Open Source AI is AI we can Trust — with Soumith Chintala of Meta AI

**Alessio** (0:05)
Hey, everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO and resident at Decibel Partners, and I'm joined by my co-host, Swix, founder of Small AI.

**Swyx** (0:15)
Hey, and today we have in the studio Soumith Chintala, welcome.

**Soumith Chintala** (0:17)
Thanks for having me.

**Swyx** (0:18)
On one of your rare visits from New York, where you live.
You got your start in computer vision at NYU with Yann LeCun.
That was a very fortuitous start. I was actually listening to your interview on the Gradient Podcast. So if people want to know more about the history of Soumith, history of PyTorch, they can go to that podcast. We won't spend that much time there. But I just was marveling at your luck. Or I don't know if it's your luck or your drive to find AI early and then find the right quality mentor. Because I guess Yann really sort of introduced you to that world.

**Soumith Chintala** (0:51)
Yeah, I think you're talking about extrinsic success, right? Like a lot of people just have drive to do things that they think is fun. And a lot of those things might or might not be extrinsically perceived as good and successful. I think I just happen to like something that is now like one of the coolest things in the world or whatever.
But if I happen, you know, the first thing I tried to become was a 3D VFX artist. And I was really interested in doing that, but I turned out to be very bad at it.
So I ended up not doing that further. But even if I was good at that, whatever, and I ended up going down that path, I probably would have been equally happy. It's just like maybe the perception of, oh, is this person successful or not might be different. I think after a baseline, your happiness is probably more correlated with your intrinsic stuff.

**Swyx** (1:44)
Yes, I think Dan Pink has this book on Drive that I often refer to about the power of intrinsic motivation versus extrinsic and how long extrinsic lasts. It's not very long at all. Anyway, now you are an investor in Runway, so in a way you're working on VFX.

**Soumith Chintala** (2:00)
Yes, I mean in a very convoluted way.

**Swyx** (2:03)
It reminds me of Ed Catmull. I don't know if you guys know, but he actually tried to become an animator in his early years and failed or didn't get accepted by Disney and then went and created Pixar and then got bought by Disney. And created Toy Story. So you joined Facebook in 2014 and eventually became a creator, maintainer of PyTorch. And there's a long story there you can refer to on the gradient. I think maybe people don't know that you also involved in more sort of hardware and cluster decision affair. And we can dive into more details there because we're all about hardware this month.
Yeah, and then finally, I don't know what else, like what else should people know about you on the personal side or professional side?

**Soumith Chintala** (2:40)
I think open source is definitely like a big passion of mine probably forms a little bit of my identity at this point. I'm irrationally interested in open source.
I think open source has that fundamental way to distribute opportunity in a way that is very powerful, like I grew up in India. I didn't have internet for a while. And in college, actually, I didn't have internet except for like GPRS or whatever. Like, and like knowledge was very centralized, but like I saw that evolution of knowledge slowly getting decentralized, and that ended up helping me learn quicker and faster for like zero dollars.
And I think that was a strong reason why I ended up where I am. So like that, like the open source set of things, I always push regardless of like what I get paid for. Like, I think I would do that as a passion project on the side.

**Swyx** (3:35)
Yeah, that's wonderful. We'll talk about the challenges as well that open source has, open models versus closed models.
But maybe you want to talk, touch a little bit on PyTorch before we move on to the sort of meta AI in general.

**Alessio** (3:45)
Yeah, we kind of touched on PyTorch in a lot of episodes. So we had George Hotz from Tiny Grad. He called PyTorch a CISC and Tiny Grad a RISC. I would love to get your thoughts on PyTorch design direction as far as, I know you talk a lot about kind of having a happy path to start with and then making complexity hidden away, but then available to the end user. One of the things that George mentioned is, I think you have like 250 primitive operators in PyTorch. I think Tiny Grad is four.