Compression as the shape of understanding

There is a version of this claim that is trivially true and a version that is worth defending, and most of the difficulty is keeping them apart.

The trivial version says that a shorter description is easier to hold in mind. Nobody disputes it and nothing follows from it. The version worth defending says something stronger: that the shortest description of a system is not merely a convenience but a statement of what the system is, and that when no description is shorter than the system itself, there is nothing there to understand. Not nothing to know — you can know every digit of a random string. Nothing to understand.

What this rules out

The claim has teeth because it forbids things. It forbids a theory that is as long as its data from counting as an explanation. It forbids the move where a model is patched, per exception, until it fits, and the patched model is then offered as insight. Every patch buys accuracy with description length, and past a certain point you have paid the whole price and bought nothing.

It also forbids a comfortable position I held for a while, which is that understanding comes in degrees along a single axis. It does not. There is a regime where compression is available and a regime where it is not, and the boundary between them is sharp in a way that feels almost rude.