other people drop their shame — so they can do the thing only they can do, whatever it is. That may not be significant in any cosmic ledger. But it is at least a prediction that survives contact with the data.
Drafted with AI assistance from my personal journal (2025-10-21), then edited by me. See On AI Assistance.