00:15:51 · Oliver Vince
[00:15:51] S1: That is correct. I mean, we are, for want of a better word, like, trying to build, like, general biological intelligence. Like, we're trying to build models that that get generally smarter, and I think other people have approached the problem a bit more, like, linearly. So a bit more saying, like, look. Here's how the drug discovery process works today. We're gonna build a model for this and a model for that and a model for this, and we're gonna sort of treat each one like a bit of a game, and we're gonna play each one, we're gonna get really, really good at each phase, we're gonna stick them together. And what that means is in each of the individual things, they get very, very good at that, but they sort of, in our view, you know, I think that's very promising, I don't, you know, I don't want to be negative on any of it, but I think that has certain limitations in the fact that you're replicating the existing process. Now we've taken a bit more of a leap and said, look. What if you just train this thing to understand biology and you learn the ability to talk to it? What would it be able to do for you? Right? And so an example of that would be normally when you treat antibiotics, you would take the pathogen, would you do a bunch of modeling on the pathogen, you'd find the target, you'd find the mechanism of action, you'd find the thing, you do the docking, and you do all these steps, and then you design a peptide or an antibiotic that worked against that. And then you, you know, you test the sort of five steps in that process, and you build a model for each one. What we did is we just built this model, and we say... We showed it lots of pairs of pathogens and, like, antibiotics and just said, here's a new pathogen. Give me a new antibiotic. And it just jumps from that, and ninety seven percent of the time it works for... In the lab, and it works, you know, equivalent to last line in in mice and that sort of thing. And we've no idea in principle, like, how it did that jump, but we know that it's able to jump from a to b. Right? So in kind of in the way that you're not quite sure exactly exactly how ChatGPT took your prompt and jumped to something at the output, but you know that it didn't, you can check that it did. We... Is...
00:17:40 · Oliver Vince, Sebastian Johnson
Rather than saying, the equivalent with ChatGPT would be saying, we're gonna try and understand, you know, what what your sentence structure is, and then we're gonna try and, like, understand the meaning of each word, and then we're gonna, like, reason our way through all those steps. We've just said, look. Train it to be big enough, scale it, and then ask it a question and see what it can do. And quite funnily and quite hilariously, like, when it's smaller, it can't do a lot of these things. Right? So you ask it the question, it just... It spits out nonsense. Yeah. And so it's kind of very simplistic in some senses. Obviously, it's, you know, a bit more... The team wouldn't like me saying [00:18:14] S0: It is scary that it's like black box of, like, we're not quite sure quite sure how it works, but it's just almost magical. Like, what was that like for you and the team, I guess, when you had those great 90% results come back?
00:18:26 · Oliver Vince, Sebastian Johnson
[00:18:26] S1: It's it's it's a pretty... Like, it's hard to... When you exist in a world that requires all that scientific rationale and has previously built on, like, you know, very, very rigorous experimentation to just take a new approach and just say, look, we're gonna do this to see what happens. Like, you know, it's a high risk kind of... But, you know, it's a rewarding way to be because it's not even rewarding in the result that you get today. So it's not a... The nine seven is is good, but it's not the reward. Right? The reward is that you have a direction of something that has got there in a different way. So, you know, you can get the same result by putting block... Lego blocks together, but the fact that you get to it using something that is completely unsupervised means that you you have a a technically feasible path to giving it more and more complex questions. And then back to my earlier point on cancer being a personalized, complex, and a moving target, it gives you the potential for something that one day could understand all those different parameters, think about them in a way that we don't understand that's far beyond our current comprehension, generate a solution, and, hey, presto. It works. Right? And we're not there yet, to be clear, but, like, that's the sort of, like, landscape that we open up by having a sort of unsupervised scaling. [00:19:40] S0: Yeah. I mean, it it demonstrates a real level of intelligence versus just a sort of, like, training it to respond to certain questions, I guess, or information retrieval. [00:19:47] S1: Exactly. Yeah.
00:19:47 · Oliver Vince, Sebastian Johnson
[00:19:47] S1: Exactly. Yeah. [00:19:48] S0: And so what... Why wouldn't this work? I mean, let's just say three years time, it doesn't work. Where will be, do you think, you know, the areas where this would have failed? Is it the data itself maybe not ever... Not quite having the scale of Atlas that you really wanted it? Is it the intelligence that maybe the intelligence can only be scaled so far? Could it be in the application itself? Where do you see at this point the biggest chances of failure being?
00:20:15 · Oliver Vince
[00:20:15] S1: I think they are realistically tactical. Right? They're they're... Do you overspend in some... You know, do I I think it will work, and I think if you had infinite money and infinite time, I would be a 100% sure that it's gonna work. The question is whether you... From an a company building standpoint, whether you can do that. And I think, like like, more specifically, right, like, it's kind of... Back to being very, very simple, all you're doing is learning to read DNA and write DNA. And basically, the bigger the model gets, the longer piece it can read with a higher accuracy and the longer it can write with higher accuracy. And so basically, the question is like, how much does... How big does a model need to get to write, you know, 20 letters, 30 letters, 40 letters? And then what can you do with 20 letters, 30 letters, 40 letters? And how valuable is that? Right? So, like, is it... If you need to operate in a world where you need to monetize 20 letters for, like, five years and then monetize 30 letters for five years, you can do it, but it's a very different approach to saying, look, the fact that we've gone from 20 to 30 to 40 shows that we will get there to a 100 or a 150 or a thousand. But, you know, like that requires a bit of a leap of faith, which ultimately is where the tech companies are interesting because they've taken that leap of faith and they've seen it before, which is very, very different in some senses to many of the sort of established ways of developing medicine.
00:21:42 · Sebastian Johnson, Oliver Vince
[00:21:42] S0: Have you faced any cynicism or skepticism from that more traditional industry versus the tech industry? [00:21:50] S1: Of course. Of course we have. Right? And, like, I mean, to be clear, like, I... You know, many of us have... I have background in... My PhD was in cancer research. I have... Like, many of our team have... Are are, like, trying to be realistic and optimistic at the same time. And so we're aware of our own flaws and why this doesn't work better than anybody else. But it really is just like, look, there is a promise there, and therefore, it's worth doing. And you kind of... If you can get people to accept that, it's a very, very, very exciting and very fulfilling thing to do because you're not guaranteeing anyone anything, but you're saying, look, here's the potential. It's unequivocally huge if it works. Now let's go on the ride and see if it works. Right?