No Priors: Artificial Intelligence | Technology | Startups
00:43:20 · Sarah Guo, Mark Zuckerberg
[00:43:20] S3: Yeah. I mean, that resonates with me as somebody who, you know, talks to and hires a lot of research scientists. They want to know if you have the data, if you have the tools, if you have the compute, if you have the talent, and then what the mission is. And so I actually think that's super competitive. [00:43:34] S0: The other thing is that you don't need a very large team. Right? So I think it's like an interesting thing about the world is that people care about different missions, and that's good. I think that's part of the whole I mean, part of why building these tools and giving people the ability to explore what they care about, whether it's across science or just across everything, is such a powerful way to make progress in society is that people care about different things. And in order to make progress in AI, you don't need many, many hundreds of AI researchers or thousands or anything like that. I think you can really make progress with a very strong group of a dozen or a couple dozen people. And, yeah, mean, finding people who care about this mission is not a particularly hard thing. I mean, this is like a super important thing in the world. So I think that that's yeah, it's just kind of a cool thing about the world is that people obviously are drawn to different missions.
00:43:34 · Mark Zuckerberg, Sarah Guo
[00:43:34] S0: The other thing is that you don't need a very large team. Right? So I think it's like an interesting thing about the world is that people care about different missions, and that's good. I think that's part of the whole I mean, part of why building these tools and giving people the ability to explore what they care about, whether it's across science or just across everything, is such a powerful way to make progress in society is that people care about different things. And in order to make progress in AI, you don't need many, many hundreds of AI researchers or thousands or anything like that. I think you can really make progress with a very strong group of a dozen or a couple dozen people. And, yeah, mean, finding people who care about this mission is not a particularly hard thing. I mean, this is like a super important thing in the world. So I think that that's yeah, it's just kind of a cool thing about the world is that people obviously are drawn to different missions. [00:44:28] S3: So I think the simplest mental models that folks have, even if they're paying attention to the space, are essentially like, okay, you know, structured prediction models for proteins and protein protein interaction models. And then, so there's this one piece, which is fundamental understanding. And then there's this like theory of someday we're just going to be able to like zero shot things into either the clinic or the clinic with much, much better hit rate. What needs to happen for us to go from ESM Fold two to this other piece? Is that feasible?
00:45:01 · Alex Rives, Sarah Guo
[00:45:01] S2: I think that's a great question. I mean, I would say that I'm really optimistic on that. So I think, you know, on the one hand, you know, these are problems that historically, you know, people could spend kind of an entire career working on. Like, how do you how do you figure out how to effectively optimize a drug? How do you get it, you know, get it through preclinical? How do you do the early safety? I think that, you know, when you have a new scientific paradigm, kind of, you know, questions that were once hard kind of become simplified through the new paradigm. And so I'm very optimistic that kind of many of these core problems will be solved kind of in an emergent way through these models. Mhmm. And I think one great example of that is is toxicity. Whereas if if you can kind of really digitally digitally kind of simulate everything and be able to predict, you know, where a drug is going to distribute and bind across the human body. You know, like you kind of have the beginning of a solution to that kind of problem. So I think that once you have these kind of accurate representations at the molecular level, we're going to start to see really rapid progress on a lot of these core problems. [00:46:13] S3: What is the most exciting use or experimentation with the models you've seen in the last week since release?
00:46:13 · Sarah Guo, Alex Rives
[00:46:13] S3: What is the most exciting use or experimentation with the models you've seen in the last week since release? [00:46:19] S2: Yeah, I mean, it's just been great to kind of see it get integrated in all kinds of things. I think one of the really interesting things that we've been seeing is people kind of connecting it with AgenTex systems to just kind of do automated design and kind of just automate that whole process. So it's really, I think, another example of how you can kind of see bringing together AgenTeq and Frontier AI with, you know, the ability to have a world model for biology and actually reason about biology and, you know, really kind of start to automate the entire design process. [00:46:53] S3: How do you decide what the next step in the research agenda is? It's like world model for biology, and then I could I'm just gonna be very coarse here. Like I could scale it up, I could add more data, I could add, like adding data is a nontrivial thing in terms of new methods and domains. Like what is, do you take input from the larger ecosystem about, you know, how people are using it and what would make it more useful? Or is it really like we we understand, like, the next step of structures or coverage that we're looking for? [00:47:25] S2: I mean, I think there's two things. So, like, we have a view on kind of the next big challenge, which I think is, you know, the the virtual cell. Mhmm. And, you know, really being able to kind of ladder up the hierarchy of biological complexity to the cell.
00:47:25 · Alex Rives, Sarah Guo
[00:47:25] S2: I mean, I think there's two things. So, like, we have a view on kind of the next big challenge, which I think is, you know, the the virtual cell. Mhmm. And, you know, really being able to kind of ladder up the hierarchy of biological complexity to the cell. [00:47:39] S3: Sorry. Very basic question. Yes. Virtual cell model. Like, what is the input and output I should expect? [00:47:44] S2: Yeah. I mean, I think there's different views on that. But I think kind of what you ultimately want is a system that can really model each of the levels of complexity. So, you know, the the proteomic layer, the genetic layer, the transcriptomic layer, and connect that to the phenotype. And you need enough generality so that you can ask the model questions about a new intervention in a context that it hasn't been trained on and and kind of get an answer from it. And, you know, the gap that we we need to close as a field is being able to really make those predictions that can generalize. So that's gonna require an enormous effort to generate data.
00:48:27 · Mark Zuckerberg, Elad Gil
[00:48:27] S0: Yeah. And then, I mean, in terms of what you decide to do next, I think this is like you know, a pretty normal process of constraint management. Right? I mean, it's like like, I think every lab in every field across the world probably feels compute constrained. I think that that's probably true here too. Right? It's like so, I mean, I know, like, you know, there's always questions. It's like, okay. Should we double down more on advancing the protein piece? Should we do more of the cellular stuff? I think those are kind of ongoing debates in terms of how you sequence that. And then, yeah, within that, there's kind of being at the Pareto frontier about how much you wanna train the different models in order to like, and and the size of the models is also dependent on the scale of the data that you have because you have yeah. For for obvious reasons. So, yeah, I mean, I think it's there's some of that as just where you wanna be on the curves and then normal constraints, but I think that this is, like, probably the same process that, like, any research organization goes through of, like, you wanna go in all these different directions, and you're just trying to constraint optimize and make enough progress to do world class work at one thing at a time while planting some seeds that can blossom over the next couple years as well. Yeah. This has been the most dynamic [00:49:37] S4: period of technology, at least, I've seen my career. Mean it's so exciting in terms of everything that's happening with AI. Every week there's something new that's changed.