What on Earth is Extropic Actually Doing
Based Beff Jezos, Guillaume Verdon, won’t shut up about his little startup and at the same time is incredibly confusing and vague about what he and his merry team of “totally cracked physics engineers” are doing. I think I have a sense of it, though, so let’s clear away some underbrush.
First, though, let us take a second to acknowledge that there may be nothing whatsoever there. The amount of tap dancing and smoke, the number of visible mirrors, these are not encouraging metrics.
Second let’s keep in mind the amount of money on the table: $14M is what Extropic seems to have raised to date, almost 5 months ago. This is not a lot of money and a lot of it is gone already. Also, Guillaume keeps doing podcasts instead of writing papers and doing engineering.
Ok, let’s set all that aside and think about what Extropic might actually be doing.
There’s a way of doing neural-networky/machine-learningish stuff called Energy Based Models (EBMs). The details don’t matter much. The underlying mechanic is this, though:
you set up your system by defining a collection of probability distributions according to whatever it is you’re trying to compute. I have no idea what the details are here, don’t ask. The point is that these distributions are customized to the problem.
you sample from these distributions, that is, you generate some random numbers (“values” let’s say, actually, it might just be a voltage or a weight or a temperature, whatever, it doesn’t matter.)
these random values you’ve pulled out of the system feed back into the system to adjust the probability distributions that you’re sampling from.
run this… a lot? the feedback goes in, more values come out, (are massaged?) and they feed back in. The distributions wobble around, pulsate, whatever.
Eventually you can read out something like an answer somewhere.
You can certainly imagine an analog computer being built like this. You need more or less two things:
You need a highly programmable random number (“value”) generator, something where you can adjust the probability distribution it obeys with a fair degree of precision, and speed.
You need a way to sample the random numbers (“values”) and feed them back into the programming of that system. Extropic’s thesis seems to imply that the values are to be fed back in directly, without a digitization step. Pure analog at least in the “inner loop” of “sample and feed back.”
Extropic seems to be working on a variation of a pretty standard random number generator. We’ve had random number generators that sample thermal noise for decades, and they work pretty well.
Note: this has nothing to do with thermodynamics. Thermal noise is variations in voltage that are measurable and which arise because the material has a temperature. Shit is wiggling around inside the silicon or whatever it is, and as a result, you get small random fluctuations in the voltage across this point and that point, if you measure it carefully. Why Extropic insists on talking about thermodynamics at all is a mystery, especially since “thermodynamic computing” is an established term that means something quite different from what Extropic is trying to do. This is one of several red flags.
What Extropic is trying to do is build some electronic widgets that work pretty much like a lot of other hardware random number generators, except also programmable in distribution.
Then they have to wrap that object in a bunch of other electronics to feed the output of these things back into the programming logic, so that feedback can adjust the distributions.
If they can do this, then they can just set up the initial distributions and let the thing run. It will compute answers to problems in roughly the same way that a soap bubble computes a sphere and an artillery shell computes a parabola.
Given the funding situation and the gibberish Guillaume spouts, I speculate that what Extropic is doing with their ~6 months of funding is building 1 or 2 devices to see if they can actually get a degree of programmability in the probability distributions. They’re trying to build a random number generator that accepts some number of electronic inputs that shape the distribution of measured thermal noise voltages in predictable ways, with a sufficient degree of precision to implement EBMs.
I don’t think they’re doing anything else. I don’t think they can be on a $14M seed round.
If they can in fact demonstrate a random number (“value”) generator that is programmable at all, with anything close to the degree of precision they’ll need, I think they can go back to the well for more money. It feels pretty tight to me, though. Even if the engineering is doable, I dunno if they can do it on the budget they’ve got, and there’s no guarantee that they can get enough programmability.
I have no opinion, really, on whether they have any chance at success. I’m not a physicist, all I have to go on is observation of the personalities involved, and I waffle back and forth vigorously between “lol not a chance” and “well maybe.”