It's interesting because, you know, I always read stories and it's hard to separate, you know, what's real and what's just grabbing headlines. But, you know, I've read a lot over the last year or so that, you know, some industry experts believe that maybe every single drug in the future will start from an AI discovery. Is that the case? Is that where we're headed? Are there still going to be offline scientists and offline researchers not using AI to discover new drugs? I

mean, look, there's tons of hype, no doubts, right? There's tons of headlines that aren't real. There's tons of stuff being tried that isn't going to work out, like for sure. At the same time, it's also fundamentally true, I believe, that we're going to not even say AI drug versus not. It's just going to be used everywhere. Like, maybe to kind of drive parallel with genetics, right? There was a time when people said, like, is genetics useful in the life sciences? Will we use it in drug discovery? And now it's just a given component of every drug discovery program in different places, in different ways, but there's no such thing as do drug discovery and forget that genetics exists, right? And the same thing is, I think, probably almost already today true for drug discovery, right? The degree will increase, but it's just going to be kind of part of the fabric of drug discovery. You

know, one thing, one person that I do trust on this is, you know, NVIDIA CEO Jensen Wong said a while ago that he thought that AI-driven digital biology would be the next kind of amazing revolution. And, you know, he said that it's going from a science to more of an engineering discipline. Is that where the industry is kind of headed? That is this going to be more of an engineering versus more of a, you know, traditional scientific discovery? Yeah,

I mean, I think it's going to be more data driven, right? And I think we see that even within biology and chemistry themselves, right? When I was doing my PhD, you could get away with quantifying using some plus signs in your paper. This was two plus signs, it was four plus signs, right? And now, of course, no one does that. It's precise analytics on any microscopy image, for example. Right. So it's certainly going to become much more data driven and structured and many more engineering and data science tools being deployed. Yeah, for sure. No doubt. I'm with Jensen.

So looking at the bigger kind of topic for today that I want to shift into a little bit, you know, how does all of this ultimately help companies, you know, help medical institutions put patients first, right? Is it just, you know, oh, we're going to have so much data or we'll have better drugs. But like how do patients ultimately win from the type of work that you're doing?

Yeah, totally. I mean, one is timelines, right? Moving faster. People are waiting. People are dying today. They don't have a treatment, right? We just had someone come in, we have patients come in and give talks, both people waiting for treatments and people who have been matched with treatments. And there was someone in just the other week sharing this powerful story about a rare disease that he has, and that if he had been diagnosed two years before, there would have been no treatment, right? He would not have been there telling that story. So one parameter is just time, right? Real people waiting for real cures. Cost, you touched on that at the beginning, I think. The medical system is complicated. Things are expensive. And one lever, there are many pieces to this, but one piece is how crazy expensive it is to get a drug to the market, right? So if you can bring costs down through better data-driven decision-making along the way, that's a huge component. then ultimately, you know, one day we hope that we can make, we in the industry as a whole, better medicines, right? Where today there's lots of studies showing, for example, that many medicines do things in the body addition to the ones we know about, like the better we understand things at the systems level, hard to grasp within the human mind, but can be modeled with machine learning models, with knowledge graphs, etc. We hope in the future in a way where more of these unintended consequences of medicines can be factored out. But yeah, long journey, and I think a meaningful one. All

right. You said one of my favorite keywords there, models. Let's talk about that. Because large language models aren't necessarily new, but obviously they're growing, they're exploding. You know, all of a sudden they're being used in every domain imaginable. But even within your industry, how have large language models both changed from, you know, how you were maybe previously using traditional machine learning and artificial intelligence? And, you know, where do you see right now, you know, the excitement with using these domain-specific models for, you know, highly specific use cases?

Yeah, totally, totally. I'll give you two, what I think, cool examples with the large language models. But I put it in context to the underlying architectures of large language models, transformers generally, are now used by Recursion and many other companies to solve many problems outside of language. So in chemistry, in biology, in imaging, et cetera. kind of in the everyday conversations about AI, that the same watershed moment that we all appreciate with chat, TTP and kind of other equivalent tools, it's not just in language, we're seeing similar transformations in all of these other technical fields. So super cool. And then the second piece there, along with transformers, there's all just richness of other new models. It used to be kind of one AI model to rule them all maybe half a decade ago. And now, you know, you have your pick depending on what it is that you want to work with. And we have some proprietary ones in how to recursion. Many other people do too. And many open source ones. With that, you want still two LLM examples?

Yes, let's do it. Let's do it.

Let's do it. One is just reduce toil. And so this is just this like, it's not just about the future. It's not just about hope that the drug is better later on. We're saving our scientists material time on things like literature review. So, you know, we just put out a statement a little bit ago on, we've reduced the time that our scientists need to review literature by 60% before we go into expensive, more time consuming. It's called hit to lead activities, kind of a standard step in drug discovery. That's really key and core. And so that's a big deal, right? That's time that they're spending thinking about the details of the science, not searching for the right paper, not kind of doing arbitrage on what it is you're focused on versus not, right? So like just a huge win in putting human minds on kind of the most thorny problems, not on kind of the simpler tasks. The second example is using LLMs to compare what we know internally with our huge data sets and experiments, and using the LLMs to compare that with what the world knows. Doing that then also at massive, massive scale, imagining that we can do that, and we do, over trillions of trillions of different kind of relationships and in combinations, how genes interact with genes, how potential drugs interact with genes, how genes relate to diseases, etc. And so this is just this huge combinatorial space that you couldn't go after manually, just wouldn't be feasible. It'd be, you know, you'll get some 0.0001 percentage with a lot of effort. And this allows us then to make sure that we as a company focus not on what the LLM knows from the outside world, but on what recursion uniquely knows, right? We don't want to be working on the exact same things as every other pharma company, because we want to make sure that patients have shots on goal for treatments in areas that currently aren't being worked on, on reusing angles and ideas that haven't already been tried 10 times, right? And so that's a cool version of LLMs, I think, that focusing not on what is known, but uncovering what's unknown. And

speaking of, you know, uncovering new things, you know, I'm not going to ask you to, you know, look into your crystal ball, but, you know, you've already kind of walked us through very briefly how these models have changed, right? It was just first this, you know, one giant model and everyone was enamored and then you figure out, okay, it's actually not that great for what we need it for. And then you just said, hey, right now, some of your researchers are saving 60% of time to review literature. What's that next big phase or next iteration that maybe you're excited about, or maybe your colleagues are exciting about, or the industry? What's that big next jump for generative AI? Yeah,

I mean, a big thing that we really have our eye on for our models at recursion is, you know, for the last decades, we have been generating huge data sets fit for purpose of machine learning to then build the best models we can, industry-leading models. We're starting to see a shift where we think on the horizon is a shift in that world of data first, models, to having so-called world models, models good enough to be your starting point. And then the experimental world is really focused on validating with depth the most important insights from the models. So models first, then data. Yeah.

And let's talk about the concept of world models because we haven't talked about it a lot on the show, which is crazy after like 500 episodes. Generally, when we talk about world models, it's more for like AI video generators, right? It's like, oh, can open AI understand physics, right? And can Google understand how a person moves from point A to point B? What does it mean in your field when we talk about world models? Can you explain that a little more? Yeah, good call out, not video

games. It's models that aren't created to tune to a single individual predetermined problem, but are created to understand a large domain. And so in our case, this is about human health and disease. And it's about doing it at multi scales from what proteins do. Some of you might have seen the alpha fold news on the news a few years ago, being able to predict protein structures up to what cells do into data sets from patients so that insights can be linked between these different domains and reasoned about how do not just one protein act, but networks of proteins and so on.

So you've given us, you know, quite a few good examples and illustrations on, you know, what you all are doing and how you're using generative AI to, you know, put patients first, right? And just, you know, the ability that now, you know, certain patients have drugs that they can use, right? And they've been cleared and how generative AI can help expedite the drug discovery and drug trial process. But what about if we take the word patients out of it? And how can this help in the future, maybe humans not become patients per se, right? So on the predictive end, how can this help maybe, you know, people not get diseases or people not get sick? Is that another area that you're looking into? And if so, how does AI help with that?

Yeah, totally, totally. Yeah. I mean, we believe by understanding like the different, you can imagine that the same model concept, world model concept be used, can be used to model any state, for example, a human cell, right, from all the way to deceased, all the way to healthy. And we already do this in a certain degree at the company. You can imagine also identifying molecules or treatments or behaviors stop the progression into a deceased state. Right. In fact, we've had some exploratory work in senescence in the idea of stopping unwanted aging pathways on a cellular level. Right. And so conceptually, I think that is it's totally something that these approaches are going to be used for, whether at recursion or at other companies.

Yeah. And, you know, speaking of aging, I think it was the which I was shocked when I when I read this headline kind of recently. Actually, the Anthropic CEO said he could see with advancements in AI, the human lifespan doubling, I think he said in like, as little as 10 years, which I'm like, okay, that sounds nuts, right? I mean, is that a potential future, right? Like, I think these models, you know, as they become more powerful, as you have world models, as you have the equivalent of, you know, the world's best model, but for protein building and duplicating proteins and all these things. Is that a realistic thing to say in the future, life spans might be way more than they are now? I

mean, I am not of a biologist to know about what kind of the inherent limitation of human aging are, right? But kind of more fundamentally about like the potential on the AI side, right? I think often about this, there's this famous Andrew Inquo, he's kind of one of the big names in AI, right? And he, around 2016 or so had this quote, everyone was using, you know, that if a normal person can do it in less than a second, then you can automate it with AI. And lots of people were like, no way, that can't be done. You're overstating it, right? And now you look back and you say, well, that feels like really undershooting it, right? And kind of the famous Andrew Inn quote from last year is AI is the new electricity, right? So just to say, like, I think we're in this like big inflection phase where it's kind of hard to know where the big directions is going to go, right? Yeah,

yeah.