World machines, not world models

The phrase “world model” has quietly become the thing everyone agrees is missing. It is doing a lot of work for a term whose commitments nobody states. I want to pull one distinction out of it, because I think the two halves point in genuinely different directions.

A world model, in the sense the term usually carries, is a representation held apart from the world it represents. You build it, you keep it, you query it, you correct it against observation. It is a map, and its virtue is fidelity.

A world machine is not held apart. It is a process whose dynamics are coupled to the dynamics it is about, closely enough that running the one tells you about the other. It is not a map. It is closer to a scale model in a wind tunnel, or a tidal predictor made of brass gears — an object whose behaviour stands in for the world’s behaviour because it is made to move the same way.

A map is consulted. A wind tunnel is run. The difference is not fidelity, it is where the work happens.

Both are legitimate. But almost all of the current effort goes into the first, and the first is the one that carries the representational overhead I keep running into.

If your world model is a representation, then every use of it costs an inference: encode the situation, run the model, decode the answer, act. The model is inert between queries and the cost scales with how often you ask.

If your world machine is coupled, the cost structure changes:

• There is no encode/decode boundary to pay for on every use.
• The dynamics run whether or not you are watching, so “inference” is closer to reading a dial than to executing a program.
• Fidelity is bounded by the coupling rather than by the parameter count, which is a very different scaling story.

That last point is the one I keep turning over. Scaling a representation means more parameters and more energy per query. Improving a coupling means a better-matched substrate, which may cost nothing per query at all.

Wind tunnels do not generalise. A brass tidal predictor computes tides and nothing else, and the history of analogue computing is largely a history of beautiful machines that did one thing until a general-purpose digital computer did it adequately and everything else besides.

This is the strongest argument against the whole line of inquiry and I do not think it has been answered. The honest version of the thesis is narrower than the slogan: some computations are worth coupling rather than representing, the set is probably larger than the current split suggests, and the way to find out where the boundary sits is to build small things on both sides of it and see which ones stop being interesting first.

Which is the whole point of the barbell. Theory at one end to say what would be surprising, experiments at the other end small enough that being wrong costs an afternoon.