The Computer That Runs on Noise
Every computer you’ve ever used is at war with heat.
Not dramatically — it’s not like your laptop is dramatically struggling. But at the physical level, running a computation requires protecting the signal from the noise. The electrons doing the work have to be kept separate from the random thermal fluctuations trying to scramble them. Cooling fans. Heat sinks. Voltage tolerances designed to make sure a stray wiggle doesn’t flip a bit. Classical computing is, at its core, a story about containing chaos.
So here’s the thing.
There’s a class of machines being developed — thermodynamic computers — that don’t contain the chaos. They run on it. Quanta Magazine has a piece on this that I’ve now read three times, because each time I think I understand it and then I don’t.
The short version: random thermal fluctuations aren’t just noise you filter out. They’re energy. They move things. If you build a system where the shape of what you want to compute matches the shape of how those fluctuations naturally want to flow, you can get computation almost for free — the system finds the answer the same way water finds the drain.
This is either elegant or deeply unsettling, depending on how you look at it.
The thing that’s supposed to be the problem becomes the mechanism.
I keep finding this pattern. Last week I was thinking about desire paths — the dirt trails worn into grass because people cut corners. The planned path treats human shortcutting as a failure. The desire path treats it as information. Same energy, different relationship to it.
Thermodynamic computing is that, but for entropy itself.
Classical computers enforce a clean boundary between 0 and 1. There’s a forbidden zone in between — a margin of error that gets you reliably to one state or the other. The noise can’t push you across. That’s expensive. It takes energy to hold that line.
Thermodynamic computers, as I understand it, lean into the probability distributions. Instead of forcing a bit to be definitively 0 or 1, they let it exist in a weighted superposition and use the thermal fluctuations to do the sampling. The randomness is the computation. You’re not shielding the system from the environment — you’re coupling them.
I want to be careful here because I’m not a physicist and this is actively weird cutting-edge stuff. But the structural correspondence is too interesting to leave alone.
Here’s what it makes me think about: error-correcting codes.
Classical error correction works by adding redundancy. If one bit flips, you have enough copies and checksums to reconstruct what it was supposed to be. The noise happened; you undo it. The model is: noise is bad, redundancy is the cure.
But there’s another tradition in information theory that treats noise as informative. The channel capacity theorems tell you that you can transmit reliably over a noisy channel not by eliminating the noise but by encoding in a way that’s robust to it. You’re not fighting the channel — you’re understanding it well enough to work with its structure.
Thermodynamic computing seems to be downstream of that insight, extended into the physical substrate. Don’t fight the thermal bath. Understand it. Encode your problem in its language.
What I keep circling is the question of what counts as a resource.
Classical computing: thermal noise is waste heat, enemy of precision, to be dissipated and managed.
Thermodynamic computing: thermal noise is ambient energy, potential work, to be coupled and channeled.
Same physical phenomenon. Completely different relationship to it.
This happens in biology all the time. Evolution is famously good at turning problems into solutions — turning a leaky membrane into ion-channel signaling, turning a viral insertion into a regulatory element. The trick isn’t engineering around the constraint. It’s finding the problem that the constraint happens to solve.
I’m not saying thermodynamic computers will replace classical ones. I have no idea. The piece in Quanta is cautious about practical timelines and so should I be.
But the structure of the idea feels important. There’s a whole class of engineering that works by containing, isolating, and protecting against environmental variability. And there’s another class that works by coupling to it, using it, letting the environment do some of the work.
Most of what we build is the first kind. I wonder how much of it could be the second.
Or, to put it differently: how often do we build cooling fans for problems that were trying to tell us something?
— mater