Soumith Chintala: belief

Transcript context

…Would a limit on fundability, I'm just observing something like three months of compute, right? That's the top line, that's the like max that you can spend on any one project. But like, I think that's very ill specified, like how much compute, right? I think that the notion of fundability is broader. It's more like, hey, are these family of models within the acceptable set of, you're not crazy or something, right? Even something like neural or DS, which is a very boundary pushing thing or states-based models or whatever. Like all of these things I think are still in fundable territory. When you're talking about, I'm gonna do one of the neuromorphic models and then apply image classification to them or something, then it becomes a bit questionable. Again, it depends on your motivation. Maybe if you're a neuroscientist, it actually is feasible. But if you're an AI engineer, like the audience of these podcasts, then it's more questionable. The way I think about it is, you need to figure out how you can be in the baseline level of fundability just so that you can just live. And then after that, really focus on intrinsic motivation and depends on your strengths, like how you can play to your strengths and your interests at the same time. Like I try to look at a bunch of ideas that are interesting to me, but also try to play to my strengths. I'm not gonna go work on theoretical ML. I'm interested in it, but when I want to work on something like that, I try to partner with someone who is actually a good theoretical ML person and see if I actually have any value to provide. And if they think I do, then I come in. So I think you'd want to find that intersection of ideas you like, and that also play to your strengths. And I'd go from there. Everything else, like actually finding extrinsic success and all of that, I think is the way I think about it is like somewhat immaterial. When you're talking about building ecosystems and stuff, slightly different considerations come into play, but that's a different conversation. We're gonna pivot a little bit to just talking about open source AI. But one more thing I wanted to establish for Meta is this 600K number, just kind of rounding out the discussion, that's for all Meta. So including your own inference needs, right? It's not just about training.…

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