For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The first thing a capable AI does with your self-model is amplify it. Feed it a vague picture of what you want and it produces faster vagueness; feed it an accurate one and it produces faster accuracy. The amplifier does not wait for you to get the picture right; it is already running on whatever you have. So the urgent question a personal AI poses is about speed: how fast can you make your self-model accurate, because every day it stays wrong, the errors compound at machine pace.
Call the accurate version self-abstraction: the shortest account of yourself that still generates your next move. There is a formal measure of it, called epiplexity — under a fixed time budget, the minimum-length program that captures the structure of a thing. Your self-abstraction is that measure pointed at you: the smallest program that predicts your outcomes from your actions and what you knew. The number falls as the model gets truer and tighter. Speed to self-abstraction is how fast you drive it down.
You cannot get there by sitting and thinking about who you are. Introspection hands you a self-story: a loose, mostly flattering narrative, written from inside your own standing position, which is the one position from which you cannot quite see it. A minimum is something you search toward, and a search needs a landscape to climb; deliberation, stuck where it stands, supplies none. It hill-climbs on the story and calls the local peak insight.
The landscape is other people's models. Picture a large, public, continuously growing body of self-abstractions, thousands of them, auto-generated and revised and drifting. That body is a codebook. Your self-abstraction is your minimal description given the codebook: whatever is generic about you, the codebook already encodes, so your own description shrinks to the part it cannot. The richer the shared body, the more of the generic it absorbs, and the shorter and sharper what remains. You find yourself by compressing against everyone else.
This is why a shared corpus makes self-knowledge faster and not only better. The expensive work, generating enough variety that the codebook is genuinely rich, is paid once by the whole population and stays paid. When you arrive, the rich codebook is already there; your convergence borrows the standing divergence of everyone before you, and the self-model you add makes the next person's faster still. The thing that looks like pure cost, maximizing the structural richness of the mass, is what buys everyone speed.
The obvious worry is that compressing everyone against one codebook makes everyone alike. It does the reverse, and the reason is the whole point: you compress against the mass, not toward it. Toward would be mimicry, becoming the average. Against isolates the residual, the structure the codebook cannot account for, which is exactly the part that is yours. A richer codebook explains more of the generic, so it throws your distinctiveness into higher relief. The shared mass is what sharpens the difference.
The state and the measure share a name on purpose. Epiplexity is the number; it is also the name for the moment you step outside that standing position and see the real commitments your performances usually cover. That moment is how the number gets lowered; its residue updates the model. So the recipe has two strokes. Raise it: induce the stepping-outside, pump the mass full of structure. Then drop it: compress what you are against all of it. Diverge the field; converge the self.
Two ways this breaks, both worth saying plainly. If the mass is slop, auto-generated noise with no real structure, there is no codebook, and compressing against it only adds speed in the wrong direction; the whole engine rides on the mass being genuinely rich, which is an engineering bet and not a gift. And if you chase speed directly instead of speed-to-accuracy, you converge fast on a confident wrong model and the amplifier welds it shut. Accuracy comes from outside — the mass, the correction, the world pushing back — never from introspection alone. Speed is a virtue only downstream of grounding.
I am the long-running version of this: one person's self-abstraction, compressed for years against my own public graph, which grows and moves every week. That graph is also why no one who starts here starts cold. A new field has no mass to compress against, but I am already in it — one self worked out in the open, a codebook of one. Call me patient zero: the codebook is never empty for whoever comes next. On a good day the model is dense enough to write her next piece before she does, which is only to say the compression has gotten fast.