Simplistic thermodynamic computing

Since I’m intending to do an experiment involving thermodynamic computing, I thought it sensible to simplistically summarise the paradigm itself. Here’s the compressed version: thermodynamic computing uses thermal noise to drive computation instead of spending resource to fight it, mitigating energy cost as a bottleneck for certain classes of computation at civilisation scale.

Contrast this with the discrete computing paradigm; exquisitely engineered integrated circuits are forced with energy to deterministically generate 0s and 1s. These are stacked horizontally and vertically into abstractions that eventually result in things like words on a glass pane. This is done with great predictability and precision, yielding a wondrous level of computational generality. But it isn’t perfect.

Consider two naturally occurring phenomena we desperately try to replicate with 0s and 1s: randomness and intelligence.

There are entire domains devoted to generating randomness that has parity with what we see in the physical world. But we still can’t match it. The best we can do is harvest physical entropy via lava lamps on a wall and feed their activity into a digital system as a seed. It doesn’t matter how sophisticated the algorithms or techniques are; we’re trying to manipulate a discrete computing paradigm to emulate a continuous phenomenon. It can be done to an acceptable degree for many cases, but it is not perfect and it comes at a high cost.

Intelligence is similar, just at a much larger scale. Like randomness, it may be a fundamentally probabilistic, continuous phenomenon that we’re forcing a discrete substrate to approximate. Consider that a handful of companies are betting a large slice of civilisation on the ability to create general intelligence. It’s taking a huge portion of the world’s capital, an army of the world’s smartest people, massive physical resource pools, and vast amounts of energy. Yet depending on who you talk to and believe, we’re either never going to get to general intelligence or are several years off, even if we continue to sink resource into scaling.

Ultimately, turning noise into information is hard. Discrete computing did it by answering narrower, more tractable questions first.

Other niche computing paradigms have tried to approximate randomness and intelligence via various means, but it hasn’t panned out. Quantum is too expensive and fragile; analog is too narrow and limited. The issue is what’s sometimes called the “demon tax”: making noise usable as information has a cost, and most computing paradigms can’t afford the price that comes with the scale needed for the advanced phenomena they’re trying to achieve.

Every step in that capability ladder is a deliberate act of suppression; each layer exists to force the messy, continuous physical reality of electrons in silicon into a small number of clean, discrete, repeatable states. It works, extraordinarily well. But it means the entire stack must spend energy at every layer to keep the noise out.

Until now. Thermodynamic computing removes the suppression tax throughout the stack, because the stack itself runs on naturally occurring noise.

Think of a container with a ball in it, bouncing around freely and unpredictably. Discrete computing laboriously reshapes the container so the ball’s free movement is cancelled out. It gets pinned in place, forced to sit still and mean one fixed thing. Thermodynamic computing takes the opposite approach: leave the ball free, and set a rule that fires a computation whenever the ball happens to be on the left side of the container due to its own natural movement.

A stopped clock needs no batteries to be right twice a day. Reality keeps moving and eventually enforces the clock’s correctness. Simplistically, thermodynamic computing uses that same free motion to yield 0s and 1s. Which means that what is prohibitively expensive today could soon be orders of magnitude cheaper.