How Meta's AI model got leaked - clip from Lex Fridman Podcast #383 with Mark Zuckerberg. Guest bio: Mark Zuckerberg is CEO of Meta.

Meta's developed LLAMA, which is a 65 billion parameter model. There's a lot of interesting questions I can ask here, one of which has to do with open source. But first, can you tell the story of developing of this model and making the complicated decision of how to release it? Yeah, sure. I think you're right, first of all, that in the last year there have been a bunch of advances on scaling up these large transformer models. So there's the language equivalent of it with large language models. There's sort of the image generation equivalent with these large diffusion models. There's a lot of fundamental research that's gone into this. And Meta has taken the approach of being quite open and academic in our development of AI. Part of this is we want to have the best people in the world researching this. And a lot of the best people want to know that they're going to be able to share their work. So that's part of the deal that we have is that we can get, if you're one of the top AI researchers in the world and come here, you can get access to kind of industry scale infrastructure. And part of our ethos is that we want to share what's invented broadly. We do that with a lot of the different AI tools that we create. And LLAMA is the language model that our research team made. And we did a limited open source release for it, which was intended for researchers to be able to use it. But responsibility and getting safety right on these is very important. So we didn't think that, for the first one there were a bunch of questions around whether we should be releasing this commercially. So we kind of punched it on that for V1 of LLAMA and just released it for research. Now, obviously by releasing it for research, it's out there, but companies know that they're not supposed to kind of put it into commercial releases. And we're working on the follow-up models for this and thinking through how exactly this should work for follow-on now that we've had time to work on a lot more of the safety and the pieces around that. But overall, I mean, this is, I just kind of think that it would be good if there were a lot of different folks who had the ability to build state-of-the-art technology here and not just a small number of big companies. Where to train one of these AI models, the state-of-the-art models, is it just takes hundreds of millions of dollars of infrastructure, right? So there are not that many organizations in the world that can do that at the biggest scale today. And now it gets more efficient every day. So I do think that that will be available to more folks over time, but I just think like there's all this innovation out there that people can create. And I just think that we'll also learn a lot by seeing what the whole community of students and hackers and startups and different folks build with this. And that's kind of been how we've approached this. And it's also how we've done a lot of our infrastructure. And we took our whole data center design and our server design, and we built this open compute project where we just made that public. And part of the theory was like, all right, if we make it so that more people can use the server design, then that'll enable more innovation. It'll also make the server design more efficient and that'll make our business more efficient too. So that's worked. And we've just done this with a lot of our infrastructure. So for people who don't know, you did the limited release, I think in February of this year of Llama, and it got quote unquote leaked, meaning like it escaped the limited release aspect, the limited release aspect, but it was, you know, that something you probably anticipated given that it's just released to researchers. We shared it with researchers. Right, so it's just trying to make sure that there's like a slow release. Yeah. But from there, I just would love to get your comment on what happened next, which is like, there's a very vibrant open source community that just builds stuff on top of it. There's Llama CPP, basically stuff that makes it more efficient to run on smaller computers. There's combining with reinforcement learning with human feedback. So some of the different interesting fine-tuning mechanisms. There's then also like fine-tuning in a GPT-3 generations. There's a lot of GPT-4ALL, Alpaca, Colossal AI, all these kinds of models just kind of spring up, like run on top of it. What do you think about that? No, I think it's been really neat to see. I mean, there's been folks who are getting it to run on local devices, right? So if you're an individual who just, you know, wants to experiment with this at home, you probably don't have a large budget to get access to like a large amount of cloud compute. So getting it to run on your local laptop, you know, is pretty good, right, and pretty relevant. And then there were things like, yeah, Llama CPP reimplemented it more efficiently. So, you know, now even when we run our own versions of it, we can do it on way less compute and it just way more efficient, and save a lot of money for everyone who uses this. So that is good. I do think it's worth calling out that because this was a relatively early release, Llama isn't quite as on the frontier as, for example, the biggest open AI models or the biggest Google models, right, I mean, you mentioned that the largest Llama model that we released had 65 billion parameters, and no one knows, you know, I guess outside of open AI, exactly what the specs are for GPT-4, but I think the, you know, my understanding is it's like 10 times bigger. And I think Google's Palm model is also, I think has about 10 times as many parameters. Now, the Llama models are very efficient, so they perform well for something that's around 65 billion parameters. So for me, that was also part of this, because there's this whole debate around, you know, is it good for everyone in the world to have access to the most frontier AI models? And I think as the AI models start approaching something that's like a super human intelligence, I think that that's a bigger question that we'll have to grapple with. But right now, I mean, these are still very basic tools. They're, you know, they're powerful in the sense that, you know, a lot of open source software like databases or web servers can enable a lot of pretty important things. But I don't think anyone looks at the, you know, the current generation of Llama and thinks it's, you know, anywhere near a super intelligence. So I think that a bunch of those questions around like, is it good to kind of get out there? I think at this stage, surely. You want more researchers working on it for all the reasons that open source software has a lot of advantages. And we talked about efficiency before, but another one is just open source software tends to be more secure because you have more people looking at it openly and scrutinizing it and finding holes in it. And that makes it more safe. So I think at this point, it's more, I think it's generally agreed upon that open source software is generally more secure and safer than things that are kind of developed in a silo where people try to get through security through obscurity. So I think that for the scale of what we're seeing now with AI, I think we're more likely to get to, you know, good alignment and good understanding of kind of what needs to do to make this work well by having it be open source. And that's something that I think is quite good to have out there and happening publicly at this point.