Solving the compute crisis with physics-based ASICs

Hello, welcome everyone. It is August 6th, 2025. We're live in active guest stream number 116.1 with returning guest Max Afer, who will be discussing this paper, solving the compute crisis with physics-based basics. Max will go over the paper, talk about a few different things, then I will check in the live chat and read any questions that people have. So looking forward to people's questions. And Max, thank you again for joining. Looking forward to the presentation. All right. Thank you for the introduction. Yes, I'm Max Afer, and I work with normal computing. And as Daniel mentioned, this is my second time on the stream with the Active Inference Institute, the first time I was talking about a paper called Thermodynamic Bayesian Inference, which is an example of one of our algorithms we've been developing in thermodynamic computing. And this time I'm going to be talking about this new perspective article we've written called Solving the Compute Crisis with Physics-Based ASICs. And just a little word about the background on this. This kind of emerged out of an industry workshop that Normal Computing, the company I work at, we hosted at our office in New York, because we had been focusing up to that point on, and still are, on the ideas of thermodynamic computing and applying those to machine learning, and in particular, you know, probabilistic machine learning. But we noticed that there's a lot of other people, both in industry and academia, working on different approaches to unconventional computing. And we figured, well, there must be something that we can learn from each other. So we hosted this workshop. And it turned out, I think, that a kind of more, a larger, like unifying framework started to emerge, that there's a bigger picture of something that these different approaches have in common. Even though, even though when we move beyond the standard digital paradigm we're familiar with, we get kind of a zoo of strange, exotic new components and circuits that we're using, there is still some kind of pattern or structure that we can identify and categorize to think about these things. And so that's kind of where this paper came out of, that us and the other people who were at that workshop thought it was worth writing a paper on. And so I'm going to, after I kind of finish, you know, giving a little bit of the background and overview, I'll kind of scroll through the paper and give more detailed explanations and commentary on it. And also, of course, I encourage, you know, anyone to go and check it out for yourself. So this is a perspective article and we're not going to be presenting new, like, theoretical results in here, per se, or simulation results. But it's more our view of the connections between these different computing paradigms and where we see the field going. But in a nutshell, I guess it's the reason why I think we see so many of these new exotic computing paradigms starting to take off now is because there's kind of a new challenge for computing, which is the amount of compute that's required for AI is just a very large amount of compute and it requires a very large amount of energy. And that's not, you know, not, I guess, a new challenge for computing in a way and that there have always been new algorithms that demanded more compute. But there is something different about it, which I'll get into when I talk about the end of Denard scaling in that we'll have to solve it in a different way than it was solved before. And so I guess that at a high level, you can kind of think of it as instead of just trying to make chips bigger or cram more transistors onto a chip. People are now looking into ways of using a different kind of fundamental structure for computing or replacing, you know, using something else instead of a transistor to do computing logic gates or storing information or taking those transistors we have and using them in a different way, modeling their behavior in a more complicated way. And of…