Ep 72: Co-Founder of Chai Discovery Joshua Meier on 99% Faster Drug Discovery, BioTech’s AlphaGo Moment, Building Photoshop for Molecules

In this episode, Jacob sits down with Joshua Meier, co-founder of Chai Discovery and former Chief AI Officer at Absci, to explore the breakthrough moment happening in AI drug discovery. They discuss how the field has evolved through three distinct waves, with the current generation of companies finally achieving success rates that seemed impossible just years ago. The conversation covers everything from moving drug discovery out of the lab and into computers, to why AI models think differently than human chemists, to the strategic decisions around open sourcing foundational models while keeping design capabilities proprietary. It's an in-depth look at how AI is fundamentally changing pharmaceutical innovation and what it means for the future of medicine.

• 0:00Speaker 00

  Biology is inevitably one of the most exciting applications of large language models. You talk to folks, anyone from Dario and Anthropic to a bunch of leading researchers, and everyone is excited about the potential for the impact on human health and lifespan. Today on unsupervised learning, I sat down with Joshua Meyer. He's the CEO of Chai Discovery. Chai is one of the biofoundation model companies. There's been a wave of these companies that have raised a bunch of money and have published some really impressive results. People should check out the Chai2 paper that Josh and team released, just incredible outcomes on antibody design. I think Joshua is kind of the perfect person to just talk about what's happening in AI bio today and what's ahead. He was the chief AI officer over at Absi. He did a lot of really interesting research over at Meta previously and has one of the most interesting companies in the space today. I think his passion and just knowledge of the subject really comes through in this. It's a subject area I've always been personally passionate about.

• 0:45Speaker 00

  I used to work in pharma-facing products over at Flatiron Health and think it's just one of the most exciting applications for how AI really can improve humanity. So without further ado, here's Joshua.

• 0:57Speaker 00

  Well, thanks so much for coming on the podcast.

• 0:59Speaker 01

  Thanks for having me, Jacob. Excited to be here. One-time listener, so fun to be on the other side.

• 1:03Speaker 00

  Well, I've been looking forward to this one for a while. I feel like everybody, you know, from Dario and Anthropic down to, you know, the person you talk to on the street is, you know, I feel like the world is convinced that biology is just this unbelievable use case for AI models and really going to have some of the biggest impact of anything that folks are working on now in terms of just the impact on humanity. And so no better person to come in and educate our listeners on what's happening in the space.

• 1:27Speaker 01

  Oh, I'm humbled. Definitely true, though. You know, AI is a perfect tool for biology and excited to dig into that today.

• 1:34Speaker 00

  You know, AI drug discovery, the idea's been around for quite some time, and I'm sure our listeners are familiar with some of the proper nouns in the space like recursion, you know, one of the early kind of startup successes in this world and AlphaFold and some of the interesting, you know, breakthroughs that happened on the DeepMind side. I think to start, it would just be helpful to contextualize for our listeners, what has this like journey been like over the last decade in AI drug discovery? And maybe bring us to like where we are today.

• 1:59Speaker 01

  A lot has happened over the past decade. Think if we look about 10 years ago, so 2015, you know, deep learning had just started working on GPUs. I think the AlexNet paper was like 2012 and I might Kind of got started building neural nets in 2013. Very different time, you know, the tooling was not in a great place. Things were not moving as fast as they were today. But 2015, I think, was really the time when you started to see the first deep learning models being applied for protein structure prediction and protein design. It's actually my co-founder, Mac McPartland, built some of the first deep learning based models for protein folding.

• 2:31Speaker 00

  And at that level of compute, could they do anything?

• 2:34Speaker 01

  Well, actually, it turns out you could do a lot. So this is a funny thing, too. You look at some of the recent models like GPT and BERT, you know, like language modeling, what's a language model? It's like kind of predicting the next word in a sentence. Or the way people did it a few years ago was mask out a word and try to predict what's there.

• 2:49Speaker 00

  Right.

• 2:50Speaker 01

  The biologists have actually been doing this for like two decades. This is how people used to do protein folding pre-deep learning. You would take, you would take like a sequence you're trying to predict the structure of. You'd find a bunch of similar sequences and you'd look for some patterns. And it turns out the math that they were using is basically the same math we use in language models today.

• 3:08Speaker 00

  Well, I thought healthcare was always like 20 years behind everything else. It's good to hear that we're, we're ahead on something.

• 3:13Speaker 01

  Yeah, yeah. So I think, I think the biologists essentially discovered language models way before we did. This interlink between AI and biology is actually quite strong.

• 3:21Speaker 00

  Yeah, and so I guess you had these early models on like limited amounts of compute. And then, you know, maybe talk through, obviously, there was kind of this first wave of tech bio funding and some folks that got started there. What worked in that era, what didn't work?

• 3:33Speaker 01

  Yeah, so if we look at the first, that was more in the research side, right? So thinking about how AI was used for protein building. If we now look on the commercial side, like startups and companies that were getting started, there's been a couple of like starts to this space, right? So maybe the wave one tech bios were companies that were trying to just create, like focus on data, right? Everyone used to say the problem with like AI and biology is the data. We're going to go produce a bunch of data. And actually, that is correct for some problems. You mentioned recursion, for example, they did some really interesting stuff on what we call phenotypic screening. So what they would do is throw a bunch of random molecules at some cells, take pictures of them and then use neural networks to try to understand what's going on. So this is a bit of like, let's boil the ocean, try a bunch of molecules and have a neural network filter out what works on top.

• 4:19Speaker 01

  The second wave of companies, I think, was trying to do stuff that's closer to what we're doing at Chai today, but with much weaker compute and models. But it was like, let's actually try to model the molecules themselves. So you saw this wave of AI small molecule companies especially, that were trying to predict various properties of molecules and use that to design better ones. And one of the challenges there is like, that actually makes a lot of sense and it's sort of what we're doing today, but no one really knew where the technology was going to go. So it was, I think, the right bet in terms of, like, this actually is a perfect problem for AI. But many of those companies were started like pre-GPT-3, right? Actually, even pre-GPT-1 in many cases. So you're doing this without like the latest and greatest machine learning technology today.

• 5:01Speaker 01

  And now you have what I guess I like to call the wave three companies, of which Chai is a part of, the companies that have started in the last couple of years, you know, post the big like AI craze, actually being able to borrow a lot of those lessons from NLP and computer vision and video generation, and then using that to actually build models that are really working and working with incredible reliability. Like you see with our latest tools at Chai, you know, success rates that are like 10, 20% plus. You know, you choose 10 molecules, you're actually going to get a couple that are even working on your first try.

• 5:31Speaker 00

  I think it would be helpful maybe just to contextualize like the state of the union today. I mean, obviously, people talk about, oh, you know, we're excited for AI models to be designing personalized cancer therapeutics for, you know, for everyone in some short order. Maybe just help us understand what like, what actually works today? Like, what have we proven with some of these newer models? And then what's, you know, what are the milestones still in the come that are, that will be exciting when we get there?

• 5:51Speaker 01

  So biotech's a really broad space. It might even be worth just like backing up into the different components, right? You've got target discovery, like, let's figure out what we want to make a drug against, drug discovery, so let's take that target discovery, a drug against it. And then we've got to go put those drugs into clinical trials. And there's been a lot of work around trying to do clinical trial optimization.

• 6:09Speaker 00

  Right.

• 6:10Speaker 01

  So I think there's really amazing AI work happening at the forefront in each of those areas. Maybe let's start backwards. You know, in the clinical trial space, you have companies that are using LLM, you know, the same chat GPT-style stuff to write regulatory documents faster. There's a ton of paperwork that goes into this. It's actually kind of funny, you see folks like the FDA now saying that they're going to be using AI to actually read some of these documents.

• 6:32Speaker 00

  Of course, AI on each side.

• 6:33Speaker 01

  Exactly, yeah. So AI is going to be like judging the documents. And then, of course, now the pharma companies like, let's use AI to write the documents. So it's kind of funny. We'll have AI talking to AI right now. I hope that ends well. But that's like what I think the frontier in the clinical trial space. And I think that that's super exciting because it's just like, let's take these efficiency improvements and try to move faster. There's a lot of bureaucracy and also rightfully so. Something that you do have to regulate. You're putting a drug into a patient. We want to make sure we're doing that carefully. But there's been a lot that's kind of been built up over time there that can slow things down. So an area of AI, super excited by like, you know, the clinical trial optimization. And I think the folks that will probably have major impact there are the ones that are building on top of, you know, the Open EIs and anthropics of the world.

• 7:16Speaker 00

  So obviously, then I think across, you know, target discovery, you know, finding the actual compound you want and then optimizing that compound. Maybe talk a little bit about, you know, where we are to that.

• 7:23Speaker 01

  So that's the area where we're playing at Chai. So we're looking at drug discovery. So this is where we take a therapeutic hypothesis, maybe that's been figured out in, you know, the lab over the past 10 years. You have some idea, if I go and make a drug against GLP-1R, maybe that will be useful for diabetes or for metabolic disorders or for weight loss. And the question is, how do I actually get the best drug possible for the patient? And that's the problem that we're focusing on at Chai. And I think it's also a perfect problem for AI as well. And the reason for that is, at the end of the day, when you think about what a drug is, it's a set of atoms that are organized in a certain way, and a either some synthesis process to put those atoms in those form, or just like a

• 8:05Speaker 01

  DNA sequence that encodes a protein that kind of comes up with that form. So if we can actually learn how those atoms are interacting, again, on that fundamental level on an atomic scale, we can use that to go in and invent lifesaving drugs.

• 8:19Speaker 00

  Can you walk us through the sequencing? So like, obviously, you had AlphaFold, which, you know, was really around like structural prediction, right? And you're kind of like, you know, taking this a step further and going from not only structural prediction, but actually to de novo design. And so, you know, and obviously, it feels like that's happening across the field. What was actually required to to make that transition from, you know, the AlphaFold breakthrough to some of the stuff we're seeing today?

• 8:40Speaker 01

  Yeah. So if we look at AlphaFold, amazing breakthrough, I think 2021, AlphaFold 2, I'm referring to, actually, I think AlphaFold 1 was even before that, 2018. What those models were doing is you would take a single sequence, a protein sequence, and then try to predict what the 3D structure looks like. That's actually incredibly useful for basic biology. I can take some sequence that I'm studying, something I find in the lab, and now I can figure out what that 3D structure looks like. That's a process that can cost hundreds of thousands of dollars to actually do. These days, it's actually even come down, maybe let's call it like $30,000 to kind of get a 3D structure. But if you can take 100 million proteins and go predict the structures, right? Like, you're talking billions of dollars worth of research that basically can be done. I don't want to say overnight, there's a lot of compute that goes into it,

• 9:22Speaker 01

  but you know, running your GPUs for a couple of months, you can really start to predict the whole protein universe. And what's happened since then, so the past two years or so, let's call it, is we've graduated from predicting the structure of single proteins to predicting the structure of entire macromolecular complexes, which means how does a protein interact with another protein? Or how does a protein interact with a small molecule or small molecule drug? How do DNA and RNA interact with those things as well? And that's been becoming transformative now for drug discovery, because drug discovery is not just about understanding a certain protein. That can be useful for our underlying biology knowledge. But as we start to move into design, right? Your question, how do we go from protein folding into design? With design, we're not trying to just change an existing protein.

• 10:05Speaker 01

  We're trying to invent a new molecule that will perturb or modulate the existing molecule. And this breakthrough of being able to actually predict how these things interact with one another is immensely useful then for actually going and designing those things as well. It's almost like we've now built atomic-level microscopes. And in the patronage, under that atomic-level microscope, we're designing these atomic-level drugs that can actually modulate those proteins in a certain way. And that's the breakthroughs that now we're moving into 2025 and beyond. There's different modalities of therapeutics that people create. And like, I'm sure these models work better for some than others. Or, you know, talk a little bit about the various challenges involved in each of the different types of therapeutics.

• 10:48Speaker 01

  There's two main classes of therapeutics these days: antibody molecules and small molecules. There's also other ones as well. There's gene therapies you might be reading about in the news. People are building things based on CRISPR. But the biggest classes by far are, again, antibodies or biologics more broadly. It's a protein, it's created by biologists, what we call it that. And then small molecules, which is really chemistry, so going and synthesizing those molecules. And as a result, that's also been the main area where a lot of the AI bioaction is happening today, just because today, at least, you've got like the biggest markets around those.

• 11:22Speaker 00

  I think traditionally, like if you look at a pharma company, they're really good at one of these modalities. And they have a whole team that's like spent their whole lives in that. Obviously, one of the powers in the general, you know, foundation model spaces when these models generalize really well. How have you and other folks thought about, you know, hey, are we building models that are really good at one of these types of drugs versus, you know, the ability to, hey, you know bio, you know bio, it can generalize across the board?

• 11:44Speaker 01

  Yeah, it's funny because you're starting to see a transition now. It used to be people would segment these companies into I'm an AI small molecule company, or I'm an AI antibody company, or I'm doing AI for gene editors. But really, the models are modality agnostic. If, again, let's take this, these models that understand things at the atomic level, right? We have a model that we started with Chaiwan, which is all atom model, right? It can predict all these different modalities. People were calling it multimodal. But yeah, at the end of the day, it kind of is like, I guess it is multimodal. You're placing atoms, you're also understanding sequences. But, you know, these models, they take an input and they give an output and they don't really care whether it's a small molecule or it's a protein or something like that.

• 12:23Speaker 01

  There is a question about maybe where your focus, your data generation efforts, but they're even potentially transfer across these domains. You might have some data on small molecules that helps you get better data for your, for your antibodies. So I think you might still see some of these commercial focuses just because the manufacturing process is different. Maybe if you decide to partner with biopharma, the teams you're working with might be different. So there still might be some focus at Chai. We've been really focused on antibodies so far, but I think long term, the models work across the stack. So there's really no reason why on the AI level, you know, the models need to do one thing or the other. It's kind of like even for LLMs, like if you're good at doing, if OpenAI is good at doing text, should they also be able to do video?

• 13:05Speaker 01

  I mean, it is a slightly different model, but it's the same AI researchers you can usually build across those different domains.

• 13:11Speaker 00

  Well, it's been interesting to see too, right? I feel like the, in the, in the general model space, it felt like for a while this generalization thing was completely true. And it's like you get better at any set of tasks and it made you better in aggregate at everything else. And now it feels like maybe we're moving, you know, you have the latest anthropic models and they're like unbelievably better at coding. And it's not clear they're that much better at other things as well. Like, does it feel like as a space, folks are seeing these signs of generalization across today? Or do you think we're in an era of, hey, let's go really deep on, on, you know, a specific space?

• 13:39Speaker 01

  So again, I think just on the base model level, especially if you're just training things on, let's say publicly available data sets like we did with our first open source model, Chaiwan. The models like again, you train, it makes sense to train on more data and bring together these different modalities. So we haven't seen necessarily that the models are much better in one domain than another. And maybe one example of this is like, turns out that predicting the structures of DNA and RNA is a little bit harder than predicting how proteins and small molecules interact.

• 14:10Speaker 00

  Why is that?

• 14:11Speaker 01

  Yeah, so that might be a data problem, where there is much more data on the 3D structures of proteins interacting versus like protein and DNA. But I mean, that can change, right? You can build new data sets around these things. I don't know if there's like a fundamental reason why one should be harder than our, there could be, right? Like maybe one of these things is very dynamic and it turns out that we struggle to get dynamic data in lab around these things. So there could be fundamental reasons why one is harder than another. But I think the story is actually still being written around that today.

• 14:40Speaker 00

  Well, I mean, you kind of alluded to this and when you were talking about the first wave of AI bio, you mentioned so much of it was around, you know, generating data in labs and really innovative ways to kind of examine and interpret that data. I feel like, you know, a lot as people think about the opportunity for AI and biology, material sciences, these other spaces, a big question is around this like this loop of kind of running experiments in the real world and feeding that back into models. And, you know, a lot of companies are thinking how to get really good at this. You know, how have you thought about the role that data generation plays and then also building that loop of kind of real world experiments into kind of the core models?

• 15:12Speaker 01

  AI models need data. So you always need to be thinking about that.

• 15:15Speaker 00

  We don't have the whole internet in biology, do we?

• 15:17Speaker 01

  Well, actually, it turns out there is a lot of data available. And I think if you just count the number of like even DNA base pairs that are just available on the internet, you know, you've got like double digit petabytes of information to train on, right? And that's actually getting more than than what we even train some of these language models on. So I actually don't know if it's like the volume of data that's the problem. I think it could be the quality or the diversity of data that could be problematic. So it depends what problem you're trying to solve. Actually, one really interesting thing to consider there is the role of like pre-training versus post-training. So now LLM space, right? You you pre-train on all this data that's available on the internet. But the early models, like let's go back to like GPT three, GPT 3.5, those models hallucinated a lot. Part of the reason there is like these models are trained on Reddit, for example,

• 16:00Speaker 01

  for question answering, right? And on Reddit, nobody answers with, "I don't know." Like, that's just not really a valid answer.

• 16:06Speaker 00

  You surely don't get upvoted if you do.

• 16:07Speaker 01

  Uh, right, right. Yeah, I guess that that's true too. Yeah. Uh, good point. So, uh, if you've been in a question on some of the early language models, it would never answer, "I don't know," if it doesn't know the answer, right? It would just like start hallucinating something because it's like, "Yeah, if you have a question, you should give some kind of answer." The latest models have gotten a lot better there, and part of this is, uh, is the post-training recipes. So really interfacing with the models and saying, uh, you can just imagine what happens there. Like, uh, you have multiple users who are building training data, like talking to the LLM. The LLM, you know, has a question, and then you rank the answer. If it gives an incorrect answer versus saying, says, "I don't know," you know, "I don't know" is better than incorrect answer. So you start to build this post-training data. You start with your base model, you build on top of it. And I think biology actually

• 16:49Speaker 01

  has a very similar analog, where in pre-training, there's all this data that's like available on the internet, a lot of like academic studies. Uh, think about where that data is coming from. There's this, this great paper I love. Uh, I think it's from, from New York University, where they actually go and swab the New York City subway system and, uh, go and sequence all of these bugs that they're finding. Uh, and there's a lot of really interesting data that that's created there. I think that really helps the models understand what makes a protein a protein, but it doesn't tell the model like what makes a protein a drug, right? That's a kind of data set where you're not going to find in the New York City subway system. Um, if you want to understand what makes a protein a drug, there, there's probably other kinds of data sets that you need to start building, and that starts to get you closer to like the post-training world. Coming back to, um, uh, the question about like where data sets arise, I think you just

• 17:34Speaker 01

  want to just ask like, what is the problem you're trying to solve? I think we should be really ambitious and, and really hopeful and optimistic about what those baseline data sets can do. Like, we should think really hard about what the, you know, those bugs, bacteria on the subway are, are teaching us, because if there's a, you know, an answer that can be, a question that can be answered from that data set, don't go and like reinvent the wheel and build up your data set. Like, don't build data for the sake of like having the biggest data set. Build data in places where like, the data just doesn't exist, or the models are not performing well. So I think that's why you've, you're, you're seeing out with the Latest wave of companies, much more of an algorithmic focus. Let's be driven by the models. Let's not just have our view of what data is needed. Let's start with the models, see what they can do, and then build up the data sets we need in order to make them better.

• 18:18Speaker 00

  Yeah. Do you think the future of this looks like, I mean, obviously, one very large, pre-trained, powerful model with like, specific post-trains for, like, what is the level of specificity by which you would post-train for, you think, in the future version of these models?

• 18:31Speaker 01

  It's a great question. I don't know if I have the answer yet. If we look at what has happened in LLMs, I think the jury is still out there as well, right? You have maybe most people using or building on top of like a baseline LLM, right, a GPT-4 or soon a GPT-5. But companies will still take these models and try to fine-tune them for their specific applications. So there's a question, how long fine-tuning lasts? There's less and less fine-tuning happening as the base models become more powerful. I think the same thing will happen in biology as well. But to this post-training point, you might have a pharmaceutical company, for example, that has a lot of data related to a certain class of molecules that they like working on, and they might have the most data around that in the world. And it might make sense, therefore, to take one of the base, let's say, protein design models,

• 19:15Speaker 01

  and then fine-tune that on the data that they've generated in-house, or find a way to prompt the model with that kind of data they've generated. So I guess I don't know the answer to the question yet, but I think it's quite possible that you do end up with, you know, post-trained versions of these models in these different organizations. But it really depends how far we'll be able to get with the base models.

• 19:32Speaker 00

  So how does that impact how you think about your focus, right? Because you've obviously built these incredible models. And I feel like, given that, what you just talked about, there's tons of different things you could do with them, right? You could just keep building a really powerful base model. And there's tons of things that people can build with those base models long-term. You could, you know, build the powerful base models and partner with companies to, you know, that have their own data sets and help them kind of do post-training or fine-tune these models. Or you yourself could find some, you know, like, hey, like, this is an interesting end market. We could build up a bunch of data and fine-tune and post-train for that. Not that these are all mutually exclusive. How do you think about allocated resources across some of these different approaches?

• 20:06Speaker 01

  Well, you're exactly right. They're not mutually exclusive, and it really makes sense to be building across the board there. I think one thing I've learned in my career is not to bet against progress in AI. And that means we should probably continue to bet on the base models getting even better and build a strategy around that. But at the same time, I've also learned that you need to be pragmatic and you need to solve problems. And if you can solve a problem today, you probably should. So there's a certain data set that's out there and requires post-training a model on it. And maybe it's not as elegant as like a single base model to roll them all. If you can go and use that and you can now create a really meaningful drug that can get to patients, like I think you almost have like a social obligation to do that. And so they're not mutually exclusive and I think it really makes sense to be building

• 20:51Speaker 01

  around all those strategies.

• 20:53Speaker 00

  Totally. I feel like this tension exists in like anyone building something that's actually being applied in the AI world where there's always this feeling that like, yeah, could you go to a beach and wait five years for AI model progress and you probably get it for free long term? Maybe, but in the meantime, like there's so much you learn from actually deploying these things. And in your case, bringing a drug to market five years sooner is a pretty powerful moral imperative. So I guess on the data generation side, how have you thought about like, you know, there's tons of people that raise money and build big wet labs and like actually, you know, do all these experiments in-house. There's a really great network of folks to partner for, you know, these days to get data. And then you can also just get raw data sets themselves. How have you thought about like wanting to have data generation efforts in-house?

• 21:33Speaker 01

  Yeah, so we've really been driven by the models in this regard, right? And figure out what data do the models actually need? It turns out that a lot of the data sets that these models need can be built with somewhat commoditized methods, right? So especially if we look at antibody engineering, post-COVID, there was a lot of capacity that kind of got built up throughout the industry. And a lot of that capacity is actually underutilized right now. And we've been able to leverage it at Chai. So if you look at the Chai 2 report that we put out, we actually partnered with multiple different labs on the studies there. Part of that was also just like the credibility and just like trusting our results. You know, you can test something in your in-house lab, but it's also really good if you can have external parties that can validate what you've done. And then moreover, if we just think about the elastic capacity here, we were able to

• 22:16Speaker 01

  like really focus on building various versions of the model. And then, you know, when the new checkpoint is ready, you can go send a lot of capacity out there throughout the ecosystem. So I think that It makes a lot of sense to tap into the resources that are out there and to partner, you know, when possible. There was a trend in a couple of years ago where people really wanted everything in-house. It was almost like a cool thing to do to have like the most people you can attract to your company and the biggest lab that you can get. And I think these days it's almost cool to do the opposite, right? Like keep the leanest team possible, really maintain that kind of focus and figure out just what you need. So I think if we just play this out over the next couple of years, it probably ends up being a mix, right? There's if there's data sets or there's methodologies that are commoditized and you

• 22:59Speaker 01

  can build on top of other folks, probably makes sense to do that. If there's something that's more unique, maybe there's a certain disease model that you create, or maybe there's a certain really smart data generation strategy that you come up with. Those are probably the kind of things that you'll build in-house. So I don't, we're not really dogmatic about either of these. We're just driven by like, what is the model need? And let's figure out the most pragmatic way to get that data.

• 23:20Speaker 00

  Totally. Well, speaking of the models, I mean, you made this really interesting decision to open source, you know, some of the model work. And I guess like, was that something from day one you knew you wanted to do? Or talk about like, what led up to that and kind of how you think about the role of open source in the company?

• 23:35Speaker 01

  Yeah, great question. So we've really talked about two main models at Chai. We have Chai one, which we released open source out late last year, and then Chai two, which we just announced a couple of weeks ago. And if you look at Chai one, we open source that, right? So this was the model Chai one at a high level is the atomic level microscope. So predict how atoms interact at the atomic level. And then Chai two is a model for actually being able to design those molecules, right? So if you have your atomic level microscope, now let's actually, you know, move things around like under that microscope. For Chai one, we actually expected that models at that, at that capability level would become open source anyway. So we decided that like, we would be the ones to do it. We really just needed Chai one for what we were building next with Chai two.

• 24:19Speaker 01

  You know, if you can't really, you don't have the microscope, it's much harder to go and build the design tools on top of it. So that's why we started by building a microscope. We really made a bet when we started the company that the protein, protein interaction stuff was going to get a lot better. If you just look when we started the company, I think the state of the art methods at the time, they would only predict high accuracy antibody antigen complexes. So antibodies are about half of the drugs that are being developed these days. And only about like 10% of those interactions were predicted with high accuracy.

• 24:48Speaker 00

  So you kind of, you mentioned Chai 1 is this like atomic level microscope. And that's obviously something that you put out in open source and it's been used in all sorts of cool ways that folks should check out on Twitter and online. And then you, you know, more recently you released Chai 2, which is really around molecule design. Talk a little bit about that, about the breakthrough, some of the results you published.

• 25:07Speaker 01

  Chai 2 is basically Photoshop for molecules. So it's a system where you can put in a prompt. We have a target protein that we'd like to make a molecule against in a certain site and maybe some other properties about the underlying molecule. And that comes out a set of molecules that we go and build and test in the lab. And really the breakthrough of Chai 2 was not even just like setting up that whole prompting system, but it was just the success rates that we saw in lab. One in five of the molecules that we tested actually bound to the target as we expected.

• 25:34Speaker 00

  And that was like really surprising for you, right? I mean, your initial, your initial expectation was, was quite a bit lower.

• 25:40Speaker 01

  Yeah, well, our, our target for the year was to get to a 1% success rate because that would be transformative for the field.

• 25:46Speaker 00

  Nice to beat your OKRs by 20 times.

• 25:49Speaker 01

  Yeah, exactly. And, and I think it just says a lot for the field, like what's going to happen now that we can just be building molecules on the computer and just expect them to work in the lab. We didn't believe the results for, for weeks when they came out and we, this is actually one of the reasons we did this across 50 different targets. Most people doing these AI biology projects test like one to five projects. And, and we did 10 times more just because we didn't, couldn't believe it. And we did a bunch of follow up experiments to really characterize that the molecules behaved as expected. And nowadays, you know, it's almost like suspended disbelief. We have these crazy ideas that we're working on at Chai these days. And I, I kind of, I used to just like not believe that they would work, but I'm, you know, now we just like expect it to work. It just changes the way that you work on things.

• 26:34Speaker 01

  If you can, it's essentially like, if you can design it, it'll probably work in the lab, which is, is just like a, a pretty crazy way to, to be, you know, doing drug discovery.

• 26:43Speaker 00

  Did you like see a piece of data that, was it a leap of faith or did you see a piece of data or something that, you know, obviously you were on the inside of this before? Like what gave you the confidence that that was about to change?

• 26:52Speaker 01

  You know, it's a bit of both. You're always taking a leap of faith in science, but that leap of faith is hopefully grounded in data. So we did see signs of life that this was going to be possible. And part of that was just seeing results on the design side. So the models were getting more and more robust designs. So if we started to design even just like a part of approaching, for example, if you just look at the literature and designing parts of an antibody, not designing like the full drug, but just like optimizing existing drug. We saw the success rates going up over time, which we felt was an indicator of the models becoming more accurate. And then the structure prediction side, there was actually a little bit less work. You know, it's been interesting about the structure prediction foundation models is they've been actually more discreet. You know, they've been like big leaps in terms of the quality that are out there.

• 27:37Speaker 01

  And what we realized is that you start with like looking at single proteins, like with Alpha Fold 1 and Alpha Fold 2. Alpha Fold Multi-Mer took this to predicting it's like multiple, like a complex of molecules. How do two proteins interact with one another? And we realized that you could extend beyond that as well. So why can't you bring in small molecules and DNA and RNA? So we started to see that just on the capability side. And the last thing I want to mention is just like the progress we saw on the machine learning side. So the previous, like the Alpha Folds of the world, where they looked a lot more like the older image generation models. You would put in some sequence, you would basically get out a single structure. But diffusion models were coming online and we thought that that was going to be really important for getting the best quality structure prediction methods.

• 28:20Speaker 01

  Because then at the day, like you have some ensemble of hypotheses, like at least we thought the model would have that. If the model was unsure about what the protein structure should look like, it would just choose like one form or the other form. But with a diffusion model, instead of kind of collapsing to the mean there and like averaging what the guesses are, it could just sample different guesses. And then you could have another model that pulls out the best guess. So we saw diffusion models coming online and we felt that that could be a big driver of stronger performance for these methods.

• 28:51Speaker 00

  That's fascinating. It's like in previous generations of models, it's one answer. So you just kind of like compress a lot of different ideas to kind of an average of everything that maybe is therefore not a super interesting hypothesis, whereas Diffusion models, you're able to generate a ton of different ideas. And it's actually in that brainstorm space and all these new ideas. You can then have other models evaluate whether those are good ideas. But the most interesting part is like of those many new ideas, some of them probably make no sense biologically. And some of them actually are quite novel and the ability to distinctly generate a bunch of them is really unique.

• 29:19Speaker 01

  Right. So diffusion models are very good fit for biology.

• 29:22Speaker 00

  Totally. I mean, I assume that's how traditional drug hunting also also works in many ways. Lots of lots of trial and error.

• 29:28Speaker 01

  Yeah, exactly. It's funny too, because if you just think about the early days of like protein folding and structure prediction. People were using these like energy based models. There's this long, rich history of this tool called Rosetta started in David Baker's lab. And Rosetta is basically an energy model. It's like a bunch of terms that describe the energy of a protein. That's sort of what a protein folding model is doing now too with like diffusion models. They're sort of like this is deep connection again, like I call that earlier. The language models and like the the way people, you know, used to do protein folding back in the day, like in a sequence based way, statistical based way. Then there was like the physics based way of doing things, like coming up with these energy based models. And that's actually very similar to what we see with diffusion models nowadays too.

• 30:07Speaker 01

  So again, very deep link between the old days, I mean, all like 20 years ago, way of understanding proteins. And like what's now been rediscovered by the machine learning community.

• 30:18Speaker 00

  Yeah, fascinating. Well, I think one of the cool parts of having the open source model, right? Is you get all these folks using it. And I think you've publicly talked about some of the cool ways. But I love like, what are your some of the favorite or like most counterintuitive or things you wouldn't have expected ways? That you've seen people use the open source model now that's been out in the wild.

• 30:35Speaker 01

  Well, the models to think reason about things different than we do. The design models like that we've done with Chaitu most recently, I think that's even crazier to think about. Cause you're giving the model a prompt and then it's coming up with some answer. The open source model, right? It's a little bit easier to think about because again, it's like a microscope, right? So things might be surprising if you have the wrong hypothesis. But when you're designing things, the models just come up with these solutions that, I mean, you look at the map, you're like, this is just so creative. And they just think differently than our than, you know, the chemists and the biologists do. And you look at these solutions the models coming up with, you're like that, that is, they're like, this is a good idea, but like, will this actually work? And you go test it in the lab and we find out two weeks later, like it worked and it's just crazy when that happens.

• 31:19Speaker 01

  And there's these days, like just that, I think the, you know, the intensity and energy level of the company has just been going up over time because we just see these wild things that the models are coming up with. And then you go test them in the lab and they work.

• 31:32Speaker 00

  Does it feel like we've already had this like AlphaGo move 37 moment of in bio today?

• 31:37Speaker 01

  Well, I think so. That's what the, the solutions that these models are coming up with, right? Like we just prompt them in a pretty natural way. Like I want you to make a molecule that causes like XYZ function to, you know, to this protein. And you know, the models are not really even reasoning through the biology themselves. We're more translating a therapeutic hypothesis into a prompt. And then the model is like solving that, like with here are some sequence that you should go make in the lab. And then it ends up playing out. So maybe that's like the move 37 style thing, right? Like the model's just coming up with those creative solutions. But like the creativity is not what matters at the end of the day. What, what matters is can you solve problems that were just not possible to solve before? Or can you make the processes way more efficient, you know, by orders of magnitude than, than what was possible before?

• 32:20Speaker 00

  And it sounds like in a lot of these, you're, you're taking a therapeutic hypothesis that exists. There are some people have like speculated about and you know, you're obviously testing, you're finding ways to test that hypothesis and really good ways to, to, to basically, if that hypothesis were true, you know, you've, you've had the perfect, you know, molecule or, or compound to test that. What about like the hypothesis generation itself? Like are, are, you know, are you messing around with that at all? Are people thinking about that at all? Or, or is this kind of given the model capabilities today, the right level or the right place in the drug discovery process for us to be applying these models?

• 32:48Speaker 01

  No, that's exactly, I think the right way to think about this. So in biology, actually validation can be a lot easier than generation. And I don't mean like clinical trials, right? Like if you want to validate in a human, that can actually be a lot harder. But if we just think about like in the lab or even in an animal, validation is much easier than generation, than finding something. Right? Cause how do you generate something in the drug discovery world, like pre-AI? There's a reason why we call it drug discovery. What you'll usually do is take a bunch of random molecules, literally billions of molecules and try to find one that has the function that you're looking for. That is a very slow process. It's actually a very inefficient process and it is not always a very successful process.

• 33:30Speaker 01

  Right. Because billions might sound like a lot, but the search base here is like there's more possible protein molecules than there are even atoms in the universe. So just trying a billion random ones, like most of them are just going to be duds. If you can now have an AI model that owns that generation process and then we can go straight to now, you know, characterization in the lab and like the verification verification is a lot easier. So now that generation has become easier as well. The new bottleneck becomes like your ideas and your biology. But the cool thing now is you can start to brute force a lot of those hypotheses quite quickly as well. So, you know, at Chi, like it's almost crazy. Like sometimes people are like, how did you decide to try this idea versus this other idea? And we're often just like trying like tons of them in the lab because it's just easy enough to do it right now. So, so don't spend too much time like being a perfectionist.

• 34:14Speaker 01

  Just go try your 10 biological hypotheses and just brute force it now in the lab and see which one works. So there was this, actually, there's this question that people often would ask us. Are we going to see more lab experiments or fewer lab experiments going forward? Or is it going to stay the same? You might've thought that with AI, there'll be fewer lab experiments, but I think that-

• 34:30Speaker 00

  It just spit out the answer. It's like, this is the, this is the cancer I'm trying to attack. They're like, this is, this is the molecule. Now go do this.

• 34:36Speaker 01

  Well, I'm saying we don't even have to be at that point, right? The models don't have to understand the biology. We can just move that capacity in the lab over to just brute forcing a bunch of different biological hypotheses. And I think ultimately this will allow us to do more creative biology and come up with better drugs and go after things that were undruggable, things that were just, you know, really intractable to do with existing methods. We can now just like try a bunch of ideas and try to find much better medicines than we were, than we were ever able to discover before.

• 35:04Speaker 00

  So how should people think about using chat too? Like, I mean, is this something that should be a part of almost every product at this point? Or are there particular areas where you're like, hey, if you're working with a problem with X contours, it's a really good fit for, for that?

• 35:14Speaker 01

  Yeah, well, so CHI2 is specifically focused on protein and molecule design. So things like antibodies, about half of the drugs that folks are working on right now. And what I mean, I would recommend if you're working on an antibody project, you should come talk to us and figure out if CHI2 can accelerate that work. I think there's both the easy projects, so things that you're pretty confident that you would be able to design a drug today. Let's speed up that process. Like, why not do it with AI? It might actually allow you to go and explore better versions of the drugs than you would have been able to discover before. And then the second area, which I'm even more excited by, is let's use CHI2 to go in and solve problems that were just intractable before.

• 35:56Speaker 01

  You know, there's companies that have spent years trying to discover a drug to address some therapeutic hypothesis, and they just can't find something. But with CHI2, we might be able to discover it at the click of a button. So I think those are the two no-brainer areas where CHI2 should be used today.

• 36:11Speaker 00

  I'm curious, like, there's these, like, legendary, you know, drug hunters of the past that have, like, had, you know, managed to have multiple hits and I feel like are in industry lore. Like, you know, and obviously, they had a very specific skill set. I'm curious if you could just riff on a little bit, like, there's clearly going to be some set of other legendary, like, AI-driven, you know, drug hunters in the future. So what extent are those skill sets going to be similar? Like, what, you know, what do they look different? Like, as we write the biographies of those people 20 years from now, curious what you think on that?

• 36:38Speaker 01

  Yeah, well, I think actually, drug hunters can actually be, will be really important in the story now, right? Because if we can now, again, brute force through those kinds of scientific hypotheses really quickly, now people who can be creative about what problem we're trying to solve can actually become quite valuable as well. So we'll see what happens as the models continue to improve, right? Like, we're being surprised every day. I think that even, like, the general-purpose LLMs, I think, are actually becoming quite good at coming up with biological and scientific hypotheses. But the folks who can now build on these models and then use them to solve really meaningful tasks, I think, have a lot to gain.

• 37:16Speaker 00

  In preparing for this, you know, I listened to your co-founder, Jack, on another AI podcast, and one thing he was talking about, which I found interesting, was just like the importance Of being a product company and not just having the models, but a product on top. And obviously, this is something we've seen in OpenAI and Anthropic and other folks in the kind of foundation model space, you know, building, building product on top. Can you say a little bit more about that and how you're thinking about, you know, building the product area out and like what it will look like in the future for a researcher to use a Chime model?

• 37:45Speaker 01

  Yeah, we think really carefully about product. At Chime, what product really means to us, like when we think of ourselves as a product company, it ultimately means we're trying to solve a real problem. We're not just building models for the sake of building models or because they're cool or because we think AI should be used in biology. At the end of the day, like if we don't need AI and it turns out that just like, you know, plain old software is like what you need to solve these problems, like amazing, right? Like, doesn't matter to us if it's AI or it's physics-based methods or something else. We ultimately just want to articulate the problems that exist today and then go and build solutions to those. And that's really what we mean by product at the end of the day. It's not even like you need a fancy web app or you need to have something that's like particularly sticky. It's just like, what does it mean to build a product? It's working backwards from the problem.

• 38:30Speaker 01

  Even in the drug discovery world, a product might, usually people, by products, they mean like a drug that gets marketed at the end of the day because that's a product that a patient is taking. And in our case, maybe it's Chime looks like a product that a drug hunter is using to make a drug. So it's really working backwards from the problem and that is core to our DNA. And that's what really the whole team is always thinking about is how do we solve this problem and being really open-minded to whatever solution we need in order to find the answer.

• 38:56Speaker 00

  Zooming out, obviously, you're part of this wave of like, there's, you know, just a ton of excitement and funding pouring into this kind of like, you know, foundation model for biospace today. And I'm wondering, you know, for our listeners, if you could just kind of articulate how you think about the landscape and maybe some of the differences in approaches across, you know, some of the big players like isomorphic and profluent evolutionary scales, IRO, and kind of, you know, where you kind of fit Chime into this broader landscape?

• 39:21Speaker 01

  This wave, we kind of articulated the different waves of AI biotechs, right? There was like the wave one and the wave two. And I think ultimately the wave three is probably going to be the most successful, right? And become the enduring companies, in part because the technology is really working today. I think you can kind of segment the landscape into A, like the approaches that people are taking, and then B, into the way that companies are monetizing the advancements that they're building. Think if we look at drug discovery, like just think about how the industry works. The revenue of the whole industry is ultimately like the big bucks are generated in having an approved marketed drug. Right? Patients are buying a drug and a single drug product can make billions of dollars a year. Some of the biggest ones are double-digit billions a year even.

• 40:07Speaker 01

  And I think you got to think about that and work backwards from there. And you could either go for that long shot, try to fully own your drug. You could partner on developing those drugs with other folks. Or you can build a kind of tooling to make it easier for other folks to kind of find those as well. I think that's sort of how the landscape gets segmented. But at the same time, I also think we're very early in terms of how things are unfolding. Ultimately, it's still at the end of the day, you know, there's the whole system of how payers kind of end up paying for those drugs and where that revenue ends up going. But along the way, you know, if you can really accelerate the process, things can really start to change. So I think it's going to be really exciting just to see how the landscape evolves. And I honestly think it's really hard to even talk about each of these players because I think everyone is kind of taking this as it goes because the technology is moving so quickly.

• 40:52Speaker 01

  I think the business landscape is probably going to move a little bit slower than the technology, but I think ultimately it'll probably start to change as well. So it's going to be really exciting to see how the next couple of years unfold.

• 41:01Speaker 00

  I feel like people look at like the kind of history of folks building for Pharma over the last 10, 20 years. The common trope would be, hey, look, to really capture the value here, like you've got to be a therapeutics company. Like ultimately, you're never really going to get anything remotely resembling those economics without taking an asset kind of end to end. You know, obviously, it seems like the industry is in for a big change. Wondering how you're thinking about like how you can capture some of the value that clearly is going to be from these models, you know, without kind of taking therapeutics end to end.

• 41:32Speaker 01

  Yeah, well, again, I think people have actually, most biotech companies technically don't take therapeutics end to end. Most of them exit way before a drug gets approved.

• 41:41Speaker 00

  Right. With assets, though, that like people, other people are valuing on taking intent.

• 41:45Speaker 01

  Yeah, the point is that the value in this space is creating a drug, right? And you need to have some way to really demonstrate that you are increasing the odds of having that drug approved. Are you opening up new kinds of drugs that can be created? Like just ultimately, if the value is being generated in assets, you need some way to kind of inflect that. And that could be like bringing a drug to phase one and then selling it, right? Like that's one way. Obviously, now I've got some drug that has a, it's much lower risk than some idea I wrote down on a piece of paper, right? I've de-risked it and then now I've created some value and people might buy that essentially data package and IP in some cases for billions of dollars already. So I think you just have to work backwards from there.

• 42:25Speaker 01

  But I think what's going to be different this time around is that the, one of the previous approaches in this space just really haven't worked well enough yet. Like imagine trying to build a business around GPT-2. You can do it and people tried. It just didn't work as well as like the current wave of AI companies that are building because the technology is just working a lot better. So imagine trying to be like a partner to pharma on GPT-2 or something like that, or on an AI technology. It's not actually competitive with the methods that they already have. Like some folks would, would buy it, some would try it out. Would they really come back to you for more? I'm not really sure. So I think a lot of these assumptions that people have about the space might start to change. But I, again, I think the story is being written now. I think we're going to see, clearly folks are taking this seriously because the models are really starting to work.

• 43:09Speaker 01

  And I'm, I'm really excited to see, you know, if and how the, the industry changes around that.

• 43:15Speaker 00

  What are like the, the milestones that matter in this space going forward? Like if you could say to our listeners, like, hey, if you read this in the news, like that means the space has really had some sort of like step change. How should people be thinking about when, you know, what to be looking out for?

• 43:27Speaker 01

  If you asked me this a couple of months ago, then I would say, especially for like antibody generation, can we get models that don't require us to do the high throughput screening in the lab? Because when you do the high throughput screening, things are just not scalable. You can't ship like a high throughput screening process to hundreds of drug programs that are being done in the industry. What you can ship across the industry and what you can do in a scalable way, or like AI models, right, or software. So the big bet that we had this year at Chai was being able to create AI models that could design antibody drugs with an accuracy or a success rate higher than 1%. Because if you have that, like 1% was really the mark that we were going for. Because if you could do that, then you could just start to brute force, you know, those ideas, right?

• 44:11Speaker 01

  If your success rate is like 0.01% or 0.1%, if you're talking about screening thousands, tens of thousands, millions, you need to build some fancy thing in the lab to do that. Or use some fancy thing that's been discovered elsewhere, right? Not as scalable. But if you can get to that 1% success rate, and now you can start to brute force your ideas, the way that you discover drugs can really start to change. So that was, I think, what we want to be looking for. We want to be moving the discovery process from the lab into the computer. And the crazy thing is that's like now happening, right? Like with Chai 2, we announced that we can actually design these things with a success rate way higher than 1%. We have a 20% success rate in the lab. And now the molecules that are coming out of Chai 2, they definitely need more work before, you know, we feel comfortable putting them into a patient. The molecules aren't perfect.

• 44:56Speaker 01

  But now we can move from, you know, those success rates into can the model be, you know, listening to our hypotheses in a more accurate way? Can it be coming up with more sophisticated hypotheses? So I think we should really be measuring how much is happening on the computer. Because ultimately, that's way more scalable. It also compounds on itself. Because as the models get better, if we think about data generation now, let's actually use the model to generate data. Or let's use the model to enumerate hypotheses and kind of bring that back into the models. So I think that's really the thing to be looking at, is how much is being done on the computer? And then how is this actually going to compound? And like really understanding that to kind of predict what the rate of progress is going to look like.

• 45:34Speaker 00

  Obviously, I think, you know, a common thing people talk about is, oh, the day the first AI-designed drug is approved, like that'll be this like insane milestone. Any like interim milestones in between here and there that you also are like, hey, these are important as well?

• 45:47Speaker 01

  I think it's less important to me to think about like what the first AI-designed drug is like. There arguably are AI-designed drugs that are in the clinic, or you might even argue approved. There's approved therapies that use like computation as part of the process, right? So, I mean, is that AI or is it software? I mean, I'm not sure it really matters. I think the bigger question is, can we actually deliver immediate value? Can we go and use these technologies to create drugs that were not possible to be discovered before? I think that's really the thing to be looking at. And maybe that's even the right answer to your previous question. Like, how do you measure the rate of progress? What I was saying before is maybe how like a technologist should look at this to kind of think about like what's happening next with the AI models.

• 46:28Speaker 01

  But if you think about this from like a pharma perspective or a life sciences investor perspective, right? How is the industry going to change? I think it's more about like, are we solving new problems that weren't solvable before? That's ultimately how you start to see like the market growing, right?

• 46:41Speaker 00

  Is it possible that like we get to models that are actually like, you know, the future we've always dreamed of? But because it takes five, 10 years for, you know, to go through full trials and really know what are these drugs work in humans? That like, we may get there and then, you know, a decade later realize we got there, I guess. What is the, the length of time that it takes to, you know, to realize that?

• 47:01Speaker 01

  Well, you need to have metrics that don't require you getting all the way through clinical trials because the feedback loop is frankly just too slow. And then this has been one of the problems as well. Like, one of the reasons why progress is moving so much faster now is when more of the design happens on the computer. Instead of waiting months in the lab to figure out whether your thing works. If I go design 100,000 antibodies, it's going to take me like, you know, weeks to months to figure out whether those designs were successful. If I only have to test 20 because my model is so good, I can find out in two weeks. So it's really important to have tight feedback loops if you're going to make progress. And that's why I think we need to be translating, you know, ultimately, um, let's say some mechanism we want to achieve in a human.

• 47:43Speaker 01

  We need to figure out a good way to measure that in the lab so that we can get the right feedback loops going.

• 47:47Speaker 00

  It's interesting. I feel like sometimes I'll like, there, there's so much excitement about these models. And sometimes I'll, I'll talk to someone in the traditional bio world and maybe they'll, they'll throw a little cold water on it. Be like, well, look, you know, you can find things that bind to drug, that bind to targets. That's cool, but like, you know, let me know when they've run a full battery of studies and it's equal or better than something else that we've generated before. I guess I'm curious, like, what do you think's fair about some of the critiques from folks who may be a little more skeptical about this? And, and what maybe do they get wrong about what's happening in the space today?

• 48:14Speaker 01

  I think that's totally fair, first of all, right? If, if someone discovered a drug and then you just discovered that same drug again, but you used AI. Like, why does that matter? How does that actually change the game, right? Uh, that's why I think what we really should be shooting for with these methods is just upleveling the kinds of molecules that that can be made. So take antibodies, for example, uh, we tried to, we've shown that you can take a specific site on a target and then go generate an antibody that binds that. But imagine if we can now go beyond that and we can say, well, since it's now so much easier to discover those molecules, what if we actually make molecules to several parts of the target and then stitch them all together? That's something that we might call, that feel called a bi-power topic, for example.

• 48:57Speaker 01

  They're much more difficult to discover in the lab, but now if it becomes a lot easier to discover them, I think you might see an increase in those kinds of molecules that are being created. So I think just like up-leveling the kinds of therapeutics that we go after is like the right thing to focus on. The other thing, too, is like the critique about, oh, do these molecules look like the same molecules that I would discover the Trishel first? Like, that's actually getting resolved, I think, much faster than we would have thought, but it's still not going to generate that kind of impact. I think what people want to see is like, how does this actually change the game for like what I'm looking at? Like, I'm a life science investor, or I'm a pharma company, I want to acquire a drug. Are you creating a drug that like just wasn't, you know, possible to be discovered before? Those are really, I think, the things to be focused on. Or can we make better drugs that just weren't possible before?

• 49:38Speaker 00

  It feels like we're going to move to a future where these models are ubiquitous in drug development and really change many things about how drugs are developed, the difficulty, the specificity of that. And I'm wondering then the ripple effects for what that has for the broader pharma industry and how, you know, what it means to be a really successful pharma company, which today I think is, you know, mostly acquiring early-stage IP and then being really good at running it through trials and sales and marketing. Maybe that stays the same, or maybe the ecosystem has new players that are needed or evolves in some way. I'm sure you've thought about this before. What do you think the evolution looks like?

• 50:09Speaker 01

  It's a matter of timeline, first of all. So again, if we look at the spectrum from like where we are, let's call it two months ago, right, pre some of the recent advancements, and then we also projected out a couple of years, like imagine it would be, it would be possible to build what we think about as shy as zero-shot drug candidates, like your first pass out of your model, you actually, maybe it doesn't have to be 100% success rate, but within the first batch of molecules you get, one of those is the one you're actually going to put into a clinical trial. You know, now the question becomes like, what is your therapeutic idea? Many people will have that idea and they can kind of translate that into molecules. So I think ultimately like patients win here, right, because they get better molecules. I do think pharma wins here as well, because if we think about them as capital allocators who are

• 50:52Speaker 01

  Acquiring a lot of the drugs that are out there or working on them their own pharma these days about like two thirds in licensed drugs and acquisitions and about a third like internal pipeline. But I think they start to win just by being able to create better medicines. I think it's really interesting to think about what happens with biotech. You know, how do you create like a competitive advantage here? If it's now much easier to discover molecules, so you might actually see. You could see this going in two directions. You could see pharma actually doing a lot more discovery because it's become easier like. Do we really need biotech anymore to be innovating on the discovery or? Are AI models doing it? So do we just go partner with like an AI company to make this happen? Or you might see the opposite. You might see that it becomes way easier to discover drugs. So pharma is like.

• 51:34Speaker 01

  Hey, like, we don't want to touch this. Let's just let all these guys kind of like discover the molecule test in the lab. We'll buy it from them. So you can really see it heading in either direction. I think regardless, it's probably not going to be like night and day here. There's probably going to be some balance. I think you'll probably still have a healthy ecosystem of biotech. And pharma kind of deploying these methods. So I think you'll, I think change management is going to be really hard in these organizations as well. So I'm not as convinced that, you know, throwing a model over the wall is going to work here. It's kind of goes back to like the more product mindset.

• 52:05Speaker 00

  I'm not sure that really works in any industry, maybe besides selling to developers or something.

• 52:09Speaker 01

  Yeah, but even there, right? Like developers also like. Throwing some tech over the wall, like you want to get those continued updates and you want to have something where you can kind of orchestrate all the tools that are coming in. So. I think people, again, what's even harder now in biotech is you're not selling to developers. You're also selling to an industry where these tools have not worked for a long time. And also where things just haven't changed that much, like the actual discovery process has not changed that much in the past like 20, 30 years. It's very different in tech, right? Tech people are used to the ways there was. There was mobile and now there's been AI and like people are just used to change and they really adopt that. And biopharma hasn't always worked that way, but at the same time, I think people are just going to look at the data and they're going to say, hey, like.

• 52:50Speaker 01

  Oh, Pfizer got access to this like really good AI tech by working with a company like Chai. How can we not get access to that? Right? Like, imagine if you had, if you were the one pharma that doesn't have access to this kind of tech. You're going to fall behind. So I think we will see kind of this like slow progression at the start and then it's going to like. Rapidly start to expand because once a couple of folks like show. That this can really make a difference in the bottom line. Other folks are just gonna have to like hop into it as well.

• 53:18Speaker 00

  Totally. I mean, it would be it would be crazy not to not to use it. I mean, obviously, you said there has been maybe, you know, some some false starts in the past in some of these spaces. You know, when you're coming in and kind of pitching, hey, this time's different. Is there like a killer slide or like is it mostly demoing on their own data or like what's like the the one killer thing you do that people realize, oh, yeah, like this this really is different?

• 53:39Speaker 01

  Yeah, I think the killer thing is hard challenges. So show examples. There are these canonical examples that are have been really hard for people to design before. And I think whenever you can show data around those, it can be very compelling. So that's like the number one thing. If someone has worked on a project for five years and then your AI model does it in two weeks, that that's pretty cool, right? There's one example I think of. There is a company that spent like five, $10 million like working on a project. And then we prompted the AI model with, you know, that therapeutic hypothesis and we got to the answer within like two weeks. Yeah, just to answer your question, like, what is getting people excited? It is showing them problems that that they couldn't do before.

• 54:21Speaker 01

  Or things where, you know, if they spent six months to a year trying to discover anybody and we did it in two weeks. That's again, it's another thing that that's pretty mind blowing.

• 54:29Speaker 00

  Well, this has been a fascinating conversation. We always like to end our interviews with a with a quick fire round. And so maybe to begin, what's one thing that's overhyped and one thing that's underhyped in the AI bio world today?

• 54:39Speaker 01

  Underhyped, I think, would be discovery of molecules. This is working really well now and people should be paying attention. I think what is overhyped would be, you know, AI designed molecule in the clinic or something like that. Doesn't actually matter at the end of the day whether the molecule came from AI. What matters is like, is this industry actually going to improve? Can we get better drugs to patients?

• 55:00Speaker 00

  I assume you, you know, you take a lot of lessons from other parts of the AI world. What like other areas of research outside of bio are most relevant to the work you're doing and you kind of pay most attention to?

• 55:12Speaker 01

  Yeah, we're doing very interdisciplinary. So of course, all this stuff on LLMs, video modeling, proteins can be dynamic, and then even physics, just really understanding how atoms interact at the end of the day. It gets down to the underlying physics and chemistry.

• 55:25Speaker 00

  Yeah. So there's the famous, I guess, Erem's Law, which is Moore's Law reversed, which is how much less efficient we have at discovering and making advances on the biotech side. Do you think that reverses in the next five years?

• 55:40Speaker 01

  I hope so. That's the goal here.

• 55:42Speaker 00

  Yeah. If you were a betting person, though, you think more likely or less likely that in five years, we're able to stand the tide there?

• 55:48Speaker 01

  Maybe that's also, I should add, is like an overeied thing. It's not just probability of success that matters. It's like, what is the value that you are generating? If you take a drug that, if you just make a copycat drug, there's already a therapy there and it doesn't really change the standard of care, it doesn't matter as much if the probability of success is high. So I actually think we want to be ambitious. We want to be taking risky projects. The industry needs to be taking more risk, not less risk right now. So probability of success, like Erem's Law, it's kind of like the wrong thing to look at. I think what we need to be looking at is like standard of care for patients, like work backwards from there.

• 56:20Speaker 00

  Well, this has been a fascinating conversation. I'm sure there's all sorts of threads people want to pull on. I want to leave the final word to you. Where can folks go to learn more about you, about Chai? Anywhere you want to point people, the floor is yours.

• 56:31Speaker 01

  Yeah, I mean, check out our website, chai-discovery.com. We've got a couple of links there to some of the breakthroughs that we worked on. We've got a video and a technical report about our recent breakthrough, Chai 2. You can go to our GitHub repo and you can find the open source model, Chai 1. And, you know, reach out to us if you're excited about what we're doing.

• 56:50Speaker 00

  Amazing. Well, thanks so much. That was a ton of fun.

• 56:53Speaker 01

  Thanks for having me, Jacob.