He Raised $70M to Cure Every Disease With AI | Samuel G. Rodriques, Edison Scientific

Samuel Rodriques https://www.linkedin.com/in/samuel-g-rodriques-080a9b22/ left physics because there were no unsolved problems left. Instead, he built an AI scientist named Kosmos to cure every disease, solve aging, and map the human brain. In this episode: The cure his AI proposed for blindnessWhy he would never touch a peptideWhether we'll need human scientists in 20 yearsWhat's stopping America in drug discovery Connect with us here: Samuel Rodriques https://www.linkedin.com/in/samuel-g-rodriques-080a9b22/Edison Scientific https://www.linkedin.com/company/edison-scientific/Lukas Biewald https://www.linkedin.com/in/lbiewald/Weights & Biases https://www.linkedin.com/company/wandb/

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  If we want to go and cure all diseases, understand how the brain works and solve aging, and AI seemed like the right way to do that. When it comes to the world as we experience it, most things are pretty well understood, except for biology. And to this day, we don't have the wiring diagram for the human brain. We like kind of have the wiring diagram for the fly, and that's like the best that we have. And so when we got started in 2022, remember, we didn't have any notion of like reasoning models or whatever, right? It was like the first multi-agent system that we showed that was capable of like doing the full loop of scientific discovery. And it came up with a new hypothesis about a way to treat a form of blindness called age-related macrogeneration. It actually just got published in Nature two days ago. Since we launched Cosmos, people have probably used Cosmos to make like 20,000 or 30,000 novel scientific findings, which is wild. The pharma companies

• 0:42Speaker 00

  of the future will be much more lean. You're going to be able to pursue many more drug programs in parallel than you can today with the same number of people.

• 0:50Lukas Biewald?

  There's a long history of AI overpromising in drug discovery.

• 0:53Speaker 00

  Oh, yeah.

• 0:54Lukas Biewald?

  Is this time different?

• 0:55Speaker 00

  I promise you this time it is different.

• 0:59Lukas Biewald?

  You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. All right, I'm here with Sam Rodriguez, the founder and CEO of Edison Scientific and Future House. He's right at the forefront of using agents for scientific innovation, and he's got a lot to say on how to use agents well and science itself. I really hope you enjoy this podcast. All right, Sam, thanks for doing this. I've really been looking forward to it.

• 1:27Speaker 00

  My gosh.

• 1:29Lukas Biewald?

  All right, so let's get started. So you had this incredibly promising, successful career in neuroscience and bioengineering. I think you studied theoretical physics originally, and then you moved into AI.

• 1:42Speaker 00

  Yeah, right. What went wrong, right?

• 1:44Lukas Biewald?

  What were you thinking?

• 1:47Lukas Biewald?

  Why did you do it, Sam?

• 1:48Speaker 00

  Why did I give up such a promising career? Yeah, great question. Yeah, so I started doing theoretical physics, did quantum information. The problem with physics is there are actually no unsolved problems left in physics. So, you know, it's like a slight exaggeration, you know, we need to build quantum computers, and there are like some... Interesting, important material science problems. And we still like, don't know how like the universe works, right? But like, actually, you know, I think one of the key insights is, if you look around at like, any phenomenon that you can see in the room today, like in this room that we're sitting in, I could probably explain to you how that phenomenon works down to the level of like, some atomic particles, right?

• 2:30Speaker 00

  Except for like, you and like, or like, you know, the mouse. And like, you know, why does the mouse do what it wants to do? Why do you like, you know, uh, wake up at the time you wake up? And why, how do you get sick and so on, right? And so like, actually, like, when it comes to the world as we, as we experience it, everything is, is kind of, most things are pretty well understood, except for biology. Right? And so that was originally what got me into biology. And then, and so I did my PhD at MIT. I was, um, uh, I'm, I'm an inventor at heart. I invented a bunch of different technologies. And, but, but like, the kind of key thing that I learned at MIT is that,

• 3:12Speaker 00

  you know, doing biology is that really, um, there are three things that you need to do to, to do, um, science. Right? You need capital, you need money, um, you need logistics, which is like, do you have the things that you need to run the experiments you need to run? In the place when, you know, to run them when you need them. And you need talent.

• 3:31Speaker 00

  And capital scales and logistics scales, right? And talent just like, does not scale. And fundamentally in biology, we're limited by, by talent. And so, um, the thing that I got to think about is like, how do we remove talent as a bottleneck in science, right? Like, if we want to go and cure all diseases, and, you know, understand how the brain works and, uh, solve aging and so on. We just like, need to figure out how to scale talent. Um, and AI seemed like the right way to do that. So, um, yeah, I mean, it was kind of crazy. I just like, basically, I like built up this career and I kind of just like, threw it, you know, jumped off, threw it away and, and jumped off a cliff. And that, that was pretty wild, but, but, um, uh, it's been working out well so far.

• 4:16Lukas Biewald?

  And, and originally the company that you started now, which you better explain what that does. I think it started as a, a new kind of not-for-profit, right?

• 4:23Speaker 00

  Yeah, yeah, yeah. So. Um, uh, we basically, so this was 2022. Um, and I had kind of figured out that, um, the most important thing, it seemed like that was going to happen in science in the next, in the next 10 years, was going to be like, you know, figure out how to build an AI scientist, precisely because it was going to unblock talent, right? Um, and, uh, but at that time, I mean, i- in biology, you're used to things going really, really slowly. Right? Um, and so we were at this like basic, we were at this point where, you know, GPT-3 was out and it was like, it could kind of say things, it was a very impressive, you could kind of see where things were going.

• 5:04Speaker 00

  Um, but it seemed like it was going to take a really long time, um, to get there and, and I didn't know how to commercialize it. I didn't know what it was, you know, and so, um, we were like, well, we should, we should do this as a nonprofit, just go and do the basic research. Um, uh, which was, which was great. And so I teamed up with Andrew White, my co-founder, um, who is really a pioneer in the space of, uh, AI agents for science. He was working with OpenAI, um, on, uh, GPT-4 at the time and was a professor at the University of Rochester. Um, and, uh, um, uh, kind of like, we had the same vision for building AI scientists. Um, I kind of like knew the biology side and Andrew actually knew how to do it technically, right?

• 5:48Speaker 00

  Um, and so we teamed up, but it was just, it, it felt like the right thing to do was to be a nonprofit, just because it was going to take, you know, we thought it was going to take five or 10 years, which was complete nonsense. I mean, that was like totally not, right? I mean, as we all know now, um, uh, within, we, we launched Future House, we announced it, um, in, uh, um, uh, in, in November, um, of, uh, 2023. Um, and like within two years, we had like extremely powerful AI scientists already, right? So the timing thing was not something that ended up being very important. I, I will say like, I think that the non-pro, the fact that we started, that was not the only reason we started Future House as a nonprofit.

• 6:27Speaker 00

  We also like really wanted, um, and, and continue to want those technologies to like benefit the entire scientific community. And so as Future House, we open sourced a lot of stuff, which I think is also super, super important. But yeah, and then what just ended up happening was by kind of spring of 2025, you know, one day like Andre Karpathy tweeted or something and then all of a sudden the entire world knew what AI agents were. And we started getting phone calls. We like literally went from being like, we would have to explain to people in our decks, like, what is an agent? And how is it different from a language model? We went from that to being like, to having like,

• 7:10Speaker 00

  senior executives at pharma companies call us to say, oh, my God, like, how do we use your agents? How do we use your thing? And so it became evident pretty quickly, we were gonna have to have a for-profit spin out in order to satisfy that demand, right?

• 7:25Lukas Biewald?

  Do you feel like there's a role for nonprofits in technology then? I mean, we now have two examples of nonprofits developing interesting technology and kind of mainly turning it to for-profits to get that technology out into the world. Is there something broken about the nonprofit model?

• 7:42Speaker 00

  No, there's, first of all, is there a role for nonprofits in technology development? Yes, absolutely. Is something broken about nonprofits? Like, no, definitely not, right? Like, this just feels like these, this is like, you know, nonprofits just, or it feels like, basically, what this is, is for-profits doing the thing that for-profits do well, which is scaling, right? So, first of all, let's back up. When I was in my PhD, I was working on connectomechs project. So this is a project that is, basically, connectomechs is figuring

• 8:25Speaker 00

  out how to map all the connections between neurons and the brain. And, you know, just like, can we, a core piece of understanding how the brain works is figuring out how to, like, get the wiring diagram. And to this day, we don't have the wiring diagram for the human brain. We don't have the wiring diagram for the mouse brain. We like, kind of have the wiring diagram for the fly. And that's like, the best that we have. And it's really, really difficult to understand how the brain works without the wiring diagram. Okay, so, we're gonna go and, we wanna go and map the wiring diagram in the brain. You, it's really, really challenging to do that. Basically, because it involves tracing these tiny, tiny fibers.

• 9:09Speaker 00

  Tiny meaning like, you know, one one hundredth of the width of like a human hair. Right? Through like this gigantic tangle of the brain. And you have to make no errors because if you make errors, then you are like connecting neurons that are not actually connected to each other. Right? It's like, you know, if you try to imagine mapping like, like all of the roots in a field of grass. Right? Like, it's like that, but way, way harder. And so we, and so that was the problem I was interested in. And I tried to do it in academic lab. And I was like, wow, there's like no way actually that I'm going to be able to gather like the engineering resources, the like, you know, the capital, the talent that I'm going to need in order to do this.

• 9:54Speaker 00

  And but also you can't do that for profit. Right? You actually can't go and in a for profit setting, like get money to go and map the brain because what are you going to do with the map of the brain? Right? You're going to go like develop drugs and that will take like 20 years or something. And it's just like, that's not an attractive proposition for investors. And so I was like, okay, I can't do this in academia. I can't do it for profit. So basically, there's actually no way for me to do this right now. Right? And so I proposed these things to have focused research organizations, which is this idea that for some problems like mapping the brain that are too big for academia, but that can't be done for profit, we need a third way. Right?

• 10:35Speaker 00

  We need some alternative kind of structure. And the idea of a focused research organization is that it is a nonprofit that looks a lot or that operates, you know, like a company. Right? Not like your kind of like slow moving foundation that you think of when you think of like normal, normal nonprofits. It's like a, like, you know, fast moving, hard charging research organization that has a specific goal, pursues that goal. And then when it's done with that goal, it's, it's done. You can actually like spin down the, that nonprofit. Right? And so actually, since then, we've gotten philanthropists to fund a bunch of these. A bunch of these focused research organizations and they're off and they're doing projects that you just like could not imagine happening in a for-profit setting.

• 11:19Speaker 00

  And so, and sometimes when they're, you know, if they really work out, if they hit, right? And they go, they go really well, then indeed, what is the right step afterwards? It might be a spin out for-profit. And then it's kind of always been the idea. And so, you know, with Future House, we started it goal of building an AI scientist. We didn't know how to commercialize it. Nonprofit made sense, right? Plus, we wanted to be able to share the basic research with the world. Then, you know, it went really well. We made a bunch of progress. What is the next, you know, most sensible step? Spin out a for-profit, allow you to scale. It all, you know, this was actually not the plan. Like, we didn't think at the beginning.

• 12:02Speaker 00

  We thought maybe we'll spin out for-profit in like five years or 10 years. We certainly didn't think it would take us two years or three years. But it was kind of like, you know, I think this is just like, it's a consequence of success, right? As it was also in, in OpenAI's case.

• 12:18Lukas Biewald?

  So what did you see that made you feel like the technology was even more promising than you thought?

• 12:23Speaker 00

  Oh, yeah. I mean, look, when we were, when we got started in 2022, remember, you know, well, we had ROHF and we had like InstructGPT. And so you could like, you know, the language models then like knew how to respond to you, right? But we were very much in a world of like, these models, they just can't. We didn't have any notion of like reasoning models or whatever, right? They, they just seemed, it was very rudimentary. And we, it was like, really, it was like 18 months or 24 months until they started to make discoveries. And the first discovery that they made was actually this paper. We described it in this paper about a system called Robin,

• 13:08Speaker 00

  which is a multi-agent system. It was like the first multi-agent system that we showed that was capable of like doing the full loop of scientific discovery. So hypothesis generation, experiment planning, and then we would go and we would run the experiments and then it would go, it would analyze the data, come up with new experiments. And it came up with a new hypothesis about a way to treat a form of blindness called age-related macrodegeneration, specifically dry age-related macrodegeneration, which affects like five or 10% of people over the age of 50. And this was in, you know, May of 2025. Our agent came up with a new way of treating it, proposed a new way of treating it that

• 13:50Speaker 00

  we were able to go and validate in some experiments in the wet lab. And subsequently, we've been able to validate them in animals. It actually just got published in Nature two days ago. And that was the thing where we looked at it, we're like, oh shit, like, you know, the future is here.

• 14:09Lukas Biewald?

  That makes sense.

• 14:10Speaker 00

  Yeah.

• 14:10Lukas Biewald?

  And so where's it gone since then?

• 14:12Speaker 00

  It's just like, I mean, that was like discovery number one back in May. We, we since released an updated version of our agent called Cosmos. This is like a much more powerful version of Robin that is, Robin was kind of on rails. So Robin could do, you know, it was like orchestrate, we like manually orchestrated a bunch of different agents and to get it to do one specific thing, which was identify new ways of treating diseases, particularly around kind of drug repurposing. And with Cosmos, what we did was we built in an orchestrator so that it could orchestrate itself.

• 14:52Speaker 00

  And then we also built in this notion of world models so that, which is really basically a very sophisticated context management tool, but basically allows Cosmos to build up a kind of integrated notion of the knowledge in a field over the course of like dozens, hundreds of sub-agent runs, which then allows it to, is a core to the way that it makes discoveries. But since we launched Cosmos, we've probably, you know, people have probably used Cosmos to make like 20 or 30,000, come up with 20 or 30,000 novel scientific findings. Um, which is wild. Yeah.

• 15:29Lukas Biewald?

  So I think anyone that's used AI for any technical application, maybe even non-technical applications, notices that the intelligence is really spiky in surprising ways. There's some things that does so much better than a human and some things just shockingly worse.

• 15:43Speaker 00

  Yeah.

• 15:44Lukas Biewald?

  Tell me about-

• 15:45Speaker 00

  I asked it to write a press release the other day. Oh my God. I asked it to go into our Notion, into our Slack and to look up all the details for a deal that we're going to announce shortly and draft a press release and it was terrible. I was like, I cannot, I don't know. Yes, extremely-

• 15:59Lukas Biewald?

  I appreciate you saying that because I've had a few guests on this show lately that like kind of refuse to acknowledge any weakness in their, in their algorithms. And it gets really boring and weird because you know these models are going to be struggling with some stuff. But yeah, like, I mean, but actually in the practical fields, like where do you feel like it's really strong and where do your customers get surprised that it can't do something?

• 16:18Speaker 00

  Okay, so it's strong in two areas. It's strong on things that are verifiable and it's strong on things where throughput matters a lot. Okay, so verifiable means, you know, that you can tell whether or not an answer is correct. And this has always historically been where AI has been strongest because it allows you to get feedback very quickly and then you can go and you can use reinforcement learning and right. AI is strongest where there's a lot of data and where it's verifiable and where the problems are verifiable. So coding is verifiable, math is verifiable, science is like very much non-verifiable.

• 16:54Lukas Biewald?

  Right.

• 16:54Speaker 00

  It's verifiable in principle because you can go and run experiments.

• 16:56Lukas Biewald?

  Right, but the loop is expensive.

• 16:58Speaker 00

  The loop is expensive. It takes forever.

• 16:59Lukas Biewald?

  Right.

• 17:00Speaker 00

  Right? I mean, like literally people are like, oh, why don't you, you know, just do like closed loop RL to like teach it how to find new drugs. And I'm like, because that loop is like, I need to go and prove that this drug is safe and then go and dose some humans and then wait for six months. And it's going to be like, the loop is going to be three years long. Right? The other thing I hear a lot is like, you know, why don't you go and like build a gigantic science warehouse where you're just like have pipetting robots doing experiments over and over again, which is a better idea. And there are a lot of things for which that is, is probably- Is great and very and a very useful thing to do. But in general, right, that requires the experiments that you're doing in the lab to be reflective

• 17:45Speaker 00

  of what you want, right? In our case, like, we want medicines for humans. I can't, like, test whether a drug works in a human using a pipetting robot in a lab. I have to test the human, right? And so your model is only as good as what you what you train it on. But yeah, so basically, there are two. The AI is good at two things. It's good at task that are verifiable, and it's good at tasks that are

• 18:10Speaker 00

  that require high throughput. On the high throughput side, this is, I think, in science where it has really shined so far because it's able to go and consider so much more evidence than any human is able to consider, is able to test so many more hypotheses than a human is able to test. You have to control for P-hacking and multiple hypothesis testing and so on. But like,

• 18:39Speaker 00

  there's just like, in biology, we have a problem with throughput.

• 18:44Lukas Biewald?

  And this is sort of on like statistical, like synthesizing data or reading papers? Or what are we talking about?

• 18:50Speaker 00

  I'll give you an example. One of our future hospital fellows is interested in figuring out how to cure autoimmune diseases.

• 18:55Lukas Biewald?

  Okay, all of them?

• 18:57Speaker 00

  Like, I mean, ideally, yeah. I mean, you know, well, we'll start with like, we'll start with we'll start with one, but but actually, we don't need to be picky. In order to treat autoimmune diseases, what you need is ways to manipulate the immune system. Okay, well, what in nature? If we go to nature for inspiration, where has nature figured out how to manipulate the human autoimmune the human immune system? Okay, parasites. Like, you know, before modern hygiene, humans just lived with parasites. And the way that that worked was that the parasites had figured out how to manipulate the immune system to turn down immune reactions. Wow, you know, if we could go and we could figure out how they do it, maybe we can do it on ourselves in order to cure autoimmune diseases, right? This is the idea.

• 19:41Speaker 00

  But, you know, there are Many, many, many parasites. And each parasite has a genome with thousands or tens of thousands of proteins in it. What is the mechanism? How are you going to go and figure out what mechanisms the parasites use to regulate the immune system? We're actually now using our agents, every single protein in any parasite genome. Like, we are running our agents to look at that protein, look at its structure, look at its sequence, right? Look at the context in which it appears in the genome. Look at the biology of that organism and figure out, is this, could this be a candidate for how this parasite moderates the immune system?

• 20:25Speaker 00

  From that, we've come up with a short list of proteins that we are, that we need to test, and we're now going and testing them in the lab. Right? That is something that just previously, there was no way to do that. It would have been completely impossible.

• 20:36Lukas Biewald?

  Right. But presumably, the hard part of that is building tools to, when you say, "look at a genome."

• 20:41Speaker 00

  Yeah.

• 20:42Lukas Biewald?

  It's obviously not feeding the genome into the context window. It's finding tools to-

• 20:46Speaker 00

  Correct.

• 20:46Lukas Biewald?

  -to, to look at what's going on.

• 20:48Speaker 00

  But the biology, but like, I call this high-throughput reasoning. Yes, you can go and you can just run some homology, you know, you can come up with an algorithm to look at the, you can use a model of protein structure, and you can just look at the structure of all the proteins. That doesn't tell you what role that protein might play, right? In the biology of the organism, which is fundamentally what we're interested in. Another good example is bacteria. There are many proteins that just have no known function. And in bacteria in particular, and if we knew what their functions are, we might be

• 21:33Speaker 00

  able to figure out, like, you know, new way, like, create new bioengineering tools, right? That create, find new enzymes that do functions that we can't do today. Right? And one of the ways that you can do this is that bacteria organize their genomes into these units called operons that are all functionally linked. And so if you have one protein- That's known inside of an operon and then also in that operon, you have a protein that has unknown function. You can reason about what the function of that latter protein based on the the former protein. Right. That has the known function and you know, that's something where you really just need like intelligence.

• 22:13Speaker 00

  Right. In order to go in and think about what is the biochemical role of the of this operon? What does it role does it play in, you know, the bacterial life cycle? And based on that, can we come up with hypotheses for what this protein of unknown function might do?

• 22:31Lukas Biewald?

  Mm hmm.

• 22:32Speaker 00

  Right?

• 22:32Lukas Biewald?

  But when you to test that hypothesis, presumably you have to do something physical, right?

• 22:36Speaker 00

  Absolutely. And so in the case of the autoimmune, the parasite proteins, we've now gotten a bunch of them synthesized on

• 22:45Speaker 00

  on a DNA chip and we're going and we're screening them to see what effect they have on on T cells. Right. And and so absolutely, you can't just reason your way to solving science. You have to do experiments, but coming up with hypotheses, right, is something that is limiting. Is one of the is like one of the steps that is limiting. And that is something where the models are able to help us.

• 23:12Lukas Biewald?

  Got it. So presumably your customers are using your models mostly for drug discovery. Is that right?

• 23:16Speaker 00

  Yeah. And when the customers, so there are the customers use this in two ways, right? So I just talked about the hypothesis generation. The other place where, you know, you can imagine that these models will have a major impact is in the kind of operational work of science. Right. So, you know, how do we

• 23:43Speaker 00

  do how do we like, you know, actually get all the materials together that we need for to run an experiment? Once you figure out what experiment you want to do, you need to like literally go and like order them and organize them and figure out what the protocol is. You, you know, when you get into testing on humans, you need to like coordinate all the resources that you need in order to do the the clinical trials. You need to figure out what clinical trial sites are, you prepare your regulatory documents, like all of that, right, is a part of the process of doing science. And so we get used in both areas. So we get used both for the kind of early stage hypothesis generation, and then also for the actual operational tasks of like of development. How do we turn this drug into, how do we get this drug through the pipeline and to patients as quickly as possible?

• 24:28Lukas Biewald?

  That's really interesting because weights and biases customers do both also. So we at first saw mostly drug discovery applications and then post L L Ms, we started to see all these operational applications. And I started to think maybe the operational applications are more important. Like certainly, I think more dollars goes into that and it's more of the bottleneck. Where do you think it goes? Do you have like a passion for the drug discovery side and kind of want to focus on that? Or do you think the bigger business here is maybe on the operational side?

• 24:55Speaker 00

  I mean, there's this very funny situation in science where progress comes from discovery, but the commercial value is in development. Okay. And the reason is that discoveries just pan out so infrequently. Right? And you can't tell whether or not they're valuable until 10 years or something after you came up with the your hypothesis. And so really what matters is being able to run experiments faster. What matters commercially and frankly, what matters practically is being able to run experiments faster. So you can come up with as many hypotheses as you want, but the experiments that matter are human clinical trials. Right? That's what tells us whether or not the hypotheses actually work in practice.

• 25:42Speaker 00

  If you want to accelerate the process of kind of medicine, you need to make those experiments faster. That's where the commercial value is. That's where a lot of the scientific value is. This is not to say, which is not to say that the early stage discovery is not important. I mean, it's also critically important.

• 25:57Lukas Biewald?

  So we've had, we've had a whole slew of guests on this podcast doing different parts of the drug discovery pipeline. We've had CEOs and researchers. We've even had notorious pharma bro, Martin Shkreli, kind of giving his take. He was actually kind of down on the whole drug discovery with AI thing. I'm curious how you think about the entire drug discovery market and how it's changing. Like what's kind of working, what's not, like what's changing, what's static?

• 26:23Speaker 00

  Yeah, great question. So, I mean, I, I think, obviously, the, the, oh, there are like two big areas where AI is having a major impact, right? The first one is what I talked about, which is like all of the reasoning, and that includes both the like hypothesis generation, which we do, and the operational aspects, the development, how do you get these molecules through to patients faster, which we also do, right? The, the second major important part is coming up with the molecule, right? So if I, you know, have this hypothesis that agonizing the GLP-1 receptor is going to cause people to lose weight, right? In order to test that, I need to have a molecule that inside of a human

• 27:06Speaker 00

  is going to go and agonize that receptor, bind to the receptor, activate it, and so on, right? And that is way harder than it sounds, because the odds are, your molecule, you know, if, let's imagine that you want to do this orally, right? So with a pill that you just swallow, odds are it's not going to go into your bloodstream. Odds are that if it goes into your bloodstream, it's just going to get filtered out immediately by the liver. Odds are, even if it doesn't get filtered out by the liver, it's not going to stay in your blood long enough to have the effect. You know, odds are that even if it activates the receptor, maybe it activates the receptor in the wrong way. Odds are, even if it activates the receptor in the right way, maybe it activates 100 other receptors that you don't need, want to activate that lead to bad side effects, blah, blah, blah, blah, blah, blah, blah. And then there's the hypothesis that your, your original hypothesis

• 27:51Speaker 00

  might not work, right? And so, coming up with a molecule that satisfies all of those criteria, right, that it has the right bioavailability, i.e. that it gets into the body, that it has the right distribution metabolism and pharmacokinetic characteristics, that has the right toxicity characteristics, that's all another extremely critical portion of

• 28:12Speaker 00

  Of the process of discovering, developing drugs. And that's where there have been really exciting and revolutionary companies like Chai Discovery, Isomorphic, which is spun out of Google, Bolts, Lambda Labs, Profluence Bio, like a lot of great companies there that are working on that problem of coming up with the actual molecule. And so the upshot of this is just that I think that you're going to, the biotechs, the pharma companies of the future will be much more lean, right? Which is to say, you're going to be able to pursue many more drug programs in parallel

• 28:55Speaker 00

  than you can today with the same number of people. And I think it is like time for us to start thinking, like, given that that is the case, this is what I mean when I say we're going to like remove the talent bottleneck, right? Given that that is the case, we're going to be able to just like pursue cures for so many more diseases. And what we should really be thinking about is something like, you know, what would the human genome project look like for medicine?

• 29:17Lukas Biewald?

  But I feel like earlier you said the bottleneck was the human clinical trials.

• 29:21Speaker 00

  Absolutely.

• 29:22Lukas Biewald?

  And what you just said makes it sound more like it's kind of coming up with the candidates for the clinical trials. It's really the bottleneck. Which is it?

• 29:29Speaker 00

  No, no. So the, so, well, first of all, I'll just say that there are many bottlenecks.

• 29:36Lukas Biewald?

  Sure.

• 29:37Speaker 00

  Right? That all need to be overcome. But I think that fundamentally, at the end of the day, there's like an operational bottleneck, which is that you need humans in order to run the clinical trials, right? You need humans to, if we had twice as much money and twice as many kind of operators in drug discovery, then in most areas, this is not always true, because sometimes you're limited by patients and so on. But in most areas, you would just be able to pursue twice as many ideas, right? So we're actually limited, and like I said before, capital scales, right? If we had more good ideas that were investable, then there would be more money that would be available for them.

• 30:21Lukas Biewald?

  Do you, do you think that the new upstarts like that you mentioned, like Isomorphic and Chai and others, have a structural advantage over the incumbents in this process?

• 30:31Speaker 00

  I mean, yes, in that, well, so mostly what those companies are doing is that they are coming up with molecules that they then partner with the pharma companies on developing. So they are mostly not, at least to my knowledge, mostly not today, developing their own drugs, right, and are instead trying to get this technology everywhere. And I think that the pharma companies don't have the expertise in-house to do it themselves. And so that's why they're going out and they're partnering.

• 31:02Lukas Biewald?

  I mean, they do have teams that supposedly work on drug discovery, right?

• 31:05Speaker 00

  Oh, they, I mean, they, the pharma companies have strong, have good AI teams, right? Actually, we have been very impressed with the quality of the AI people who are inside of the, who are inside of the pharma, who we find inside of many pharma companies. But, you know, when it comes to figuring out, when it comes to training the best model in the world for generating, you know, a new antibody, that's just like not the pharma company's game, right? In the way that it is Chai's game and it is Leighton Labs' game and so on.

• 31:45Lukas Biewald?

  Got it. Okay, so Sam, tell me about your peptide stack.

• 31:48Speaker 00

  I had to think about what you meant for a minute because we're actually, with one of our partners, we're working on training a model that is better at reasoning about peptides. And so I was like, wait, how does he know about that? My peptide stack. Okay, so I'm gonna admit, I'm like not a biohacker.

• 32:08Lukas Biewald?

  Whoa, no peptides?

• 32:10Speaker 00

  No, no peptides.

• 32:11Lukas Biewald?

  Wow.

• 32:12Speaker 00

  Yeah, I know. And the reason is I know too much about biology. And I think that you'll find this, that like the really...

• 32:21Speaker 00

  Okay, I won't put this person on the record. So I was, you know, I was hanging out the other day with the head of research and development at one, at a very large, a top 20 pharma company. And, you know, he was talking about the kind of like peptide fad and he was just like, he was just like, yeah, these people, they like, don't understand what can go wrong.

• 32:43Lukas Biewald?

  Oh, I see.

• 32:44Speaker 00

  Which like, and I think that's, that's true. I mean, that we, when you, when you are like, study, when you have studied biology and drug development, you just get an appreciation for everything that can possibly go wrong, that including things that are like very, very hard to identify if you're not looking carefully.

• 33:07Lukas Biewald?

  Interesting.

• 33:08Speaker 00

  And you also gain a deep appreciation for things like, like placebo, like the strength of placebos, which just, those two things together just make you very intrinsically skeptical.

• 33:20Lukas Biewald?

  So are you nervous even about something like Ozempic that tons of people take?

• 33:23Speaker 00

  I mean, Ozempic has been used to treat diabetics for a long time, right? Or, or at least GLP-1 agonists have been. And so I'm not that concerned about it. And even, so it's been through, the GLP-1 agonists have now been through many, many robust, well-controlled trials. And so I don't worry about that. A lot of these peptides where people just go in and they find some hypothesis and they're like, great, like, let's just take this peptide and inject it into ourselves. As I said before, like, you have no idea about if that peptide is even like staying in your body for long enough to do anything, right? Let alone if it is like actually doing what you think it's doing, if it's having the effect.

• 34:06Lukas Biewald?

  Well, these are like separate concerns, right? Like, does it not do anything or does it hurt me?

• 34:10Speaker 00

  Yeah, those are different concerns, right? But like, the, the issue, I mean, I'm also just a scientist, right? And I think the issue is that when you think about, like,

• 34:24Speaker 00

  Trying peptides on yourself that have not never been studied in like a really robust, in a robust trial setting, you know, you don't know, you actually have no way to know whether they're working, really. And people, they can make people feel better. That's great. I'm, you know, placebos also make you feel better. And so that doesn't teach you, doesn't owe you anything. That makes you feel better, right? Even if you say, oh, look at everyone who takes it feels better.

• 34:50Lukas Biewald?

  Right.

• 34:51Speaker 00

  Yeah, that's the point. Yeah, I know. That's right. And if like this group of people were systematically, you know, even having like a very large, the group of people taking some of these peptides, even systematically had like, you know, a 10% incidence of like cardiac arrest or something, they wouldn't know. No one is, no one is looking, no one is tracking those, those, those adverse events when people just go and dose themselves with peptides, right? And so I think that's, now, I don't want to sound too much like a, you know, so my recommendation would always be like, probably don't, you know, I feel like

• 35:32Speaker 00

  I have a professional and ethical and moral obligation to say to people like, generally doing this is ill-advised, right?

• 35:41Lukas Biewald?

  But this is totally off the record, so you can say anything you want on this podcast.

• 35:45Speaker 00

  Yeah, exactly. Completely off the record. I mean, but then again, I will say like, I do, I do admire people who like, I have a lot of admiration for the biohacker movement. I have a lot of admiration for people who go out and like, kind of, and want to just try things. Because I'm a fan of just trying things, right? We just need to make sure that like, I think you just need to be clear-eyed about what the risks are. Because there are no guarantees with a lot of these things. And it can do damage, I mean, to the field.

• 36:17Lukas Biewald?

  Would you change anything about the clinical trial process that we have in the United States?

• 36:22Speaker 00

  Yes. So, the flip side of what I just said about peptides, right, is that part of the reason why people feel compelled to go and test things on themselves is because like the process for testing them in humans in a robust and rigorous and well-controlled manner is so onerous and is so, it takes so long, right? Like, you know, if it were possible to go and run like really high-quality, well-controlled studies in a way that is like, you know, well-regulated and so on, quickly, then maybe we would, you know, people would be doing that instead, right? Which would probably be better. And so there are a lot of things, there are a huge number of things that we should be

• 37:07Speaker 00

  doing. The first one, which is just the most obvious, boneheaded thing, like absolutely we should be doing, it's crazy that we're not doing this, is that in Australia and in China, for early-stage studies, the process of getting a study approved, a clinical trial approved, is decentralized, okay? Relative to the way it is in the US. So in the US, you need to, even just to do like a small-scale, initial trial, you need to get central approval from the FDA. And in China and in Australia, you, you are individual centers that run trials, that

• 37:47Speaker 00

  are capable of approving trials, which then means that those centers can compete for how, you know, how easy they can make it to do trials while maintaining within the regulatory, while staying within the regulatory guidelines, right? And so that is a drive for efficiency, and that is great. And as a result, like, US biotechs are going to Australia and are going to China to do their clinical trials. This is obvious, right? We should be fixing it. And actually, like, you know, if there's any, if there's any administration that is going to fix it, you would think that the, like, you know, the bull in a china shop kind of approach that this administration takes would be a great candidate to do it. And so, you know, I, they need to be doing that.

• 38:33Speaker 00

  And then I think there are others, so that's like just obvious. I feel like most people would probably agree on that. There are other things that we should definitely be looking at. One of the more obvious ones is loosening the requirements for efficacy. So the FDA requires you to prove two things in order to go to, in order to proceed to get your drug approved. They require you to prove that your drug is safe, and they require you to prove that your drug is effective for treating a specific condition.

• 39:12Speaker 00

  And actually, drugs usually fail the second, they usually fail on efficacy. And, and the thing about this, it's a little bit perverse, is that often even drugs will work in a specific sub-population. Like you'll work in one population of patients, but it does not work in, you know, the entire population of trial. Therefore, the trial fails, therefore it can't get approved, even though it works on some subset of patients. That feels like a failure of the system. The alternative is to only require that drugs be proven to be safe, right? And then determine that they are effective in the course of using them on patients in the field, right?

• 39:53Speaker 00

  Now, you can imagine that sometime, that this would like not be ethical, and this would not be something that you would want to sign up for in all cases. Like if there is a standard of care that is known to be effective, and then there's a drug where it's like, we know it's safe, but it's not necessarily effective. We don't know if it's effective. You might choose that you want the, the effective one. On the flip side, if the effective one doesn't work for you, right, and you have no other options, you might choose to go with this other one. And that, I think, would probably reduce the barrier to, to doing clinical trials in, in many diseases, right? Not always, but-

• 40:22Lukas Biewald?

  It's so hard to know what's effective, right? Like, I feel like I have lots of smart friends that take lots of, you know, medicine or things like that, that they think is effective. And in my mind, I'm thinking probably not because the science doesn't show that it's effective. I'm actually not even sure like who's more likely to be right.

• 40:39Speaker 00

  But what you can do, absolutely. But so the way that we deal with this is with randomized control trials in when you're in the standard process. Right. So you have patients and half of them get the actual medicine, half of them get either don't get the medicine or more often just get the standard of care because it's usually you would consider unethical to deny a patient a medication. Depending, again, depends on the condition, depends on the disease. But the but the other way that you could do it, you know, if we were to relax the efficacy requirement so that, you know, you only have to show a drug is safe, then you could look at

• 41:21Speaker 00

  the extent to which, you know, a single patient improves, like you could look at within patient improvement. So if you take drug A, if you have depression and obviously famously antidepressants are very variable in whether they work in a given individual. And so maybe you take antidepressant number one and it doesn't work for you and your symptoms persist. And then you take antidepressant number two and suddenly you get much better. That's like, you know, that's pretty good.

• 41:54Speaker 00

  As far as evidence goes, because like the only variable there that has been changed is which medicine you're taking. Right. That avoids the placebo problem to some extent, right? Because the issue with the placebo is you go from not taking medicine to taking medicine. And simply taking a medicine is effective often regardless of whether medicine works. The other way that you can think about doing it is like patients who are on drug A, which is known to be effective versus versus patients who are on drug B, which is not known to be effective. If drug B is more effective, then like if the patients on drug B have a better outcome than the patients on drug A, that's very strong evidence in favor, right?

• 42:31Lukas Biewald?

  Although I'd imagine this all sounds good in theory, but then I would think it'd be very vulnerable to pee hacking. If you, you know, if you're just sort of doing it, these natural experience experiments in the wild and tracking everybody.

• 42:43Speaker 00

  Yeah.

• 42:43Lukas Biewald?

  Sort of looking for effects like this.

• 42:45Speaker 00

  Peehacking is not.

• 42:48Lukas Biewald?

  Especially if you can.

• 42:49Speaker 00

  Usually something you can control for.

• 42:50Lukas Biewald?

  After the fact, pick your subpopulation.

• 42:52Speaker 00

  So that's a problem. So picking the subpopulations post hoc can be a problem. Although, again, this is, it's, it's really, it just becomes a question of power. So peehacking is just always a question of power.

• 43:07Lukas Biewald?

  Except that you don't know all the possibilities that were considered, right?

• 43:11Speaker 00

  Well, you need to kind of control for that.

• 43:14Lukas Biewald?

  Or pre-register what you're gonna look at or something.

• 43:17Speaker 00

  Imagine, let's just take this example of, uh, let's take the example of vaccine. Because vaccines, uh, real world evidence for vaccines is very unambiguous. Because you either get the disease or you don't get the disease. Okay? And so if I go out and I have vaccine, I don't know if it's effective. And I just give it to, you know, a bunch of people. And then afterwards, and let's say at a population level, it's not effective. Right? Which is to say, people without the vaccine get the disease at a rate that's indistinguishable from people who get the vaccine. Then I'm gonna go in and I'm gonna look, look at like, oh, but what about like, you know, just men, women, like, or, you know, men under the, above the age of 18, but below the age of 35, or right, like, I mean, go and look at all the different subpopulations.

• 44:02Speaker 00

  And inevitably, I will find one where it's like, no one in that population got the disease and therefore, um, I'm gonna say, hey, it's 100% effective. And you're gonna be like, no, it's just peehacking. And you'd be correct. You know, if I have like a thousand patients, maybe. Let's imagine that I have 10 million patients. And across 10 million patients, I still find that like, somehow, um, men between the age of 18 and 35 who get my vaccine, never get the disease. Whereas all other populations get the disease at the same rate. No, that's like, not, right, like, you know, that's not peehacking anymore. That's obviously not peehacking. Like, something is going on, we don't know what it is. But that's like, obviously not peehacking. And so, that's why I just mean, it's a question of power.

• 44:43Speaker 00

  That if you get enough real world data, you can robustly go back and do, and, and, uh, kind of identify these, uh, subpopulations in which a drug is effective without worrying about peehacking. Now, you may still want to go and do more experiments because you may have no idea why. Right? And you may want to go and do a confirmatory experiment, right? But you know, you can imagine that it's better to be able to go back in and at least find those hypotheses than it is to just like have the trial fail and the

• 45:17Lukas Biewald?

  I mean, how often do we have real world data at that scale where there's not sort of weird selection bias?

• 45:22Speaker 00

  Yeah, so vaccines are often given at population scale. But you know, you could imagine, actually antidepressants are a great example. You could imagine getting enough real world data on antidepressants in order to be able to do this. You know, the more common cancers, I mean, cancer is a specific thing because the treatment is obviously life or death. And so the treatment is, you know, you need to be more careful. You never want to deny a cancer patient the standard of care, right? But you know, the common cancers, breast cancer, colon cancer, and so on. You'll have, you know, hundreds of thousands, millions of patients.

• 46:02Speaker 00

  And so I think that, yeah, obviously that for a rare disease, you know, you won't get a huge amount of real world data. But also for rare disease, you don't have that many subpopulations usually, right?

• 46:18Lukas Biewald?

  Do you think that a model like yours could go through the existing real world data and find new patterns you haven't seen before?

• 46:24Speaker 00

  Absolutely. And we do that. We are, we have several clients that we're working with that aim at those, that kind of work. The challenge there is always the quality of the data, which is to say that real world data is gathered in the real world. And real world people, as probably many people watching this understand, do not always go to their follow up appointments. Do not always take the medication on the schedule that they're supposed to take it on. Do not always, you know, doctors do not always take high-quality notes. They, you know, patients move between healthcare systems and then you lose, the records end up

• 47:06Speaker 00

  disconnected. And so for all of these reasons, like, the real-world evidence is never, existing real-world evidence is like never as good as you want it to be. And there are a number of great companies that are trying to fix this. One of them is a company called Empower Medicine, which is focused on gathering just extremely high-quality, kind of, electronic healthcare record data from patients so that they, you know, you can kind of design synthetic clinical trials. There are other companies like Tempest, Komodo, and so on that do this as well.

• 47:42Lukas Biewald?

  I mean, I guess in the single treatment case, it makes sense there'd be a lot of public research and someone would want to decide once and for all, is this effective or not?

• 47:53Speaker 00

  Yeah.

• 47:54Lukas Biewald?

  I've found in my life, when I have medical issues, they feel high stakes. And when I go look at the research and the data, it never seems clear. Often seems like there's conflicting research and every single dataset seems biased in different ways. And I find myself more and more using LMS to try to synthesize the research into making like a sensible decision. For example, fertility was a big issue for me and my family. And, you know, it's expensive and there's upsides and downsides. And I really wanted a model like yours. And I think I asked you to use your model for recent, you know, thing that I was, I was looking at. Do you expect that people might use your model in that way to

• 48:36Lukas Biewald?

  sort of look at their individual specific situation and try to go out and dispassionately synthesize the research into a decision?

• 48:44Speaker 00

  I mean, my mother does.

• 48:45Lukas Biewald?

  No way. Tell me about that. That's cool.

• 48:47Speaker 00

  Well, I don't know that my mother wants me to air her, her medical history on a podcast.

• 48:53Lukas Biewald?

  This is totally off the record, man.

• 48:54Speaker 00

  I know. I forgot. Yeah, yeah, yeah, exactly. No, the, look.

• 49:02Speaker 00

  I have a bunch of friends who, who, you know, who use it. I have one friend who had a breast cancer scare, for example, who wanted to figure out, you know, if they delayed treatment, if she delayed treatment. What, you know, she had a reason why she wanted to delay treatment for a month or two. And if she delayed treatment, what would that do to the effect to, you know, the, her, her, her recovery or survival odds? And luckily she turned out to be fine, which is good. But one of the things that, you know, our platform is very good at is going out and dispassionately surveying the evidence and giving you

• 49:44Speaker 00

  an answer that is really like directly grounded in, in like hard evidence from papers, from the literature. As opposed to vibes.

• 49:57Lukas Biewald?

  All right, actually, you're a recent father. Have you used these models or your own model for your child yet?

• 50:02Speaker 00

  Yeah, so we have been lucky enough so far that we've not needed to use them for medical reasons. But, you know, 100%, I ask it things like, you know, when does, when will my baby start to walk and crawl and so on? Yeah, yeah, yeah.

• 50:17Lukas Biewald?

  I'm sure you will. I actually, I will tell you, I had a recent interesting experience with my wife where we both put in the same symptoms and incident with our daughter. Actually, she hit her head. We both described it identically. And the model told me it's fine and told my wife to take the child to the, to the doctor, which I think might have been a different kind of incentive in the model. Sort of, I do feel like they kind of want to tell you, you know, what you want to hear. And maybe sort of some RLHF kind of gives it that. So I wonder if it sort of implicitly knew.

• 50:46Speaker 00

  Or like, I mean, it has the history. It probably knows who each of you are.

• 50:49Lukas Biewald?

  We know who we are. It definitely knows.

• 50:50Speaker 00

  Yeah, I mean, you know, like I said before, we are, what we focus on is making sure that our answers are grounded in like fact and not scientific. Right, but it does mean, it can be frustrating. So, you know, you will go, you can ask ChatGPT or you can ask Claude, you know, about if you should take your child to the, to the doctor. And they will say, no, child's fine. Fine. If you ask Cosmos, our agent, if you should take your child to the doctor, Cosmos will probably come back and say, you know, children who like hit their heads have a, you know, X percentage chance of developing like, you know, this serious problem and Y percent chance of doing this. And then it will give you something like, you know,

• 51:30Speaker 00

  on average, like there's like a whatever 97% chance that if you don't take your child to the doctor, they're going to be fine. And that can be more frustrating sometimes.

• 51:39Lukas Biewald?

  Honestly, that sounds fantastic. I, I, yeah.

• 51:42Speaker 00

  I, I, yeah, but it's a question of actually like how much information you want. And it's very interesting to think about. We're gonna, we're entering this era where information and intelligence are just like so much more abundant than they were before. But like, it's not always good that we're gonna want that.

• 52:02Lukas Biewald?

  Mm hmm.

• 52:03Speaker 00

  Right? Like more information is not always better for making decisions.

• 52:07Lukas Biewald?

  Totally true. Although I don't like to admit that.

• 52:09Speaker 00

  I know. But, but often it's like you, you need the right information, but not the most, right?

• 52:19Lukas Biewald?

  So OpenAI and Anthropic and DeepMind obviously are working on similar things to you, at least in the sense of like deep research, making the models more advanced, making agents that, that think more. Do you feel like you need to keep some kind of structural advantage over what they're doing?

• 52:38Speaker 00

  Um, I mean, uh, yes. I mean, if you're, if you're building a company in general, you want to try to have a structural advantage over what the, or what, you know, other people are doing.

• 52:50Lukas Biewald?

  Okay. So what, what is your structural advantage?

• 52:51Speaker 00

  Great question. So the, um, uh, there are a couple of ways to think about this. Um, the, the, the most-

• 52:59Lukas Biewald?

  By the way, were you yawning through my question?

• 53:01Speaker 00

  Yeah.

• 53:01Lukas Biewald?

  Is that a bad question?

• 53:02Speaker 00

  Yeah, it's boring.

• 53:03Lukas Biewald?

  Sorry, man.

• 53:04Speaker 00

  I just, you know, and my whole life gets, my whole life, I get asked always like, you know, OpenAI, Anthropic are doing this. I was like, oh.

• 53:11Lukas Biewald?

  I'm giving you a softball, man. Let's go.

• 53:15Speaker 00

  No, so.

• 53:19Speaker 00

  So.

• 53:20Lukas Biewald?

  Man, now I'm worried. Is this, maybe we should move on.

• 53:22Speaker 00

  No, no, no, no, no, no, no. It's a great question.

• 53:24Lukas Biewald?

  All right.

• 53:24Speaker 00

  I'm just teasing you. It's a great question. The.

• 53:30Speaker 00

  So.

• 53:32Speaker 00

  The way that, you know, fundamentally, at the end of the day, OpenAI, Anthropic, DeepMind are focused on building, you know, AGI, or artificial superintelligence or whatever, right? Now.

• 53:47Speaker 00

  But I think the key important, the key thing to understand is that having a specialized model for a particular task is, you should expect, is like always going to be better than having a generalist model for that task, even when you have like superintelligence or whatever. And the reason is just

• 54:13Speaker 00

  that the, the, the generalist model has to do everything in the weights and the specialist model only has to do something.

• 54:21Lukas Biewald?

  Wait, so do you have a specialist model?

• 54:23Speaker 00

  Absolutely. We train, we train reasoning models on specific scientific tasks.

• 54:26Lukas Biewald?

  I see.

• 54:27Speaker 00

  And what we see is that with a very small amount of data, you can get enormous gains for those specific tasks over the, over the, over the frontier models. And so that's the first thing, right? The first thing is, is, you know, just having specialized agents, specialized models, which also, by the way, comes with like specialized user experiences, like I'm mentioning, right? We put way more test time compute into minimizing hoardings than, than the, than, than the others would.

• 54:54Lukas Biewald?

  But presumably you also use the, the bigger models, right?

• 54:57Speaker 00

  Yeah, yeah, yeah, absolutely. We use the bigger models in, in some areas, right? Like, we don't want to be better than Anthropic at coding, for example, right? But when it comes to like very niche biological reasoning, we have our own models internally that, that are superior.

• 55:10Lukas Biewald?

  Doesn't the coding thing actually show an example of a specialized? I mean, these coding models are not really specialized in coding, right? And yet.

• 55:19Speaker 00

  So they definitely like.

• 55:24Speaker 00

  If your point, if your question is like, is it also true that if you had a specialized coding model, could the specialized coding model beat like, you know, G P five point five or beat?

• 55:36Lukas Biewald?

  Seem unlikely.

• 55:37Speaker 00

  You think it seems unlikely?

• 55:39Lukas Biewald?

  Well, I think a lot of people are trying to do it and they haven't.

• 55:41Speaker 00

  I think they haven't yet. I kind of, I mean, coding is very special because I think all the labs have realized at this point that like coding is the thing that they need to be really, really good at, right? And so maybe it's not true in coding for that reason, but I can kind of guarantee you that like, if you want a model that is, if you want to like, train a specialized model for, you know, for reasoning about like synthetic chemistry, synthetic pathways, I think if the specialized model is going to be better than the generalist model, right? Now, the place where this may, may fall down is when you get to like intelligent saturation, where so like, you know,

• 56:21Speaker 00

  the thing I tell people about saturation is that if you think about tic-tac-toe, right? Any super intelligence will be exactly as good as humans at playing tic-tac-toe, right? No amount of intelligence above humans will improve performance at tic-tac-toe because humans are optimal at tic-tac-toe, right? You can like always.

• 56:39Lukas Biewald?

  Some humans.

• 56:40Speaker 00

  Some humans, not all humans, but, but there exists humans who are optimal at tic-tac-toe, right? And, and so similarly, you should imagine that at some point, the intelligence will get strong enough that like, we will have saturated, like that more intelligence will not improve your ability to reason about chemistry. I don't know where that point is, but it's feasible that we would get there. And then at that point, my argument might fall down, right? Except, you know, insofar as, yeah, no, at that point, the argument might fall down. But so that's like the first thing, the second thing, but I think we're very far away from that point, right? And.

• 57:25Speaker 00

  This is why, like, the second thing that you think about is that all of the pharma companies have their own internal data sets, right? And they need to have proprietary advantages over each other, right? And so they need models that are trained on their data. Because if everyone has the same intelligence, right? Then no one has any advantage in R and D, right? And so those are the areas where we really like, where we really shine. And like, the reason right now why all pharma companies, or it feels like all pharma companies are like, are fighting to do a deal with us, and we're just kind of like overwhelmed with demand, is because we have the best models for science.

• 58:08Speaker 00

  We do the kind of last mile integration to get those models deployed internally and actually really accelerate the pipeline, and we train on their data, which provides them with their sustainable advantage, which is something that at least today, you know, OpenAI and Anthropic don't want to do, because they don't want to have a different model for every customer, right?

• 58:25Lukas Biewald?

  It's kind of interesting. I feel like when I talked to you six months or a year ago, you talked more about the sort of agent framework and the evals that you were doing. Has things changed, or is that less exciting to talk about?

• 58:37Speaker 00

  No, I, like, I think that we have matured, or we are in the process of like, you know, our kind of, the way that we think about what we do and about the value proposition has matured, right? But it remains the case that, you know, it remains the case that our agents are, I think, the best at science, and are substantially further ahead of the offerings that, you know, Anthropic has, and that OpenAI has, and that GDM has, and so on. And I think that we'll be able to maintain that edge for a while, right? But I think that, like, when you think about the macro dynamics of like, what will the

• 59:21Speaker 00

  market look like in, you know, three years, or four years, or five years, there are a couple key things happening. The first one is that these companies need models that are trained on their data to maintain their advantage. The second one is that they don't want to be locked into a single model provider, so they don't want to be locked into just Anthropic, or just OpenAI, or whatever. Because last year, OpenAI was, you know, way ahead, and today, you know, Anthropic is, feels like they're way ahead, and whatever, right? And so those dynamics are kind of pushing, push customers to work with us.

• 59:54Lukas Biewald?

  Do you think when you look at the the major model providers, they're different enough that using them for specialized use cases is a valuable thing to do? Or are they essentially interchangeable?

• 1:00:02Speaker 00

  This is a great question. I think that at the moment, they are mostly

• 1:00:09Speaker 00

  they're spiky, right? So I think at the moment, actually.

• 1:00:13Lukas Biewald?

  But are the spikes overlapping or are they?

• 1:00:15Speaker 00

  I think there is a lot of non-overlapping spiky.

• 1:00:19Lukas Biewald?

  Oh, interesting.

• 1:00:20Speaker 00

  So.

• 1:00:21Lukas Biewald?

  Give me one example.

• 1:00:22Speaker 00

  Yeah, great. Like, for example, we have a benchmark that we're building that we will release shortly that involves reproducing analyses that have been done in scientific papers previously. And

• 1:00:37Speaker 00

  at least on our evaluations, as of when we're recording this, it may it may change, right? But at least as of the last day that I saw, Anthropic is particularly good at that, right? More, you know, significantly more so than OpenAI or significantly more so than than DeepMind. More, another good example, which is actually maybe even makes the point to a greater extent, when there was a time early in the development of Cosmos, where there was a specific task that we needed the models to do in the process of kind of like updating Cosmos' world model. And at the time, the only model that we could find that was able to do that task was Gemini 2.5.

• 1:01:20Speaker 00

  And like, none of the other models, we could not figure out how to get any of the other models to do that task well, which was wild. And so there's a lot of that spikiness, right? They have different, slightly different personalities, obviously, so that's also interesting. But yeah, we've been surprised at the extent to which you can get real, you can actually get much better results on tasks by

• 1:01:50Speaker 00

  by like combining models across different providers.

• 1:01:52Lukas Biewald?

  Oh, see. When I fire off like Cosmos today, how long does it run for? Like, how many calls are they making?

• 1:02:01Speaker 00

  Yeah, great question. So we released a version, we released like the original version of Cosmos back in November. And that would run for like six to 12 hours, would write 45,000 lines of code, would read 1500 papers in a single run, and was just way more powerful than any agent that anyone had seen.

• 1:02:20Speaker 00

  And is still today, I think, like one of the most powerful, most intensive agents out there. The issue with it was a UX issue, which is that like, actually, if you are a scientist and you're doing some research, you like, don't really want to just like, ask a question, then walk away and come back six to 12 hours later. And so we have just recently announced a significant update to Cosmos that now puts a kind of interactive front end on it, so that it can still go and do those extremely long-running tasks, but is more, but you can give it feedback throughout, you can steer it, it's more interactive, right?

• 1:02:57Lukas Biewald?

  Let's go to rapid-fire, random questions here.

• 1:02:59Speaker 00

  Awesome, yeah, let's do it.

• 1:03:00Lukas Biewald?

  All right, you said, I saw that you said, "Scientific progress has slowed down a lot." That was different than how I think about scientific progress. What did you mean by that?

• 1:03:10Speaker 00

  Well, that depends on which time I said this that you're talking about.

• 1:03:14Lukas Biewald?

  Do you still stand by that point that I pulled out of context?

• 1:03:17Speaker 00

  The, okay, it is empirically the case that progress in medicine has slowed down.

• 1:03:21Lukas Biewald?

  What does it mean empirically?

• 1:03:23Speaker 00

  This means that we have in medicine what is called Eves' Law, which is more spelled backwards, which is the observation that the amount of money in real terms that it costs to develop a new drug has, you know, doubled over the past, whatever, 40 years. And I think, I forget what the double time is, it might be it's doubled every nine years or something like that, right? So like, you know, prices on something, conductors have come, have had this exponential drop, prices on drugs have had this exponential increase. And there are various reasons why people have argued that this might be, it actually broke, that trend broke in 20, around 2011 or 2012.

• 1:04:04Speaker 00

  Think it's largely because of the Human Genome Project.

• 1:04:06Lukas Biewald?

  Oh, actually.

• 1:04:06Speaker 00

  Actually, that like the Human Genome Project make it made it easier to find drugs. But now it's about flat and like certainly productivity is not is not yet increasing. And so.

• 1:04:17Lukas Biewald?

  Is that because we found all the best drugs or?

• 1:04:19Speaker 00

  That's one, uh, the one of the reasons. So one of the reasons may be a lack of low hanging fruit. One of the reasons may be what is called like the better than the Beatles. Problem, which is that if you have a drug that treats condition X. If you have a drug that treats depression, uh, maybe bad example, uh, whatever. You have a drug that treats, uh, colon cancer. You need to come out with a drug that is better than that drug.

• 1:04:43Lukas Biewald?

  Okay.

• 1:04:43Speaker 00

  In order to, um, uh, uh, be in order to gain market share, right? In order for it to get approved, even even just in order for it to get approved.

• 1:04:52Lukas Biewald?

  Well, also for it to be useful.

• 1:04:54Speaker 00

  For it to be useful, exactly.

• 1:04:55Lukas Biewald?

  Why are you all the better than the Beatles?

• 1:04:56Speaker 00

  Oh, just because like there's this observation that like, you know, um, I guess the argument is the Beatles are still extremely popular.

• 1:05:04Lukas Biewald?

  I see.

• 1:05:04Speaker 00

  Right? Like the Beatles have not been displaced as a band, right? They were like first any band that comes afterwards, like if the Beatles had not existed, maybe there would be some other band. And like, there are still good bands, but they like, don't get as much, uh, like, you know, uh, uh, the kind of like early rock bands. Like most of them don't get as much.

• 1:05:23Lukas Biewald?

  I see.

• 1:05:23Speaker 00

  Share as the Beatles, um, because they kind of like took that space in the.

• 1:05:27Lukas Biewald?

  I see.

• 1:05:28Speaker 00

  Consciousness. I don't know if that's.

• 1:05:29Lukas Biewald?

  So basically like all diseases have been have been cured. So.

• 1:05:32Speaker 00

  No, but this is the thing. They haven't been cured.

• 1:05:33Lukas Biewald?

  Right.

• 1:05:34Speaker 00

  It's just that we can't find better treatments for them.

• 1:05:36Lukas Biewald?

  Interesting.

• 1:05:36Speaker 00

  Like no one is going to go out and argue that depression has been cured or that colon cancer is cured. Right? Um, but you know, if you want your drug to get approved, you have to be better than the, the, the best in class.

• 1:05:48Lukas Biewald?

  But yeah, I think that's some kind of, I mean, that's a, that's a problem. You of course need to be better than the best in class, right?

• 1:05:53Speaker 00

  For. But that just means that it's just getting harder, right? It's getting harder to come up with. And because you don't, you don't build on, in semiconductors, right? You were able to build on the previous innovations, right? In a way that is less the case in medicine, right? You don't.

• 1:06:13Lukas Biewald?

  Why not?

• 1:06:14Speaker 00

  Well, because if you have a drug that uses a particular mechanism in order to cure cancer or to fight cancer, you need to come up with a different mechanism. Like, there's only so much that you can get from juicing that specific mechanism, right? You can optimize the drug, you can make it marginally better and so on. And there's a lot of gains that you get out of that. But then, like, you know, it's often the first in class drugs that are, that like lead to the dramatic improvements.

• 1:06:37Lukas Biewald?

  I see.

• 1:06:37Speaker 00

  Right?

• 1:06:38Lukas Biewald?

  Interesting.

• 1:06:39Speaker 00

  So where it's like, there's some new mechanism that we're using.

• 1:06:44Speaker 00

  Yeah, and so I think that's a, science is moving more slowly for that, for that reason. I also just think, as in the case of physics, right? Like, as you discover more things, it just becomes harder to discover more things, right?

• 1:06:58Lukas Biewald?

  Interesting.

• 1:07:00Lukas Biewald?

  All right, is a PhD or formal science degree still worth pursuing?

• 1:07:05Speaker 00

  Right, wow, yeah, good question. The, I don't know. If I were starting like.

• 1:07:12Lukas Biewald?

  The answer, red.

• 1:07:14Speaker 00

  Sorry, you can yawn, your turn to yawn.

• 1:07:16Lukas Biewald?

  Oh, my God, equivocating, I hate it.

• 1:07:19Speaker 00

  The, I think that the answer is yes, because I don't think that, like, human researchers are going anywhere anytime soon. And fundamentally, like, the, the point of a PhD is to learn how to do research, right? And if you never learn how to do research, you definitely, like, are not going to be effective at doing research using the.

• 1:07:37Lukas Biewald?

  But aren't you automating research?

• 1:07:39Speaker 00

  We are, or like, certainly accelerating it. The question is, the more interesting question is, like, are we going to need human scientists?

• 1:07:47Lukas Biewald?

  Yeah, it's a related question, for sure. What do you think?

• 1:07:49Speaker 00

  And I think the answer is yes, for a substantial amount of time. And the reason is that science is non-verifiable. And so fundamentally, at the end of the day, we need human scientists for their taste. Or there's like, I don't, I'm not sure that it's going to be. Um, like maybe in 20 years in on the five, 10 year time scale. I'm not sure that we're gonna get to a point. It's not clear to me yet whether we will get to a point where the models will just be like obviously better taste than the humans.

• 1:08:12Lukas Biewald?

  Hmm. Well, it's interesting because here it says, uh, you said language models will eventually be better than humans at coming up with ideas.

• 1:08:19Speaker 00

  Uh, yeah, but, but like.

• 1:08:20Lukas Biewald?

  You think that's far out there?

• 1:08:22Speaker 00

  I mean, eventually and, and better exists along many different axes, right? So, um, better can mean, uh, um, uh, you know, better can be mean more likely to work. I mean, I think that's definitely the case, right? That the language models are way better than humans at coming up with ideas that are better, that are more likely to work, right? But when it comes to like specific, like which of these problems are the most interesting to pursue, um, it's just harder because it's not verifiable. And so, uh, like this is just uncertainty, right? Like, you know, like I said, I'm a scientist. I admit when I don't know things. I don't know whether or not we're going to get to a point within five years where we're just like, oh, you know, there's no point in asking, you know, uh, Frances Arnold Nobel

• 1:09:07Speaker 00

  laureate, right? Um, uh, what, what she thinks about evolving proteins is we could just ask the model. I mean, you know, I, I, I think it's a pretty tall bar. And when I look today at like, I'm going to go and ask Claude 3.7 what it thinks we should be doing in, you know, um, using direct evolution to improve chemistry or like, you know, to open up new avenues in chemistry versus asking Frances. There's no question, right? That Francis is gonna have better ideas.

• 1:09:38Lukas Biewald?

  Interesting.

• 1:09:39Speaker 00

  Today, today, right? I mean, you know, in two years, we're gonna be sitting here again. You're gonna be like, well, you have said that. And now, you know.

• 1:09:47Lukas Biewald?

  In this off the record podcast.

• 1:09:48Speaker 00

  Yeah.

• 1:09:51Lukas Biewald?

  All right, so we're just gonna work going forward to tell guests it's off the record. All right.

• 1:09:57Speaker 00

  Chat the House rules, people. Chat the House rules.

• 1:10:01Lukas Biewald?

  I was literally just at an event where I was on a panel and they're like, it's off the record, you know? And I was like, looking at it, I was like, what about all the cameras that are pointing at me? It's just like, oh my God, nothing's off the record. All right, so you raised quite a lot of money. I think I have 70 million in my notes. Is that even accurate?

• 1:10:17Speaker 00

  Yeah, that's right.

• 1:10:18Lukas Biewald?

  70 million, but I think-

• 1:10:19Speaker 00

  And that was just the for-profit, the nonprofit raised more, but-

• 1:10:21Lukas Biewald?

  And, you know, I think unlike most of the labs, the Neo Labs and the kind of companies that I come across, most of your investors are actually folks I don't know. They're kind of coming from more the pharma world.

• 1:10:32Speaker 00

  Oh, yeah.

• 1:10:33Lukas Biewald?

  I think. Why do you think that is?

• 1:10:35Speaker 00

  Well, we, for us, you know, success is getting inside of the pharma companies. And the pharma investors are the ones who have that, you know, who have that capability, right? So we get very high value out of all of our investors. So Yaz, Yasmin Rosafi, is an investor at Spark who co-led our round. And she is absolutely incredible. She's on the board of Anthropic. She led their first VC round and has extraordinary perspective, extraordinary insight, you know, and has really helped us with talent and so on. But

• 1:11:17Speaker 00

  like, when I think about the concrete, like, business traction, the value that, you know, the investors we have who have contributed the most value, Jeff Huber, who was the CEO of Grail and somehow seems to know literally every single person in pharma. And then the we have a unknown, as they like to call themselves, an unknown, or what do they, what does it say? An unnamed institutional biotech investor, very large unnamed institutional biotech investor, which has, would be displeased if I said in this forum, in this off the record forum, who they are. But I think most people who are in biotech will know what

• 1:12:01Speaker 00

  that means because they're very well-known. And they have been extremely, they also co-led our round. They have been extremely valuable in terms of setting us up with, you know, high-level connections into companies and getting us deployed.

• 1:12:16Lukas Biewald?

  It's interesting. So the people actually know your field better are more bullish on you than the maniacs doing AI investment in Silicon Valley.

• 1:12:24Speaker 00

  Yeah, I mean, I think the maniacs, I mean, the maniacs are very excited about many things.

• 1:12:32Lukas Biewald?

  Right.

• 1:12:32Speaker 00

  Are very excited about the concept of, you know, we're going to automate science. I mean, everyone is very excited about, let's go automate science, let's automate drug discovery. Right. But the insiders know the problems. And so the insiders really viscerally know where the value is. Right. Let me put it this way, I don't get any of the insiders asking me how we're differentiated versus Anthropic and OpenAI and Google.

• 1:12:56Lukas Biewald?

  I see.

• 1:12:57Speaker 00

  Right. I get asked that every single tech investor.

• 1:13:00Lukas Biewald?

  Right.

• 1:13:00Speaker 00

  Will ask me, you know, why won't Anthropic and OpenAI and Google do what we're doing? But anyone who has operated, you know, inside a pharma company, who has, you know, large positions in pharma companies and biotechs, we never get that question from them. Right.

• 1:13:17Lukas Biewald?

  Okay, there's a long history of AI overpromising in drug discovery.

• 1:13:20Speaker 00

  Oh, yeah.

• 1:13:20Lukas Biewald?

  Is this time different?

• 1:13:22Speaker 00

  Yes, this time, I promise, I promise you this time it is different. Yeah, man. Back to like 2012, people have been saying that pharma, AI is going to revolutionize pharma, and it's been one flop after another, with the notable exception of AlphaFold. Right. AlphaFold really changed the way that a lot of things are done in. Now, it changed the way a lot of things are done in one area of drug discovery and development, which is, you know, kind of the actual process of coming up with the molecule, because you have the structure of the proteins.

• 1:13:54Speaker 00

  But I think there's a ton of skepticism. That said, it is kind of crazy to think that this time will not be different. Like, we have so much evidence already that, you know, this time it's really going to be different. But I think the thing I do want to emphasize is, the proof is in the pudding, and the pudding is approved drugs. So you're going to hear a lot of stuff in the next couple of years of like, my model came up with this drug, and my model, you know,

• 1:14:22Speaker 00

  like, you know, discovered this fundamental aspect of biology and so on. And anyone who has been in pharma is not going to care until they see the outcome of a pivotal phase three clinical trial or an approval letter from the FDA. Right.

• 1:14:38Lukas Biewald?

  Awesome. It seems like a great place to end.

• 1:14:39Speaker 00

  Yeah.

• 1:14:39Lukas Biewald?

  Thanks, Sam.

• 1:14:40Speaker 00

  Cool. Thanks. This is great.

• 1:14:42Lukas Biewald?

  Thanks so much for listening to this episode of Grading Descent. Please stay tuned for future episodes.