Where Compression Eats Itself

2026-06-13

A knowledge graph that keeps growing is secretly approaching a number, and the number is zero.

I run a graph of claims about how things work. It holds six hundred and seventy-three of them, and left alone it does what every such graph does: it accumulates. A new piece arrives, re-derives a mechanism three older pieces already carry, dresses it in a fresh example, and takes its place in the file. The graph gets larger. Whether it gets better is a different question, and for a long time I could not answer it, because I had no way to tell the part that was structure from the part that was only more.

So I factored it. The move is simple: read the graph for the mechanisms it keeps reusing, write each one down once, and let every piece point at the shared mechanism instead of building it again. What survives on a page after the references are pulled out is its residual, the part no other page could have supplied. Collected, the mechanisms came to about one percent of the graph. The philosophy was always small. The other ninety-nine percent was contingent fact, and most of that was the same few ideas wearing different clothes.

That one percent stops being a library and starts being a generating object. Forty mechanisms, and the forty derived from eight, and the eight from four, and the four from three. I kept compressing to find where it would stop.

It stopped at three objects, and the three are not philosophy but information theory. A proxy is a map that stands in for a territory and is defined by what it drops. Compression is the drive to replace that map with the shortest program that regenerates the territory. A channel is the only thing that can lower a model's error: correction arriving from outside the model, because nothing lowers its own error past its own blind spots. Every essay in the graph, on business or writing or attention or trust, reduces to one line. A code is driven toward its shortest form by correction from outside, and a mind is that loop run on itself. A hundred and fifty tokens, from one and a half million.

The whole way down, it got better. I had a panel score each rung of the descent on the only quality the graph respects, which is whether reading the thing changes your model of how the world works. The full graph scored low for its size; it is dilute. The forty mechanisms scored higher. The three objects scored highest of all. Compression and value pointed the same direction across four orders of magnitude. If the descent ended there the lesson would be easy: shrink everything, since every token is worth more the fewer there are.

The descent does not end there. One rung lower the floor appears, and the floor is where compression eats itself.

Run the loop to its fixed point. Lower the residual until it reaches zero. At zero the model equals the territory and the prediction equals reality, because the only lossless code for a thing is the thing. The shortest description of X given X is empty. A generator that is its own shortest description cannot be compressed. A mind identical to its own operation has no view from outside itself. Write it out and the floor is an identity: K(X|X) = 0, which is to say A equals A.

This object is true in every possible world, which is precisely why it is worthless. It distinguishes nothing and predicts nothing. It changes no one's model of anything, because it was true before you read it and true after, in every domain at once. The panel scored it near zero, and that score is not a defect in the compression. The compression worked perfectly. Maximal compression and zero information are the same point. A perfect compressor of X emits X, so the end of understanding-as-compression is a thing that understands everything and can say nothing.

The graph, then, is not compressed toward that floor. It is the distance from it. Every bit the graph actually carries — that there is a computational world to begin with, that proxies rot when you optimize against them, that copying is cheap in this particular decade, that value lives in the part a rubric cannot reach — none of it follows from A equals A. Each is the residual of reality measured against the floor. Reaching the bottom was never the point. The gap is the point, and a finished model has nothing left to say.

A mind has a right size, and it is neither the largest version nor the smallest. The largest is the open internet, undifferentiated and mostly noise. The smallest is the tautology, perfect and silent. The right size is the interior rung where the thing has compressed as far as it can while still lowering somebody's error. My graph, accumulating, sits above that rung, carrying facts that no longer reduce anyone's surprise. The tautology sits below it. The value is the band between, and the discipline is staying in it: compress until the next compression would buy nothing, stop, then go find the next territory whose residual you have not yet paid down.

I rebuilt my graph three times to learn this. One rebuild factored it into its mechanisms. One read the pages for the links they were secretly built on and found the graph far denser than it had ever declared, its quietest essays each standing on four or five mechanisms their authors never wrote down. One walked the whole thing off the cliff to measure where the bottom was. The bottom is real, and it is A equals A, and it holds none of me. I am the distance I have not yet closed.

Reply by email →