00:16:22 · Sachin
[00:16:22] S1: Speaking about the AI landscape for a second, I think people are recognising the limits of LMs, you know, text based learning about the world, and now there's a whole kind of movement around working world models and trying to understand the world without just text. How have you guys kind of cracked this? Like biology is obviously such a hard problem to solve and probably very different to what you're working on at OpenAI. Can you give us a kind of look inside your kind of thinking process and how you've kind of gone around the complexities of the real world?
00:16:51 · Josh Meier, Adam Miller
[00:16:51] S0: Yeah. I think a lot of it comes down to some kind of scaling law at the end of the day. Like, there is a process here that at its core seems like it can be learned. Something like protein folding, for example, predicting the atoms in three d space, like, given a sequence, you know, if you put either learn the problem better, like use better algorithms, put more compute so the model can process better, or more data, these are levers that tend to, you know, make progress on the problem. So if we believe that these things are predictable, then it's just a matter of getting those ingredients into a place to solve the problem, and I think that's ultimately what keeps us going, you know, through the hard challenges. Cause you asked me about doubts before, right? Like, you know, it's always like, this has gotten so good, can we make it better? Right? And I think we're just convinced that you can almost always make it better, you can keep going through that scaling law, and it's not always going be easy, there's going be a lot of work to do. People often ask us like, from Chai one to Chai two to Chai three, like, how did you get these big improvements? Like, what what was the secret from one model to the next? And I wish it was as easy as one secret that we had to find, but it's really compounding, returns. It's, you know, every week or every month, you know, making five to 10% improvements that you stack those together up for a long time, and you end up in a very different place than where [00:17:59] S2: you stand.
00:17:59 · Adam Miller, Sachin, Josh Meier
[00:17:59] S2: you stand. [00:17:59] S1: How have you kind of structured the company to go about these things? Because as you said, there's not one rate limit on progress, there's multiple at the same time. And how does, like, 30 odd people solve a problem this complex? How have you thought about the organizational design and the kind of philosophies you've installed in people here? [00:18:16] S0: Yeah, we're very interdisciplinary. We all work in person together in San Francisco. And I think this means that ideas percolate very quickly. People are very, like, open and transmorph with one another about what's working and what isn't, so you're always getting feedback from the people next to you. We have very few meetings in the company because, science does not run on a pace of, you know, weekly meetings. It happens in the pace of like, holy cow, just got this new result, let me go show it to the people next to me and like have them tear it apart or think through together like how we can make this better. So I think a lot of it just comes down to the people that we have here. We have people who are really excited about what they're doing, really passionate about this problem, and also just really like working with one another, and I honestly think those are like the ingredients to success. [00:18:56] S2: That's awesome. And obviously one of the really exciting things for us today was hearing about Chai 3. Could you sort of walk us through in your own words, some of the impact, I guess, sort of results of Chai three that you've been seeing?
00:19:08 · Josh Meier
[00:19:08] S0: Yeah, well, first of all, just taking a step back where we were even just a year ago, right, we could barely design molecules on the computer. Antibody design had about a point one percent success rate in the lab, so one in a thousand, or sometimes even one in ten thousand molecules, would actually bind as expected. With Chai-2, that was about fifteen to sixteen percent, and now it's more than doubled again with Chai-3. But the thing I'm most excited about with Chai-3 is not even just the success rates, right, and the efficiency you get out of it, but the quality of the molecules that are coming out. So the molecules we got from Chai-2 were good starting points, and that's like a very big leap on the problem, but they weren't they're clearly like far from being ready to like go into patients. There was a lot of work that we would have to do in order to make the molecules better, and we've got tools in the lab to do that, so Chai two was impactful in the field. But now with Chai three, we're getting molecules out of these models that have many of the properties dialed in that we need for these molecules to actually be drugs and to go into patients. And that's incredibly exciting because it allows us to now take new ideas, create molecules quickly, and then see in the lab whether they really work, you know, do they have the kind of biology, modulation that we're looking for. If it looks good and we want to advance it forward, we don't have to do much of that optimization anymore. In some cases, we might not have to do any of it at all, and it's getting very close to this vision we have as a company of someday creating a zero shot drug candidate, like a molecule that's ready to go into patients that comes out zero shot from the model and requires no changes. And that feature doesn't look very far anymore.
00:20:33 · Sachin
[00:20:33] S1: So Josh, we've been focusing a lot on drug discovery and it feels like you guys will solve that very soon. The next stage of drug development is obviously clinical trials and this is what takes a lot of time, there's regulators involved, it's quite complex, you have to do these studies over a long period of time. Will you ever kind of more like vertically integrate into that space to speed up that process so the end customer at the end of the day can get drugs a lot faster?
00:20:56 · Josh Meier, Sachin, Adam Miller
[00:20:56] S0: Well, way that we often think about even drug discovery is we don't want to just solve drug discovery in isolation, but if we can make a lot of progress on that problem, then it can actually make the problems before and after it also more tractable. So, for example, when we think about trials, if we are creating better medicines, it's actually one way to make our trial easier. Often trials are difficult because we're looking at, you know, are we giving the patients four more months to live or five more months to live, and the error bars on both of those things are like three months, and then the whole thing, it's very hard to know whether you see some significance, but if we can create better and more impactful medicines, actually a lot of the clinical trial work becomes easier. And similar on the previous stage, like even choosing what to work on in the first place, if we can generate molecules, very quickly against many different ideas, then we can now go from idea to animal very quickly, or to a disease model, and then figure out whether the molecules we've made are having the impact we're looking for. So I actually think getting very good at solving, a, you know, one part of the problem, actually has, pretty massive second order effects in the ones that surround it. [00:21:58] S1: Can you kind of paint that future in a little bit more depth? [00:22:01] S2: Like, what [00:22:01] S1: does that actually mean for the world if that happens? And everyday person, they may be watching this that has a parent or loved one that's maybe sick, what actually does mean?