Suhail Doshi: evaluation

Transcript context

…Yeah, and I was there at YC Demo Day when you first announced it. It was, I think, the last or one of the last in-person ones, at Pier34 in Mission Bay. How do you think about that now when everybody wants to put some of these models in people's machines and some of them want to stream them in, do you think there's maybe another wave of the same problem before it was like browser apps too slow, now it's like models too slow to run on device? Yeah. I mean, I've obviously pivoted away from Mighty, but a lot of what I somewhat believed at Mighty, maybe why I'm so excited about AI and what's happening, a lot of what Mighty was about was like moving compute somewhere else, right? Right now, applications, they get limited quantities of memory, disk, networking, whatever your home network has, et cetera. You know, what if these applications could somehow, if we could shift compute, and then these applications have vastly more compute than they do today. Right now it's just like client backend services, but you know, what if we could change the shape of how applications could interact with things? And it's changed my thinking. In some ways, AI has like a bit of a continuation of my belief that like perhaps we can really shift compute somewhere else. One of the problems with Mighty was that JavaScript is single-threaded in the browser. And what we learned, you know, the reason why we kind of abandoned Mighty was because I didn't believe we could make a new kind of computer. We could have made some kind of enterprise business, probably it could have made maybe a lot of money, but it wasn't going to be what I hoped it was going to be. And so once I realized that most of a web app is just going to be single-threaded JavaScript, then the only thing you could do largely withstanding changing JavaScript, which is a fool's errand most likely, make a better CPU, right? And there's like three CPU manufacturers, two of which sell, you know, big ones, you know, AMD, Intel, and then of course like Apple made the M1. And it's not like single-threaded CPU core performance, single-core performance was increasing very fast, it's plateauing rapidly. And even these different companies were not doing as good of a job, you know, sort of with the continuation of Moore's law. But what happened in AI was that you got like, if you think of the AI model as like a computer program, like just like a compiled computer program, it is literally built and designed to do massive parallel computations. And so if you could take like the universal approximation theorem to its like kind of logical complete point, you know, you're like, wow, I can get, make computation happen really rapidly and parallel somewhere else, you know, so you end up with these like really amazing models that can like do anything. It just turned out like perhaps the new kind of computer would just simply be shifted, you know, into these like really amazing AI models in reality. Yeah. Like I think Andrej Karpathy has always been, has been making a lot of analogies with the LLMOS.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.