00:05:39 · Josh Meier, Sachin
[00:05:39] S0: Yeah, so I studied chemistry actually in school, and I was a programmer since I was a kid. I did a lot of computer science in college too, But, so at OpenAI, we were training some of the early language models, right, GPT-one, GPT-two was happening then, and it occurred to me that if language models worked for natural language, French and German and English, why wouldn't they work for the real natural language, like DNA and proteins? And that's really how I got into it. I kind of just got obsessed with that idea at that point. Even back then, you know, like the business model around AGI was not very clear, right? It's like you had these, very mediocre chatbots and you had to take a ton of generations from the models to get them to do something interesting. But in biology actually, if you could take a 100 generations and like just, well, just one of them was good, that's actually incredible. People are used to screening millions or billions of molecules just to find like an initial starting point where you go make it better. So it felt to me actually that like the ingredients in AI were actually set up perfectly to have some impact in biology. [00:06:36] S1: And this actually isn't that much of a commentate. Like traditional pharma will say that AI falls apart in biology related problems. What's your kind of response to that? How did you know that there was something here where you could apply some of things you learned at OpenAI to solving this problem?
00:06:52 · Josh Meier
[00:06:52] S0: Well, I think if you asked me that question a couple months ago, I probably would have said they're correct, right? Like on the whole, this tech has been very promising, but it hadn't really reached the inflection point for it to be useful because we have to keep in mind, like, discovering a drug is very, very hard, right? Like, we can't see things at the atomic level ourselves. The search space is massive. There's more possible molecules than there are atoms in the universe. So there are, there's so much to search through and we have very primitive tools for getting there. So as a result, have spent decades building very, very good tools to screen through tons of molecular matter. When I say tons like millions to billions, right, it's still a small fraction of the search base, it's actually AI has quite a high bar to go against. You need to do better than what people are doing in the lab today. I think it turns out that, you know, where we're at right now, it's not necessarily the highest bar anymore, right, because AI has these scaling laws that you don't have in the lab, but it took a very long time for us to get to this point, and it was really only in the last couple of months with Chai two and now Chai three, that we've actually managed, I think, to make a dent on this space.
00:07:53 · Adam Miller, Josh Meier
[00:07:53] S2: Do you think a lot of those doubters, those doubts will change once Chai three comes out, and some of the sort of philosophical reasons behind why some people say AI isn't gonna be able to, you know, tackle this problem within biology? Do you think those doubts will change? [00:08:07] S0: I think when people see the data, then it does, yeah. I think first it's seeing it in a paper, for example, like the work that we put out, then it happens with our customers actually using our product, making molecules, seeing in their own hands that the thing works. So I think, you know, these are scientists at the end of the day, and they are convinced by data, and again, they were right before, like the data wasn't there, and they were right to ask these questions. These people are responsible for some pretty big campaigns. It takes a long time to discover a drug, it's very expensive, you can't afford to just go for the latest hype, you really need to go with something that's trusted and something that's working. At the same time, also have to innovate, and in this industry, in the pharma industry, you can't sell your drug forever, you'll lose exclusivity after a couple of years, like your drug goes off patent, and you always need to invest in the next thing. So that's the bright side of this too, people are always willing to try new things, and once they see it's working, then they'll start adopting it at scale.
00:08:07 · Josh Meier, Sachin
[00:08:07] S0: I think when people see the data, then it does, yeah. I think first it's seeing it in a paper, for example, like the work that we put out, then it happens with our customers actually using our product, making molecules, seeing in their own hands that the thing works. So I think, you know, these are scientists at the end of the day, and they are convinced by data, and again, they were right before, like the data wasn't there, and they were right to ask these questions. These people are responsible for some pretty big campaigns. It takes a long time to discover a drug, it's very expensive, you can't afford to just go for the latest hype, you really need to go with something that's trusted and something that's working. At the same time, also have to innovate, and in this industry, in the pharma industry, you can't sell your drug forever, you'll lose exclusivity after a couple of years, like your drug goes off patent, and you always need to invest in the next thing. So that's the bright side of this too, people are always willing to try new things, and once they see it's working, then they'll start adopting it at scale. [00:08:58] S1: In the past, what's been, like, the rate limiter to Chai's progress? Is it the amount of compute that you can throw at the problem, or has there been other things? And, you know, maybe the context of the question is is, looking around, I was so shocked that there's only, about 30 people on this floor right now. I thought there'd be hundreds of hundreds for the amount of kind of progress you guys have made.
00:09:17 · Josh Meier, Adam Miller, Sachin
[00:09:17] S0: Yeah. So I I think, almost everything is a bottleneck in the company right now, which actually means, like, nothing is a bottleneck. It's pretty much like any lever we push right now, we're getting better results, whether that's compute or data or even customer interactions, right? So it's a very exciting time where we're able to push all of these things. So actually, to answer your question, it's, at the moment, it's not like there's any one of those things. Over the you know, past couple of months, it's it's ebbed and flowed always, right? It's like, okay, this week we need to get more compute because we discovered something new for the models, or now, like, we really need to go and double down things on the product, make it more reliable or ship some new features that our customers really want, Or maybe it's we need to go in the lab, we need to generate a lot of data. Maybe it's we need to go stress test this really hard drug design problem that no one in the industry has been able to crack, and we need set up a benchmark to see whether our models can do it. So it, you know, changes week by week, but I think that's what's so exciting when you have so many of these bottlenecks at the same time. You're going to learn something new every week, and that just creates a ton of excitement within [00:10:14] S2: the company. [00:10:14] S1: Josh, we're really curious about what diseases Chai aims to kind of target first.
00:10:14 · Sachin, Josh Meier, Adam Miller
[00:10:14] S1: Josh, we're really curious about what diseases Chai aims to kind of target first. [00:10:20] S0: One of the interesting things with these models are that they don't seem to really care about the disease that they're going after. They seem to kind of work across the board. So we actually don't think of it as like, this is the one specific disease that we wanna go after first. We actually wanna parallel track many of these, and I think that gives us the highest chance of of seeing success across the board. [00:10:37] S2: So are you saying that there isn't going to be any more difficulty in solving certain diseases than others? [00:10:43] S0: I'm just saying that one is, yeah, from the model's perspective, it doesn't think that one is necessarily harder than another. There might be some biology reasons, it might be hard to find the patients, but if we think about it on the molecular level, the targets that the model is looking at, it doesn't have a preference. It's not like cancer is easier than immunology or something like that. [00:10:59] S1: Okay. Are you going to stack rank what you prioritize according to what's most impactful in society, or has that been something that's up to pharma to prioritise? [00:11:08] S0: Well, I think that will naturally happen via our pharma partnerships, so people will approach the problems that have the deepest unmet need or the biggest patient populations, so I think that will happen naturally. [00:11:17] S2: Yep, and what other ones that you think pharma will then sort of pick and prioritise?