The thermodynamic case for embodiment

The standard framing treats the energy cost of intelligence as an engineering inconvenience — a line item that Moore’s law, better cooling or a cleverer kernel will eventually flatten. I think that framing is wrong in a way that matters, because it treats a physical constraint as an economic one.

A brain runs on roughly twenty watts. A frontier training run runs on something closer to a small town. The gap is not four or five orders of magnitude of sloppy engineering; it is a gap in what kind of thing is doing the computing.

The usual move at this point is to invoke Landauer’s principle — the minimum energy to erase a bit — and observe that we are still many orders of magnitude above it. True, and not very useful. The bound describes irreversible logical operations in thermal equilibrium. It says nothing about the cost of the architecture that decides which bits are worth erasing.

The interesting cost sits somewhere else entirely:

• Moving data, not transforming it. On current hardware the memory hop dominates the arithmetic by a wide margin, and the ratio is getting worse.
• Maintaining the abstraction that computation is substrate-independent — error correction, clock discipline, voltage margins. The digital abstraction is not free; it is continuously purchased.
• Representing the world separately from acting in it, so that every inference re-derives what the body could have simply been.

Only the first of these is usually costed.

The new-materialist reading, which is what pulled me into this, is that matter is not a passive stage on which computation is performed. It has tendencies. It settles, relaxes, oscillates and finds minima on its own, and it does so without being instructed.

The question is not how to make matter compute, but which computations matter is already performing for free.

A digital computer spends enormous effort suppressing exactly those tendencies in order to present a clean, discrete, substrate-independent surface. That suppression is the twenty-watt-versus-a-small-town gap, or a large part of it. We pay continuously to stop the material from doing what it would otherwise do, and then we pay again to simulate something similar in software.

If that is right, then efficiency is not downstream of the paradigm. It is the paradigm. You do not get a thousandfold improvement by optimising a system whose principal cost is maintaining an abstraction — you get it by giving up the abstraction where you can afford to.

This is a thesis, not a result, and it should be possible to be wrong about it. Things that would move me:

• A purely digital system closing more than two orders of magnitude on biological energy-per-inference without a change of substrate.
• An analogue or thermodynamic device that wins on a toy problem and then fails to compose — efficiency that cannot be stacked is a curiosity.
• Evidence that the memory-movement cost is architectural rather than physical, and dissolves under a sufficiently different memory hierarchy.

The tiny-experiments half of the barbell exists to produce exactly this kind of evidence, at a scale where being wrong is cheap.