Latent Space - 🔬 RL with Verifiable Rewards, but the Verifier is a Lab — Lila Sciences

Transcript

We have but one internet

0:00 · But not just TechBio, what do you do in terms of science?

0:03 · We are all in on the bitter lesson and scale. We think that methods that scale and that are general beat those that are not. You know, as Ilya said it NeurIPS last year, we have but one internet.

0:14 · It's the fossil fuel we fracked, we got every ounce of of data that we could out of the internet, but it's gone. And so the question AI is like, where is the next internet scale data set coming from?

0:24 · You know, people normally talk about different scaling axes. You have compute, you have data, and for science, data is not necessarily an infinite resource. And [music] your point is that we now want to add a new scaling axis for data.

0:35 · We think that like the lab of the future should feel like a data center. Rows of server racks as densely packed as possible and also as energy efficient as possible and things like that.

Intro & guest backgrounds

0:46 · Welcome to Latent Space Science. I'm Brandon, I'm here with my co-host RJ. Uh today we have Rafa Gomez Bambarelli and Andy Beam from LILA Science.

0:56 · Uh we'll just start off and will you introduce yourself?

0:58 · Yeah, thanks for having us on the podcast. Like, you know, long time listener, first time caller. Uh excited to to be here. I'm Andy, I'm the chief technology officer at LILA. I've been uh an AI researcher now for something like 20 years going back to the like pre-deep learning days, SVMs, random forest, things like that. Uh did a neural net PhD around 2010 to 2014 right as deep learning was taking off. Um it was clear neural nets were the thing to to back, but like auto grad libraries really hadn't been developed yet. So I did the back prop by hand, you know, back in my day. Walking uphill both ways uh kind of thing.

1:32 · Um got very interested in AI for healthcare and life sciences. Uh my wife's a physician, so I watched her struggle through different um uh things and thought that AI was obviously a natural solution for a lot of those problems. Did a post doc at Harvard um in the medical school doing early work on medical AI. Um and was really, you know, I'm in it for the AI. I was really interested in what problems could AI solve. But I've also always been like startup curious. Uh so I took a break from academia for a year and helped start a company called Generate Biomedicines, which was an early generative biology company. I was the founding head of machine learning there and got to do the fun kind of hybrid professor startup founder thing for the next five or six years. So I had a lab at Harvard, um again sort of between the School of Public Health and the Medical School, um doing methods research but also a lot of applied work. That was those those fun those were a great set of jobs, um but I got a sense that the like AI moment was changing in like a very significant way.

2:25 · And I wanted to be a part of it. So I started to think about where could I work at the frontier of AI and um on really really exciting problems.

2:33 · Um and you know, academia has a lot going for it. Um access to scaled compute is not one of the things that it has going for it or scaled resources. So I'd been an early advisor for for Lyra and got very excited once the thesis crystallized. Um but basically, you know, science is as an infinite token generator to train models at scale. Why would I want to work on anything other than creating a new frontier model that can solve scientific problems. Um so I kind of joke that like I hung up the tweed jacket um two years ago uh left my position uh at academia in academia and joined Lyra full-time as the inaugural CTO.

3:07 · Yeah, I I I go by Rafa.

3:10 · I'm the uh chief scientific officer for physical sciences at Lyra and a co-founder. I was a computational chemist back in the day.

3:20 · We used a commodity resource that is compute. So it was clear that we could scale up the compute to do molecular simulations and that's sort of something that produced enough data that in the early 2010s we realized we had a data problem and sort of things switched gear for me right around then. I was I worked with David Duvenaud and Ryan Adams in sort of blending what I think first like the felt like the first instances of deep learning for science.

3:48 · I was one of the first people to do generative AI for chemistry and you know with an auto encoder on a tokenized molecules. I'm so done deeply in love with latent spaces. We actually have a very similar to your guys' logo, uh but for molecules, um and that has taken its own life, that figure.

4:06 · This is the one that will be on your tombstone.

4:07 · Exactly.

4:07 · And my students have a Slack channel just to uh post it when it shows up in the wild. So, [laughter] uh not not as much of a story that's uh as Andy, but the same converting the in the 2015-2016 era.

4:22 · I spun out a company out of my postdoc at Harvard, a computational materials platform company, uh and then went to MIT, uh where I started my group in material science and engineering. And And the group there was sort of working at the interface of molecular simulations and AI with things like generative models for material structure, uh autograd for sort of really cool gradients that what we want to see in the molecular simulations.

4:48 · And uh by 2022-2023, sort of things were were taking sort of the turn that that Andy just mentioned, right? We had seen the bitter lesson come to a computationally generated data. That's the reason why, you know, Meta and DeepMind and Microsoft, they have teams doing AI for computational material science.

5:09 · Uh but it was clear that we needed to bridge a gap and and get this thing all the way out and make do AI for actual material science and not just the computational version. And that sort of lined up with this uh opportunity to to start spinning out something again in like I said, 2022-2023, started sort of thinking about the idea. I'm very excited now to sort of been pushing this this integrated vision of scientific reasoning across all the modalities of science we can validate in the lab.

The thesis: the bitter lesson & the infinite token generator

5:37 · All right. That brings me to what is Lyla's thesis? It seems like you have a very ambitious goal here.

5:43 · Yeah, it's a great question. Um I'll try and like give you the TLDR and then we can go a couple levels deeper. So, like Rafa said, we are all in on the bitter lesson and scale. We think that methods that scale and that are general beat those that are not. That's actually sounds straightforwardly true, but is actually counterintuitive and contrary to much of the 70-year history of AI research. But the realization that we had is that what gave rise to large language models over the last, you know, 4 5 6 years is the access of the combination of scale compute and scale data. That data came from the internet.

6:21 · It was human generated and we have used it all. You know, as Ilya said it NeurIPS last year, we have but one internet. It's the fossil fuel. We fracked. We got every ounce of data that we could out of the internet, but it's gone. And so the question AI is like, where is the next internet scale data set coming from?

6:38 · Post the pre-training era, we moved into reinforcement learning with verifiable rewards. So people talk about RL a lot. But really what RL is is a way for a model to generate its own data and the reward signal reinforces good data and penalizes bad data. So the that has been a very productive framework for problems in math and coding.

6:58 · Um but what at Lila we believe is that actually science running the scientific method and using nature and experiments as verifier is like the ultimate version of that. And so what we're building, we'll talk about these things that we call AI science factories. They are scaled verifiers for science so that we can do post-training at scale and push out the frontier of what reasoning models are capable of. Um so that's like the thesis in a nutshell.

7:24 · Your proposal is basically, you know, people normally talk about different scaling axes. You have compute, you have data, you know, you have parameters. And for science, data is not necessarily an infinite resource. And your point is that we now want to have a new scaling axis for data.

7:39 · Correct.

7:39 · So, I I to quote some of my friends at the Escalate Bio have a blog post, really good blog post, I recommend you read it. It says, "Your experiment has a run time." So, what is the run time of your data collection?

7:50 · I mean, so that is an awesome question. It's so in it obviously vary by experiment. So, like you can't make the ribosome go faster, at least to my knowledge. There is the biology sets a limit for how fast you can go. In material sciences and chemistry, there are smaller time scales, there are bigger length scales.

8:09 · Um what you're actually kind of asking is a technical question though. So, how do you train a model against feedback mechanisms that vary by orders of magnitude in terms of a feedback? So, we think about all of Lyra as being able to generate different kinds of data on different length scales. We can then synchronize how we train the model once that data has been generated. Again, for some of the experiments we do, the length scales are on the order of days or weeks. And then the question is can we multiplex, can we get more data per unit of time? But, the infinite token generator is still there.

8:43 · We just have to solve the technical problem on the other side of that to be able to line all these pieces up and train it into the model.

8:49 · So, when you say the infinite token generator is still there, what do you mean by that? Because there are many different scientific tokens you can imagine, and some tokens provide much more information than others. And certain things you can collect maybe at scale, like people who love NGS, you can basically collect an infinite amount of NGS data.

9:07 · Yeah.

9:07 · And yet, there are certain cases where, you know, another human genome is probably going to be a, you know, an incremental update versus, you know, Yeah, like my genome relative to a reference genome is like a couple kilobytes worth of information. There's not a lot of information there.

9:22 · So, that's you're exactly right. So, we don't want to generate the same kind of data over and over again. And so, the the platform that we're building is qualitatively different than traditional automation framework. So, actually the experimental platform that we're building prioritizes generalizability and flexibility over raw throughput.

9:41 · We want the model to be able to design a new experimental protocol, run the protocol, and receive the feedback even if that's not an experiment we have thought about doing ourselves. So it's the next incremental token has to be something that is valuable to the model versus yet another NGS sample to teach it something where it's already hit diminishing returns.

Inside the AI Science Factory: the "PCI bus" & the API line

10:02 · So when you say next experiment at Lyra what I think of is traditionally you would go into the lab and reconfigure the lab in whatever way and then run some experiments by hand maybe over the course of weeks or whatever.

10:19 · How does the lab get reconfigured for the new experiment at Lyra?

10:26 · The way to think about the lab is it's it's almost like a graph. And so each instrument is a node in this graph and an edge between the node indicates that there's a physical transport layer between those two instruments.

10:39 · Yeah.

10:39 · And so we'll probably have a video that we'll we'll show in a bit but we have a physical transport layer that connects almost every instrument that we have bought at Lyra to each other. These are currently planar motor systems where there's a 96-well plate that magnetically levitates over a track. You have sort of millimeter control over where that plate goes and so you can I think of it almost as like a PCI bus where each instrument I'm I'm not sure on analogies for like the It's a good one.

11:06 · I think half of the audience might not know what a PCI bus is so Yeah. So it's a universal serial bus that on your motherboard allows you to connect a new device. So if you plug a new graphics card, if you plug a new hard drive in, there's a bus that allows that device to speak to the rest of your computer.

11:25 · And this works for like bio systems and material systems and etc.

11:30 · Increasingly but um, not totally yet. So the other thing to keep up in mind about automation is there's a very long tail Yes.

11:36 · of things that you have to solve to be able to automate. And uh, to date, people have not been thinking about end-to-end automation in this like flexible kind of way. And so, um, there are instruments um, that are not connected to this now. Um, there's not a lot of high throughput automation in material sciences, for example, and we've been building custom instruments for that um, that then are brought on board. Um, but there's like an 80/20 rule I play here where things that are easy to onboard and automate are plugged directly into the PCI bus.

12:05 · Um, and then things that are not, people still move. People will still move a sample to that. Or turns out that removing a cap from a test tube is a very hard thing to automate. It's like a lot of the lab assumes that you have opposable thumbs and you're good with them. Um, [clears throat] and some of the the things that we've seen discussed about Lyra frames this as an automation company and that's like kind of the the wrong perspective to think about what we're doing. We're not automation maximalists. We are actually sort of like token generation maximalists and flexibility maximalists. So, we will over time automate things um, that make sense to automate and then again, use solutions now um, where they make sense.

12:40 · So, the the system designs the experiments. It gives instructions. There's like, oh, people need to actually do this thing. So, you get you recruit some of the staff to go and do that thing. Everything's an API call.

12:52 · Yeah.

12:52 · And so, sometimes when you call an API, there's a robot arm. Uh, sometimes uh, there's a there's a human arm that does something.

12:59 · literally below the API line.

13:00 · Yeah, well, funny thing. [laughter] Um, I think that like again, we want to spend resources where it makes to spend resources and make rational decisions.

13:08 · And sometimes it just doesn't make sense to try and automate a step when a person can do it in a tenth of a second. But what matters is that the the model has um, the ability to give uh, instructions to test hypothesis and that all of that data is visible, transparent, stored so that those tokens flow back into the model. Do you have your AI models doing entire like experimental designs which are beyond just a pre-existing protocol where you tweak like relative ratios or sources from or like what, you know, all it goes going to a pipe or pipette or something. I mean, it depends on you like your threshold for novelty here.

13:46 · Certainly for expression protocols, for some gene editing work that we've done, we have tested like the the platform's ability to do that versus humans. Model gets like 80% of that zero shot. Humans get 0% of that zero shot. Are we doing like fully open-ended ended free-form experimentation now? I mean, no, not yet. That is the goal. But, we're building towards that. That is the end state that we want to be in.

14:10 · But, we have seen the ability to do what would be an enormous amount of human intellectual labor over a very very short time horizon.

14:20 · So, when you're giving these, you know, giving your AI models kind of free reign to start designing new experiments, like how do you make sure that these are things that should be measured or that you validate that this is, you know, a good strategy or that you didn't just waste a bunch of money?

Safety, security & scientific rigor

14:36 · The first one is I there is something maybe an underlying safety question there. And I think that we've been taking very seriously from the beginning, right? Both security and safety, security of the data and the safety of the model suggestions. We have a very strong team. It's growing under sort of very strong leadership.

14:54 · That's the first layer, right? Like that we have strong safety AI safety protocols that look similar to the sort of uplift considerations that people have been looking into large language models. Only it's absolutely for We have strong AI safety protocols that look similar to the sort of uplift considerations that people have been looking into large language models. Only it's absolutely for real.

15:20 · In a lab automation setting where you're working on some, you know, biophysical material science type problem, what are actually the dangers you have to worry about? This is like I generally think of malicious actors and or you know, situations where you have a sufficiently complicated system that it could genuinely output something dangerous. It seems like from the scope of Lyla as I understand it, which we haven't talked about yet, maybe it'll come in a minute, but it doesn't seem like safety is actually going to be a a major concern at this point.

15:52 · I mean, it's something we need to take seriously from the beginning, right?

15:55 · It's something where we cannot afford to not get it right. I agree with you. Right now, it's in the hands of Lyla employees that are sort of whose interests are aligned and whose understanding of the platform is aligned with with our mission. So, I agree we we don't have to worry about malicious actors. We still need to worry to some degree about the model giving a suggestion.

16:17 · I think it's more some things that it start touching into lab safety more than malicious I don't think we're going to have emergent behavior where the model suggests an extremely toxic chemical.

16:27 · It's more about sort of pushing an instrument such that maybe it is sort of it overflows, it combines chemicals it shouldn't have. So, I think there's like a chemical EHS safety layer that needs to be there since the beginning because we're doing sort of open-ended I mean, I do think Rafa is right in that safety is not something you can procrastinate on because capability curves tend to be sigmoid shaped and it can look like everything's fine everything and then all of a sudden there's something that you didn't anticipate the model being able to do.

16:58 · So, we are definitely proactive on that that side. We have an AI safety team like Rafa said, but I I think they're also right in that like we can constrain the problem in meaningful ways in the way that like a broad-based AI system that interacts with the general public cannot. We can also lean on biosafety levels and things like that, you know, good old fashioned lab safety um to to help in the meantime.

17:20 · And of course the knows that are exposed to a particular we don't necessarily need to expose all the experimental capabilities to all the scientific questions, right? For an antibody design question, we probably don't even need to expose a model to the fact that we have gas canisters that contain gases, right? Because it's not going to need them. So we can still be creative with within sort of questions that relate to one particular area of science.

17:44 · Yeah.

17:45 · Your question though is sorry is interesting like how do you know if something is dangerous is like actually kind of hard to do or actually how do you know if it's wasteful? Um some of the work that we've been doing in like electrocatalysts, we have someone inside of Lila who's like published 40 papers on the topic and some of the suggestions from the model um initially were boring but then um transitioned from boring to what he considered to be stupid. These are non-platinum group um electrocatalyst for separation of hydrogen and oxygen from water to make hydrogen and those turns out to be our best non-platinum group um electrocatalysts um that we've made. So the the line between like obviously wrong and like quasi moved 37 surprising even to a human expert is hard to know and so we we will do wasteful things because we kind of want to know the difference between between the two.

18:30 · That brings up the question um so for like uh an experiment like what you're describing now, it is obvious whether it works or not, right? But you can imagine and there were some controversy um in previous it it the Berkeley Lab around measurements that were misinterpreted, right? How do you know that your measurements of effectiveness or whatever your your optimizing are actually correct?

18:58 · Yeah, I'm very familiar with that part of of the landscape. I would say we cannot relax our standards of scientific rigor because it's AI, right? It's not [clears throat] you know, maybe 5 years ago when you know, we started doing genetic models for X and Y, they were like, "Yeah, it's cute. It kind of works." Like you would do with a kid.

19:18 · It's like But now we're past that and we need to hold AI science to the same standard we hold regular human led science. And I think that's a I think that 2023 paper was a switch over from the community a part of, you know, the AI community AI for science people were always excited to see incremental progress and I think at that point we started collectively touching upon the rest of the community's awareness and they were like, "Fantastic, but you know, now we're going to talk about the way we do things to our highest standard." So, I think we have lots of experimentalists and I want to go back to the API point. Yeah.

19:57 · I think we've had the fortune by starting from zero to build a company where people are sort of AI aware, AI excited. We have sort of across all the people and the networks that I collaborated with uh we've managed to build a team of experimentalists and automation engineers that really believe in the mission and really want to make it happen. They're really taking this graciously, right? So, whenever AI gives something that is sort of very, very wrong and and they're there to just okay, they push the red button, watch out, this is this is a bad idea.

20:29 · Uh but they're also gracious in uh for instance, trying false positives. False positives are terrible for human scientists, right? Because you go to try something, it doesn't work, but the model is fantastic. It reduces uncertainty a lot.

20:41 · For the operator, it's kind of a bummer, right? Because you thought you were going to get something cool. And I think we've managed and going back to the point Andy made, for the you know, I think about until 3 months ago people would be sort of approving AI decisions. And I think about 3 months ago we started seeing that the models' crazy ideas has started being sort of surprising to people, but surprisingly good. It's like, "I don't know. I guess we need to try.

21:09 · And we see the switch over to I think the um the ability of people experimentally to sort of challenge the AI by being gracious. Like that interface of human and and in computers has been very rewarding over the last few months.

21:22 · Also say like um giving the model control of the lab forces you to build infrastructure to expose pieces of data that you would normally want to or care about. And you know, in in maybe um no experiment is wrong, but you want a the ability to explain the outcome. So, if you think about if you have an experiment, you fit a statistical model, what you're what you're trying to do is use variation inputs to explain variations in outputs. And so, we have the ability to explain variation in outputs because we measure so many different things because we have to expose that to the model. So, we can say, "Okay, the humidity was off in the lab that day. Maybe that explains exactly the thing." And then we can also push button and rerun the the experiment to verify. And so, we do not believe like Rafa said, in some sense we have to be more skeptical of any outcome, but we can then quickly go and rerun that experiment um because it's all it's all software effectively.

22:12 · And you find that the team spends a lot of time on verification or like how's the breakdown?

22:18 · Less and less, so I think at at the beginning it wasn't so much I mean there's an execution of okay, we got a hypothesis, we got a set of instructions that's going to go off to the API. Um we vouch for it, and then from part of the API are people doing things. And I think that will stay, right? That's the labor of it. But, I think the uh double-checking that the intuitions were right, I think we're starting to see in this super intelligence local spikes of places where where kind of you know, supporting this emergence of uh super intelligent behavior more than we are sort of gatekeeping and that the ideas are not just wasteful. So, I think that that's happening for domains. And maybe to elaborate a little on the example that that uh Andy mentioned, we we care a lot about energy and sustainability, right? Something that And you know, we're not just a biotech, we really care about energy and sustainability and materials. We're trying to make green hydrogen. And you know, in order to make green hydrogen, you need to use light to split the water molecule. A bunch of that energy you need to pay for because it's it's the energy that's stored in the chemical bond, and that you get from sunlight and electricity. And then there is some overhead that you pay that's called the overpotential, which has to do with the fact that the world is not perfect and things are lossy. And the the loss comes from something called the the catalyst.

23:36 · And today, the catalysts that are out there are okay, but they're expensive and rare. They're made out of ruthenium and and iridium.

23:44 · So, we set up a model to okay, explore what can we do to not use these two elements. And people do these papers and there's, you know, there was something couple of weeks ago that said, "Ruthenium, you know, low ruthenium alloys for XYCs." I well, I mean, yeah, it's like, sure, you can, you know, dope it down, right? So, you can water it down, but still the same fundamental problem.

24:05 · You're just using 50% less. So, we set out the model loose on this type of problem, and we have the ability to make the material, measure the properties, measure the catalysis, measure the stability. And then on the second third generation of sequential learning, right? This this interplay between sort of information and and what the model knows, we started seeing suggestions that were like, I mean, the words were fine. It was using the concepts that we use that just said, you know, I wouldn't apply that idea to that element. I wouldn't have put them together in that way. And it turns out those have been our best performing chemicals so far.

RL, reward hacking & chain-of-thought pathologies

24:41 · I do want to get to the why we're not a biotech, but before we do, one last question along this train thought train.

24:47 · RL is famous for reward hacking. You just I forget what you said. I don't know if you said you were using RL, but or you know, learning iterations. I I be very concerned that, you know, you throw some rewards and that you can really hack the you know physical sciences in a way that you can't do with a compute.

25:04 · Yeah, I'm not going to disagree with that.

25:05 · [laughter] 100% agree with What's what's the funniest example of reward hacking you've seen?

25:09 · We have lots of funny like RL fails that are not explicitly reward hacking. Well, I mean one is when we trained one of the early things we did was like can you just like make a plate map? Like can you like lay the experimental conditions out on a plate?

25:23 · Um and it got annoyed when the person would ask so you would ask the model to do a plate map and it would do it and it be like actually could you change these reagents and it would like swear. It would be like it's a 96-well plate. Come on man. Like it's not that hard. Like in the chain of thought I don't know where that came from but it would like Somewhere on the internet.

25:43 · Yeah

25:43 · yeah somewhere in the the internet's in there. I didn't care forget that. [laughter] So we've seen lots of like funny like personality quirks like that as a function of RL. Um there's like obvious RL I wouldn't call them reward uh reward hacking but pathologies like repetition. So like the chain of thought will collapse and it will just like repeat its final answer over and over and over again.

26:04 · Um for some reason that reliably sometimes leads to higher rewards. Like we're not sure exactly why pathological chain of thoughts or non-legible chain of thoughts in some cases lead to higher rewards.

26:15 · So sorry I want to interrupt. So we're talking about RL. Yeah. We're talking about the way way you're talking about it sounds like just RL on chain of thought just like everybody's doing. But your RL actually has a lab step.

26:31 · Yeah.

26:31 · If you're in a pathological loop does that mean the lab is just like doing the same experiment over and over?

26:35 · It's just not so a a chain of thought maybe just a to step back is tokens that the model uses to solve a problem. So if you were solving a math problem you would do theorem one theorem two corollary lemma you know what you you decompose the problem. In science the chain of thought there's some of that too. So there's reasoning that happens. You know, I'm trying to make a a an antibody for this target. What do I know about this target? What are the known epitopes? Like, what's my plan of attack? Um in the chain of thought are also tool calls. Um so, I'm maybe I'm going to use a structure prediction model in this case to get some read of how the sequence folds in three-dimensional space. Um so, tool calls are part of the chain of thought.

27:12 · Uh at Lyra, the the fun thing is that the lab instruments are also tool calls or a series of tool calls uh if a workflow or a but it's It's all human legible. It's all in English.

27:22 · Right.

27:22 · And so, what some of the pathologies we've seen is it just skips all the middle part, which we would think is important for solving a problem, and just goes right to the answer. And says, "I don't need to do an experiment in this case. I don't need to call a tool." And um you know, in some cases where we can judge um something because maybe we've already done the experiment or something like that. For some reason, it is actually not a bad strategy in some cases. And so, there's some some mystery there.

27:44 · It's a theorist.

27:44 · Yeah.

27:46 · Yeah. Yeah. It's done the calculation.

27:47 · And this is probably too much of a tangent, but like it actually thinks in latent space. It emits uh tokens. So, like uh the chain of thought is often a unreliable narrator for what the model the computation the model's actually doing. And so, one of the big things we're trying to think about is when we're moving into um working on a problem, you know, like Rafa said for electrocatalyst, that we actually don't know what right and wrong looks like.

28:11 · How much should we rely on the chain of thought versus just trusting the experiment, trusting the verifier, trusting the simulator as the the ultimate ground truth.

Why Lila isn't a biotech: the model is the product

28:18 · So, You know, Lyra is not a biotech company. Uh Lyra is actually fairly unique, I think, in this way.

28:25 · You know, I've I've been involved with biotechs. I've helped start biotechs.

28:27 · Often, the goal is to sprint to a clinical trial. So, you want to develop an asset. You develop the platform in service of having optionality of what space you move into. But once you do the the have the clinical asset, put everything into a medically induced coma, and you you get through the clinical trial, and if it goes well, um then other things get to um so, we are taking that option off the table. The model itself is the thing of value at Lyra. So, in that sense, we're much more of like a neo lab, trying to think of a new way to push forward capabilities of a core reasoning LM-based model.

29:01 · Not even the lab platform?

29:03 · Well, so, the lab platform is the token generator.

29:06 · Okay.

29:07 · That is the data generation mechanism that ultimately is the moat for Lyra. Is that um once that continues to scale, um the amount of data that we can generate um both per unit time, but per unit square foot will go up, and that feeds back into the model to make it smarter, that then suggests the next experiment to do.

29:26 · Um, and so, we really are focused on making this core model as performant and smart as possible. Um, uh and we can talk about how that lends itself to different commercial strategies, but ultimately, we're interested in creating this new type of AI model.

29:41 · So, I want to quote Sridhar Kota from Octant Bio who had a great tweet I really loved uh you know, a few weeks ago was, "What is the business model in ML for drug discovery? Because if you need the data to train the model, but if you have the data, what do you need the model for?"

29:54 · That is true when you are narrowly scoped. So, that is true within any given vertical of science. The analogy that I would use is like if you went back 10 years and you tried to create like a coding assistant model, you would just get coding data. You wouldn't also get Shakespeare poetry, carnitas recipes. It It just It turns out that there is spillover as the model is able to train on a broader swath of data and a a deeper cut of data.

30:19 · And so, again, the core bet that we're making is that is true for science.

30:23 · Um, that if the model is trained on an increasingly broad set of data, the amount of data that you need in a given domain, that data requirement is reduced. In some cases, it will be reduced to zero if it's adjacent to what the model has already seen before. And so, there's a data efficiency argument um that would um suggest that again, having a general platform that can create a broad swath of scientific data.

30:45 · I'll also just mention that like obviously we are using things that are already commodities. So So public data sets we use, simulators we use, and uh experimental platform is a complement to these existing uh commodity resources.

30:58 · This brings up a question in my mind about there's a concept of applicability domain where you have different scales, different, and they result in different types of completely different types of information and relationships between entities, right? So you have the the quantum realm, you have chemical realm, you have uh you know, sort of different bio realms. One concern I would have with a cross-cutting approaches is there domain transfer between these at all?

31:30 · Whereas, you know, Carnitas recipes and uh you know, chess problems have the commonality that they're written in language, whereas you almost have a completely separate not even language, right? It's a It's completely separate model between these domains.

31:46 · Human scientists work on all of those domains.

31:49 · Correct. And they mostly communicate with each other in written language using Okay.

31:53 · using tools. I would say there's a there's a common reasoning process that allows um someone to solve problems in each one of those domains. And so I think that that logic carries over to a reasoning model that we're training that again uses tools, can do math, can do code, um but it it's it's having all of that knowledge stored in one place.

32:13 · One classic example for me of domain transfer is um between complexity theory and quantum gravity, right? Where now uh a lot of the quantum gravity theories are are basically recognizing the the the the identical math behind the two of Yep.

32:32 · right? Do you have examples of this kind of sort of oh man, this domain actually applies to this domain. So, we have assembled this um reasoning data set of 10 trillion scientific tokens reasoning traces that are experimentally verified across life sciences, chemistry, and material sciences. And we have seen that this general model often beats the domain specific models. And so, it's it's hard to point to what's in the model that is making it that it what connections it has realized. But clearly having seen more data across all science beats sort of in a sample for sample kind of way domain specific reasoning models.

10 trillion tokens & why the general model wins

33:10 · And the future of of science is language, right?

33:13 · Well, so yeah, so I Yeah, future of chemistry is language.

33:17 · Yeah, yeah.

33:18 · Maybe.

33:20 · [snorts] I don't think it's necessary I don't think that's a necessary condition for a scientific superintelligence. I mean, there was this quote from Demis Hassabis what last week that it's might not be worth distilling all the ways that live in sort of you know Um there are data modalities that are so different from language and you know, and they always tell me, "Well, Rafa, English is Turing complete. So, you could express everything in English."

33:47 · Turing complete languages, but yes, yes, yes.

33:49 · [laughter] So, and I agree with that. There might be places where it's more efficient because of the nature of the I mean, you've done geometric deep learning, right? I think for for geometry and maybe you know, I call my colleague Tess Mead. I think geometry is one of those places where people feel that there might be some just the nature of the problem is more amenable to other architectures. So, if we need to call a protein folding model or we need to call an equivalent diffusion model to make crystal structures, that's fair game. So, I would say um the future of the way science talks with us for sure is through language. That the model needs to be thinking in English about chemistry all the time.

34:27 · Maybe yes, maybe not. Don't think about chemistry in English.

34:31 · They talk about it in English.

34:33 · Exactly.

34:33 · And I I agree with everything Rafa said. Like token-based reasoning with tool use is very powerful and I think the claim that we're making is we have barely scratched the surface for that in science. We're not trying to distill domain-specific models into a reasoning model. It can use those tools productively. And so it's just the combination of reasoning often in English but also in Python and things like that combined with tool use is very powerful and we're very early in science in understanding how far we can push that forward.

35:04 · I see. So can you give some examples of campaigns that you are running that are are representative?

35:14 · Actually, before you do that, let's take back. I realized we still haven't explained that you don't just do bio.

35:20 · Not just any tech bio. So I think for this is a great great lead into this. So not just tech bio, what do you do in terms of science?

Not just TechBio: materials, quantum dots & MOFs

35:28 · science? No.

35:29 · [laughter] Um so maybe yeah. So the the way that we train the model is is all I mean it's across um life sciences are DNA, RNA, proteins, cells, small molecules, different kinds of chemistries and different types of materials. So that is like where we're scoped now um which is admittedly But materials itself is also not just that's also as largely scoped as everything that's in the bio side.

35:53 · We're going to give some examples. So today we can make thin films, we can make powders, we can make quantum dots.

35:59 · We have a cute quantum dot. Are you folks familiar? Quantum dots are the luminescent technology in some TVs and you need to control to make them of exactly the same nanometer size and the nanometer size you make them controls what color they're going to be and you need the purest red and the purest blue and the purest green to make really sharp and and rich color palette for your TV and you need to make them as homogeneous. They all need to be the same. Otherwise the color gets again and blended. So we have a cute demo where our self-driving lab we ask our visitors to pick a wavelength, what color you want your quantum dot to be when they come into the office, and then we fire off the machine, the model reasons, even it sometimes we even throw in new chemicals that the model had never seen just to see what how it moves. The machine is running and by the end of this sort of hour and hour and a half tour, the machine has made maybe one, maybe more generations of quantum dots that tend to hit otherwise we wouldn't do it, right?

36:54 · [laughter] Then to hit the color that people suggested. So, we have the ability to make lots of materials. We can formulate liquids and polymers and soft matter. We care about about energy and sustainability a lot. So, we have a good chunk of electrochemistry capabilities about the interplay of chemical transformations and and electricity as a renewable energy source.

37:14 · Uh we care about traditional catalysis and we care about mechanical properties of materials. And all this comes together in programs where, you know, we make catalysts, we make high-performance uh coatings for corrosion or or aerospace high-performance mechanical applications.

37:31 · And over the last few weeks with a with an external partner, we started multiple sprints of things we weren't doing before that touch, you know, from adhesives to cooling fluids. So, we we've been able to sort of more and more spin up just exciting discoveries in sort of open-ended chemistry and material science spaces.

37:52 · Do you have any, you know, uh connection between quantum dots and let's say protein design?

37:58 · It's the same platform that does that.

38:00 · It's the same set of capabilities. And so, there's a shared infrastructure that lets us do all of those things under the same roof. If there were no um connective tissue, then our ability to do we just would not have the ability to do all those things. Um have we done like the mech and terp thing where we look inside the model and see does this insight from uh electrocatalyst inform We haven't done a deep dive on on the mech and Turk thing. We have seen that our ability to to do these programs has gotten faster as the platform has become more mature.

38:29 · So, is LMP just like a common thing along like is this something really common in your toolkit that because of this this enables like the a large fraction of these ideas that you've just mentioned? It certainly makes sense on the bio side but I I know bio much more than materials. Is that Is that like a common theme amongst your like in your lab toolkit?

38:49 · The AI science factory the more capabilities has the faster we've been able to go after new target product profiles and about new exciting opportunities because you know, the model is prepared to do more things, the lab can do more things, our scientists are more flexible and and more and faster in order to incorporate new capabilities. Adding new instruments has become faster the more instruments we have so they make it goes with sort of what type of company we are sort of the there's echoes of hyper scaling here of scaling in software is backed by scaling in hardware and the fact that we have sort of tens of thousands of of square feet of lab coming online with sort of dozens to hundreds of instruments is giving us this breath to move fast.

39:34 · There are places you folks had a my colleague Heather Kulik here in the podcast recently. One of the areas you were seeing disruption so I think the audience will be familiar with these materials. I don't need to spend a lot of time introducing them. These materials are made of the interaction of a molecule with a metal.

39:52 · And it turns out that our models had been trained on small molecule drug discovery and all of the chemistry that they had learned thinking about drug discovery carried over to start reasoning over these metal organic framework materials that we can use to take CO2 out of the air or or to filter ammonia.

40:10 · I find that fascinating like when I think so many times I've seen people work on machine learning where they train some big data set and then they move to some new domain and often times the amount of transfer you see is small. [laughter] Yes.

40:25 · So, are there like a group of I don't know how to say this, primary colors that you have that you combine together that often times result in your experiments?

40:35 · So, on biology they'd be the obvious candidates, nucleic acid competence, self-re-expression, and then downstream assets were things that we care about. So, they're core competencies that we can then, you know, sort of give rise to a factorial number of different things that you can do. And on the material side I think formulation. I it and it wasn't even it's so I don't know I don't want to say, you know, mundane, but it's so common, it's so important that it wasn't one of the first sort of, you know, super intelligent places we thought of flashier things back at the beginning.

41:09 · And it turns out a lot of people, you know, in industry and in the rest of the world care about formulation meaning mixing liquids and gooey things to make other gooey things. But that's lubricant, that's sleeping nanoparticles, that's deodorant, like there's all these things in consumer products and industrial products and in medicine, you know, gels to to skin grafts, all those things emerge from this sort of mixing gooey materials and that's a muscle that we're we're building. That's that's a very common platform that is showing up all the time the more we talk with people.

Scaling & the "bittersweet lesson" of materials

41:45 · Interesting.

41:46 · I have a rule of thumb that I often use [clears throat] when I'm thinking about scaling, which is that every time you scale an order of magnitude in a system, that your set of problems completely changes. You guys picked the two hardest problems, right? Materials and and and bio.

42:02 · You do other stuff, but materials and bio are notoriously difficult to get to market, right? They it, you know, 10-year 15-year time horizons. And the reasons are especially for materials scaling. So, how are you thinking about that? Are you just saying we're discovery or you saying that we'll get to it? Like what How are you thinking about it?

42:24 · The last academic lecture I prepared before I stopped giving academic lectures called the bittersweet lesson of scaling in materials and chemistry. Because it's this. It turns out, you know, in AI scaling is a good thing because it gives you a roadmap of what you need to do.

42:40 · And in chemistry and materials scaling is a spooky thing because it turns out only the things that that you can scale matter. So, we're extremely cognizant, right? Our product team, our lab team, we all know. For instance, in the quantum dot example, we were able to use the same recipe from a single digit milliliters to a hundred or almost a liter. So, there are places where, you know, our capability today takes bites into scaling and into technology readiness level.

43:09 · Then we are making the system such um that they can reason about what's going to matter later as they're doing the experiments now. And this will be in our rare earth free or sort of platinum group free catalysts, right? Precisely, the nature of the question is that we need to be able to scale these, right?

43:27 · So, it's supply chain conscious as we're firing off the first experiment. We've already read every paper. We've already have a techno-economic analysis agent sitting on this on the corner ready to do the techno-economics of of anything we do. At the end of the day, we're not going to do clinical trials.

43:45 · We're not going to make pilot plants for one particular process that you would put in your refinery, right? At that point, these are places where we will work with our customers or, you know, if we find something so amazing that we don't even need any instruction and we just go sell it. But typically, we will hand off just like we're going to support therapeutic discoveries for our customers, we're going to support materials innovations at the pain points our customers have. And those also have to do with scaling.

The in-vivo CAR-T proof point

44:12 · How far have you gotten so far?

44:15 · So on the life sciences um and I I think I agree with everything Rafa said there.

44:19 · The way to think about like how people would use the platform is or like just kind of like what we're building is much more of like a cloud codish kind of thing for science. So one of the things that has drawn early customers uh to us is so we're we're not an in vivo CAR-T company. There's lots of in vivo CAR-T companies. Super hot right now. We did see um you know 6 months ago with the Capsid acquisition for like 2 plus billion dollars. Um if folks aren't familiar with that like in vivo CAR-T is this very new heart therapeutic modality. Um previously for blood cancers but now increasingly for autoimmune disease. Um we did have in the internal like sort of triumvirate of capabilities that you would need to do in vivo CAR-T. So binder design obviously we we can do that. LMP formulation and then mRNA design.

45:05 · Just so people know what CAR-T is cuz it's really freaking cool.

45:08 · it. It's so cool. Can I can I talk about CAR-T as well?

45:09 · Yeah yeah yeah. Talk about CAR-T.

45:11 · Uh so CAR-T has been worked on since like the late 80s or 90s. Um really caught fire around 2010 or so for for cancers. Um the way it used to work is you'd extract someone's T cells. You would engineer what's called a chimeric antigen receptor that goes on top of that that tells the T cell to what going to go what kind of cell to go and kill.

45:28 · So you're you're basically modifying people's T cells. You take them out. You modify it so that it has this weird antigen receptor on its surface.

45:37 · seek and destroy tag. Um usually they use a protein called CD19 um which is preferentially expressed on B cells. When B cells get malignant they create blood cancers. They create autoimmune diseases. You wipe out someone's almost their entire B cell repertoire when you do this. There's a lot of collateral damage. But essentially you're telling the T cell what to go and kill. So this uh what really started to catch fire around 2015. It was expensive and slow. You have to extract someone's T cells. You have to engineer them. It's like a $400,000 per infusion and just like um it's still a miracle cure for a lots of different types of cancer. Too much of a tangent for this, but there's this child named Emily Whitehead who was treated at the Children's Hospital of Pennsylvania CHOP. Um she was one of the first cures in pediatric cancer by CAR T. She was going to be referred to hospice care, got CAR T. Um another slight slight tangent. She almost died of a fever from this initial CAR T treatment. The only reason she survived because the doctor who was treating her had a daughter with pediatric arthritis and knew that the this specific antibody would blunt her IL-6 response to to CAR T. So like there's a lot to unpack there in terms of AI for science, all the serendipity that had to happen in that specific case for all that to go right.

46:48 · And in probably if you roll that dice a thousand more times, you probably don't get that that that doctor at that moment who knew exactly what antibody to give her to make that treat that treatment curative instead of lethal. Um so again, like those are the types of serendipity things that we we'd actually like to automate. So anyway, it's slow and expensive. People then realized that actually through just an infusion, if you take an mRNA that encodes for the chimeric antigen receptor, you put it in a ball of fat called a lipid nanoparticle, you put a CD8 targeting moiety on the outside of this the ball of fat, it will then you give them in the infusion, it will go bind to the T cell, get ingested, ball of fat dissolves, mRNA comes out, chimeric antigen receptor gets expressed and presents on the top of the T cell.

47:27 · So you're just telling you're reprogramming the T cells to express these weird antigens.

47:33 · Literally programming biology.

47:34 · Yeah. And then the T cell goes and does its thing and wipes out whatever has CD19 in this case.

47:39 · Um so malignant B cells are explain a lot of blood cancer. They also explain a lot of autoimmune diseases. B cells often make antibodies in response to autoantigens and things like that. So recently, six months ago, as the result of about six years worth of work spun out of a Nobel Prize winner's lab, and about a hundred million dollars worth of R&D R&D. Um, we saw Yes, that's some good music, man.

48:03 · [laughter] Um, we saw like some of the most compelling preclinical data for in vivo CAR-T treatment of autoimmune diseases. Uh, it was by a company called Capstan. They were bought by AbbVie for like 2.1 billion dollars. So, uh, at Lyell we had been working on all three of those things in isolation.

48:21 · Um, so about six months ago, uh, a team of two or three people inside of Lyell tried to see, um, what we could do to in the in vivo CAR-T. And what we had been working on, um, was mRNA uh, design. So, like most RNA medicines, the biggest uh, knob that you can turn is expression peak and expression durability. So, how many proteins do you get per unit of mRNA when you give someone a vaccine or some other mRNA medicine? So, we have developed some monster uh, UTRs, untranslated regions which flank the the protein coding region, uh, which dictate those expression properties. Something like 10x the references from Moderna and Pfizer. And over the course of six months got to in vivo data in non-human primates where B cell depletion was significantly better than what was shown in the Capstan data, and the sort of like durability of that, um, was also the all the the characteristics that we looked at were significantly better.

The "zero-FTE startup" model

49:16 · Um, having more CAR expression is probably one of the most potent ways to improve a CAR-T therapy. The number of receptors that get expressed dictates how likely that T cell is to bind to the bad cell once it finds it, and T cells are literally serial killers and that they will go they'll kill a cell, then they'll go to the next one, the next one, and so how long they can do that is dictated by the how durable the expression of the the CAR, um, is.

49:39 · Again, we're not a CAR-T company, we're science nerds, we like to do cool stuff.

49:43 · Um, so we got to that proof point in about six months where again, all the way up to where you might think about filing an IND, um, for a new clinical asset. We're not going to do that, we're not going to do a clinical trial. Again, that would be all all encompassing. But some folks who had been around Lyra for a long time saw that as a way to do essentially like a two to three person FTE startup.

50:04 · Where there's a couple scientists who have domain knowledge and a combination of the model plus platform can do five years worth of biotech work over a six-month period for 10% of the total investment. And so a lot of the like commercial relationships we're thinking about now are essentially like the zero FTE startup model. Where someone comes with an idea they say, you know, if I if there was a CAR-T in the market that could bind to two things, if it was a bispecific, or if it had these other properties, I know the hole in the market that that thing would plug into.

50:33 · And so a lot of our commercial engagements are effectively virtual startups running on um Lyra now, where someone comes with a very well-specified problem. They don't know how to get there. Um there may be some things related to target identification and things like that, too. Um but they can effectively run that entire program over a um a much shorter amount of time with at a fraction of the cost.

50:53 · And so those are those are like a partner comes to you says, I have this idea. I don't want to build a I don't want to build a lab. I don't want to hire a team. I like I want to get there I just want to get it done.

51:02 · Yep.

51:02 · So it could be some academic at a university that's like I I have this idea. I kind of did a little bit of validation. I think it'll work. Can I do it Can I sit with you guys for six months and make it work?

51:12 · Yeah, I mean that that's the right way to think about it. Um yeah, the way that like contractually it plays out is there's like a platform access fee.

51:20 · There's like we have to pay for reagents and running the system and then some overhead and stuff like that. And then there's like some upside sharing. That is a scalable model. Where we can service as the platform gets better, instead of doing dozens of those, we can do hundreds and then thousands of simultaneous, you know, kind of virtual startups being developed on the platform where um we we have revenue that helps pay the bills uh in the near near term, but then we also have this upside partnership with folks who decide to to to build with us.

51:47 · It's amazing because this is what we're seeing is that people are more and more pushing towards getting rid of all the extraneous infrastructure and using automation and focusing on the idea.

52:01 · I mean, the way that I think about it is like most of us got into science because we're curious and want to answer questions. You know, I'm I'm a computer scientist by training and like I like to answer questions through software. However, if I had to program in binary, I would enjoy that significantly less.

52:16 · [laughter] They're They're high-level abstractions, increasingly high-level abstractions.

52:21 · You know, it used to just be Python and Java. Now it's like cloud code that help me answer questions faster. You know, the analogy is that like scientists are still programming in binary. They have a question that they want to answer. They have to just compile that down to an experimental protocol. Then they have to go and do the manual labor and get arthritis by like moving liquids from one That's the equivalent of scientific programming in binary. And so we're trying to help scientists move up the abstraction ladder where What's the like You know, maybe your idea isn't going to work. Most clinical trials fail, but you can at least get to failing fast if you don't have to do both the physical labor and also some of the intellectual labor to, you know, get all the the pieces in the right place.

Clinical translation & loading the die

52:59 · You know, most clinical trials fail you know, somewhere between 5 and 8% of clinical trials actually get from IND to approval.

53:09 · So, the the discovery is not actually the constraint. I was interested You you were you were talking about the sort of economic modeling agent. I can't remember what exactly you called it, but um I mean, that seems like the problem to solve. What How do you think about this?

53:26 · The pro- You mean the the success rate of of clinical trials.

53:29 · Well, the economic model underlying what scaling in general for both bio and materials. Like oftentimes there's this huge uh huge process. Like once you have something which is you consider final, like a IND or D development candidate Yeah. for for material. Like there's still usually like 10 years of clinical trials or you know, qualification for in the material science world to just get that into a product. And Yeah.

53:55 · often times the bottlenecks there are things about scale manufacturing, about regulatory, about safety and things that are often times just very hard to answer up front. So every time you do this you just have to through you know, you know, roll the die. And there is the typical people deal with this is, you know, essentially a portfolio model and you know, financing wise it's it's very much a, you know, the only way you can make money is if you scale with some level of like, you know, uh risk calibration.

54:24 · Yeah.

54:24 · You know, it's really exciting to hear that you can do these things specifically, but how does it feed into the larger thing where even if you solve these problems immediately, it's still only 10% of the problem?

54:37 · The reason why US biotech is losing to Chinese biotech is not because of an innovation problem. There's a regulatory framework too that has to go to enabling like fast clinical trials. The FDA has made motions towards that recently um both for the preclinical data that you have to submit in some cases, but also how we will run and monitor trials. Um so it would be crazy to think that like one company could or even like any company um combined could could change could change that on their own. So it has to be done in tandem with the regulators.

55:08 · However, the minor moves in um preclinical probability of success matter a lot. Um you know, from a portfolio theory perspective it [clears throat] makes the investment much more attractive. It means it you know, in expectation medicines get to patients faster, a fewer of them fail.

55:25 · And so I I would say like, you know, that is the area that we're focusing on now is that a medicine created by um a system that has had the benefit of in this case a million unique mRNA designs to maximize um things that are known to translate to therapeutic benefits will will meaningfully move those pre it still you know I guess I would say it's better to throw a loaded die than it is a fair die and so we're just trying to like make the die as loaded as possible.

55:51 · I guess my thinking this is the thing that I think about constantly is how do you bring you know basically tran- translation right in whatever the equivalent is in materials so how we name that that I really want to see a model that thinks about these are the factors that and and reasons about and is very good at saying I'm filtering my designs to the ones that I think are going to make it through phase three right.

56:19 · My wife is a translational scientist in biotech so I I see it reminds me very often like yeah you guys should be doing AI for translational science.

56:27 · In a sense I think that's sort of some of the echo especially in the materials and chemistry you know our our tools can call process engineering simulators and go figure out what pipe diameters and what heat exchangers you should be using in order to scale up the process for the economics to be worth it. So still maybe there is you know I don't know if we're going to gang up a filter until we go measure it but the ability to reason now about the things that will come downstream which is sort of a little bit what the translational sort of AI would do is is reason now about sort of what's going to matter because I would just say earlier when you have the IND you're locked in and it's true on the chemistry side. The molecule that the sequence you've chosen of course which population you're going to give it to and how you're going to measure success those choices you make afterwards right and I would say to the I don't work on the preclinical stuff but on the chemistry and materials that is precisely the type of behaviors we're trying to instill now for the with the verifiers and the data sources that we can access either because somebody has thought about them, either because the physics allows it or because we can measure good enough proxies now that tells us what's going to happen later.

57:45 · And to be clear like the all those things are things that we talk about internally a lot. We're already on the verge of being pathologically over scoped.

57:52 · [laughter] But I'm absolutely like the the belief that we have is that as models get smarter, as they ingest clinicaltrials.gov, as we partner with pharma companies and get access to, you know, that cookie jar these pre-trial these probabilities will meaningfully change.

58:12 · On the the biomanufacturing side, having access to, you know, biogood manufacturing processes scale-up processes too, we think the models will be able to contribute there. We've just chose to focus a lot of our commercial and collaborative activity on the sort of like frontier of science that we think we can address now, but the goal is to push past push past that.

58:32 · If I may summarize, it's it's kind of like it's a tool call.

58:36 · Yeah, it's all tokens, it's all tool call. Yeah, tokens and tool calls are all you need.

58:40 · But also the reasoning mechanisms that maybe you mentioned for that doctor that treated, you know, the IL antibody IL-6 antibody that you mentioned, well, that person had learned that from you know, a combination of lived experience and reading the literature and we get since we believe our thesis that the breadth gives us there. We will get better at those things by doing more of the things we do.

59:03 · So many counterfactual worlds that doctor was not the one treating Emily Whitehead in that case and CAR-T may have looked like it might have been yet another gravestone in Eroom's Law, you know, for you know, yet another failed drug. So I do think that like that went from like a 2% success probability to a 98% just because that person happened to be in the room. And so if we could just operationalize that, again like the you you going to move a lot of probabilities when you that's a that's a good example of where just having really broad knowledge of scientific information. So that's like that's almost like Google [clears throat] cuz that was only being used in pediatric arthritis. Like another like very niche area of medicine.

Ken Stanley & open-endedness

59:43 · I see.

59:44 · Yeah.

59:44 · So you have Ken Stanley on your team famously written wrote the book why greatness cannot be planned and is very big on open-endedness and serendipity in research. So what is the role of open-endedness at Lila?

59:58 · Oh yeah, one like Ken is awesome. So for those of you who don't know Ken pioneered area of machine learning and AI called open-endedness which I think of as machine creativity. Like how do we get models to do open-ended exploration and also like have a sense of taste about what's interesting what things we should go down. So you can't have scientific superintelligence if you're just a good test taker.

1:00:21 · So like if you think about what reinforcement learning is doing even at scale it's answering questions in kind of like a ruthlessly Vulcan-esque like you know Spock kind of way but you probably only in like limited ways would think of that model as being supremely creative. And so Ken has created or built an open-endedness team at Lila to sort of take take the outer loop or the the meta part of that reasoning challenge on. So how can we get our models to I'm not only be able to answer tough questions but ask interesting questions in the first place. And so that's really Ken's mandate.

1:00:56 · He's been building like a world-class team over the last like several months and they're you know they're in the kitchen cooking now and I think by the end of this year we'll have some cool stuff from Ken's group to share. So we're going to hop into a video here of the lab that's going to show a couple of different things. Okay, so that's probably a peeler or sealer.

Lab video walkthrough & the lab as a data center

1:01:16 · So when you move plates from instrument dimension obviously there's liquid in it most of biology is wet. So so you put these stickers on it. Um, so that that was a a um a plate being sealed. Um, all right, so here we go here.

1:01:30 · It's picking up a Yeah, so this is inside of a liquid handler. Um, let me go We'll wait till it gets to a a wider shot here. Um, so that you can see the PCI bus and see some of the the robotics. So, the liquid handlers is the magnetic Yeah, so this is the planar motor system here where the the plate magnetically levitates. This is PCI bus where the transport layer connects all the instruments. You can see benches um there where all the instruments sit. Robot arm picks it up is now going to transfer it to a different um plate to go on to the next.

1:02:01 · There's a little bit of a traffic control thing that you have to do here like they actually will go and park for a while while traffic congestion clears.

1:02:08 · Yeah.

1:02:09 · Um, and here's a long shot of the PCI bus. And again, like all that's fully controlled, all that's fully automatic. And this is a material science example.

1:02:16 · This is a physical science example, yeah, where it takes us back to the scaling point. Here, it's making our hydrogen catalysts in a scaled-up form factor by uh that's an ink that contains nanoparticles of the material. That's a spin coater as you can guess from the fact that it spins the plates.

1:02:32 · And then, this is a robotic handler moving around little piece of catalyst to test. And this nice-looking purple 90s neon uh vibe, [laughter] this is called a magnetron sputtering machine where we make atoms fly from a source and deposit on the other side of the chamber in a very thin atomic film where we can make arbitrary mixes of elements based on what's on the three four sources. We just vaporize them and make them fly over the chamber and make these nice thin films that are uh very material efficient. We can do this with very very little material and is one of the workhorses for us to design test uh design make test fast in in many applications in catalysis, in corrosion, in mechanical properties.

1:03:17 · Many things you can test in this sort of very convenient form factor.

1:03:20 · The liquid handlers and some of those machines, those are kind of off-the-shelf mostly. And then so you the and then you have this you've come up with this sort of form factor that works for lots of those those machines both for material and for bio.

1:03:38 · Quantum dot is a good example of the combination of the two. It's actually a liquid handler that we've repurposed for quantum dot synthesis and enzyme. And that you know, I think that speaks to like that you can get very far with you know, 20, 30, 40, 50 instruments. The ability they just have to be on platforms that the model can can use them. I think this a big eye-opening thing for me coming into Laila cuz I wasn't I wasn't in lab automation like in any meaningful way before coming to Laila is it's not the automation that I was hoping for. Like a lot of automation is point automation where there is a tablet attached to the side of a liquid handler where you can enter that device is not meant and then sometimes purposely designed not to talk to other things. And so a lot of what we have done has been to you know, I kind of joke that we have the world's largest collection of voided warranties in [laughter] biology because we have written our own custom drivers, our own custom firmware to get sort of low-level granular control over a lot of these instruments and make them make them talk to each other. So the video is cool because you see magnetically levitating plates. What you don't see is like the custom software wrapper that like stitches all that together. And a lot of this comes down to like really hard software hardware interface challenges.

1:04:51 · Some of the machines literally still run Windows 95. And so think about how you automate that. Like we actually have a vision language model controlling a Windows 95 machine.

1:05:00 · Yeah.

1:05:00 · Uh because that's the only way to automate it.

1:05:01 · to joke about a mechanical finger pressing buttons but joke but no, we did that. Like we actually did use a robot to push the iPad on the the side of the thing.

1:05:09 · [laughter] I I the only thing to like the other thing to call here is this is still automation made for people. Like the instruments sit on benches which are approximately chest high because there's the assumption that someone needs to reach in there to service it or to fill the reagents.

1:05:23 · Um so this is the like V0 V0.5 what we think lab automation will look like. And because we've just decided to vertically integrate and own the hardware software stack, the V2 will look very different than this where we'll be able to integrate things. This is happening already on material sciences because those capabilities just don't exist. And you know, we often think about labs in terms of like their XY coordinates. Um as we integrate like we'll have like a Z component too because we'll be able to stack things. So again, tokens per like unit volume is like what we'll be thinking about then. But we we think that like the lab of the future should not be made for people to easily walk into it. It should feel like a data center where you go and you see the rows of server racks. There's room for like a crash cart behind it to to service the nodes. Um but it should be as densely packed as possible and and you know, also as energy efficient as possible and things like that. So um yeah, so we're to answer your your question, we're using commodity things now because it makes sense to get started, but over time almost surely the form factors of those will will change quite a bit.

1:06:23 · I see. I'm I'm just a little surprised that you can come up with this common size of tray Mhm.

1:06:31 · that kind of matches your needs for a good percentage of your problems.

1:06:35 · Well, it's just working backwards. 96 well plates are the like atomic unit of experimentation in lab automation. And so we now do 96 well form factors for material sciences as a result. Not everything fits into that that form factor, but again, the the coverage that you get from adopting a 96 well or 384 well plate format 80/20.

1:06:53 · 80/20, exactly. Yeah.

1:06:54 · Yeah.

1:06:54 · I think that you can see some of those where the the pieces of deposited material were bigger. So we still use the plate shape to carry them over, but then the number of samples, right? That you have them are smaller. They I think some of them are like maybe 12 4 x 3.

Orchestration, scaling & faster assays

1:07:09 · Yeah.

1:07:10 · Um this takes me also to a point you folks asked earlier about scaling and sort of how when you scale your problems are different. And a problem I think we're looking forward to collectively at the company is the orchestration and the scheduling of a data center size AI science factory, right? When it's of all the experiments you could run concurrently, how are you going to think about the you know, the logistics and the orchestration of moving all these samples and interfacing all these instruments to create sort of the maximum value for our customers, the maximum information for our model. And that's exciting part. That problem is going to look very different from some of the other problems we're thinking about now.

1:07:49 · What we think about Rafa said is like orchestration on top of that is like a Slurm queue or something like that that lets you globally maximize throughput of the of the system that you have. But again, using those same abstractions to think about throughput, scheduling, orchestration, um as the system gets complex or gets larger, the you know, the complexity in in maximizing that throughput. So if you're like a you know, a CSP a constraint satisfaction problem nerd, like we have like one of the coolest ones to to think about. Are you thinking about scaling as like that one cluster and then you just cookie cutter that or is it like I have all of my liquid handlers and all of my whatever spin coders and all over [clears throat] in different parts of the the lab or something.

1:08:34 · I mean, currently what we have essentially is one big fully connected graph. And like that won't scale indefinitely. Just some of the material stuff use throw off hazardous fumes and so that's isolated for for safety reasons. I don't know exactly what the exact configuration layout of the science cluster of the future looks like, but I think that it will probably have fewer instruments on it than like you might guess you would need.

1:09:00 · You know, hundreds, maybe thousands. But we do think about scaling it in the same way that you would think about scaling a data center in that it's a multi-level building, occupies millions of square feet, and it's it's like a lights-out facility, as they say. It's like running 24/7, generating data um in real time. And you would want the same uptime that you would expect of a data center. Now, that's very hard to do.

1:09:21 · Uh-huh, yeah, that's like an insanely hard thing to do. Um, but that is the that's the endpoint that we that we are we're trying to work backwards from.

1:09:29 · What problems do you need to solve on the way to that like that endpoint?

1:09:32 · No, this goes back to my previous question though about like the runtime of your experiments too, because scaling means different things, and one of them is experimental design, which intrinsically scaled but maybe at the cost of signal-to-noise ratio or some other idea, but, you know, getting broad data quickly and efficiently Mhm.

1:09:49 · at some cost. Or scaling is, you know, lower throughput but just parallelizing, you know, wildly. So, like in general, I would approach those as two different sets of problems. I don't think the general the same strategy really works for them in general. So, like what types of scaling is more important for you as a scientist?

1:10:09 · I would say like round-over-round iteration is more important than like a broad, hugely multiplexed, uh highly like noisy kind of thing.

1:10:17 · So, iteration time is really the single thing.

1:10:20 · Yeah.

1:10:20 · Okay.

1:10:20 · So, does that limit the domains that you you want to focus on? Like, you know, now do you think like if we're going to try to tackle a new problem, do we ask can we just solve this problem with faster iteration versus something where maybe this the answer is we you scale up by, you know, massively multiplexing something but with like month-long turnaround?

1:10:43 · Parallelizing and multiplexing are are somewhat different, right? So, sometimes that's right.

1:10:48 · I would say pooled we love pooled.

1:10:49 · Yeah.

1:10:50 · So, pooled we love because it's a you get fast and broad.

1:10:53 · What what is pooled made from?

1:10:54 · Pooled are things like DNA-encoded libraries, where you you seek to well have a bunch of crap in it and you can sort out the crap after you do the experiment.

1:11:03 · Somehow the the form of the assay allows you to throw a thousand or a million or a billion experiments at the same time. And the way the assay is set up, the readout picks the winner. So you try a million things in one plate and you get one readout of or a thousand readouts of the thousand winners.

1:11:19 · Multiplex the whole All biotech is just mapping whatever readout you want to on the GSC and yeah, you can PS. Yes, target, you know, multi multiplex. There you go. Yeah, you can get lots of data.

1:11:33 · The elegant argument would be that see if the standard for this field is a month and it's going to take us four days, a four-day learning cycle is amazing because it's sort of it's really going to move the needle for that part of the field. This is where our automation engineers and our and our teams are thinking about uh other ways of measuring things. And you know, in coolants and in catalysis, there are places where we just made different instruments that measure a different property that turns out response a thousand times faster. For instance, in in sorption, I can tell you folks a little bit. Uh in gas sorption, people typically measure they pressurize an amount of gas. Well, for the MOF and COF materials I was talking about sucking CO2 out of the air. You know how much from the ideal gas law, if you remember high school, you know how much gas you put in the little box and then you wait for the gas to be adsorbed in the material, you check the pressure and from the difference in pressure you know how much went into the into the thing. Then you up the pressure again and you see how much extra went. And if this sounds slow, it's because it's very slow. It's called BET. This takes like a day per sample and it's very tough to parallelize because it's another gas line, another canister.

1:12:39 · Um or you can take other types of proxy measurements from other instruments that are parallelizable. And that's something we built in the lab now where instead of measuring pressure, we're measuring another property we care about that is a readout for what actually pressure would tell us, but we can do 96 well plates for 96 metal-organic frameworks [snorts] in like an hour. So, it's like maybe 2,500 times faster.

1:13:04 · So, this is a place where there's a little bit of room for ingenuity or just uh, you know, an hour is still slow compared to other read-outs, right?

1:13:12 · Other things in electrochemistry, maybe we can do in a minute. But, now we're sort of, you know, a thousand times faster than the way we were doing it.

1:13:18 · I think the answer to your question also, too, depends on how much we think the model is starting from a dead start versus a walk versus a jog. Um, so, if there's some area that we care about, some question, it's clear there's like zero knowledge in the weights of the base in the model that we're using, then we may prefer a big slow thing to to move it in. If we think that it's already relatively competent in that, then we would we would vastly prefer the the rapid serial fast iteration um, cycle. So, it will we will do both. Um, the bet is that um, the sort of like um, as as the model performance improves, the sample efficiency goes up, and therefore like the the compound interest that you get from round over round experimentation will outweigh that that you would get from a a big um, uh, noisy but broad data set.

1:14:04 · So, do you have any concern? This is I'm just thinking out loud here, but do you have concern that you're going to quickly sort of saturate the problems that you can solve using Concern or hope?

1:14:17 · Or [laughter] either Okay, okay. Okay.

1:14:19 · Like concern and hope, maybe. But, the um, maybe you have these systems that you're putting in place, and right now because they're new, then there's like a lot of green field you can go and tackle all these problems that are amenable to high-throughput experimentation. You're going to do that for a couple years, maybe, and then all of a sudden now the like everything is different, and you have to like completely retool your your like desillion dollar in I hope that that is true to be clear.

1:14:49 · Like so I hope that we don't have to measure a binding KD again in two years. Like if we didn't have to do that like I'm very pumped about that because the model has essentially mastered binding kinetics.

Instrument onboarding & the 10T-token dataset

1:14:58 · So so you would think that eventually you get to the point where the model knows how to do that you don't Let's go back to the the the PCI bus again. So like what we actually want to do is to reduce the amount it takes the time it takes to bring a new instrument on on platform. So you want that to feel a lot like a USB I don't know how old you guys but like when I was old when I mean when I was a kid you got a new device you got the drivers on a floppy disk. You had to you know beat your head against the wall to get the driver to install and two days later your printer only kind of works.

1:15:27 · Yeah.

1:15:28 · So that's kind of what like Yeah yeah yeah exactly. And if you're a Linux hardcore person you can still live that experience today.

1:15:35 · Your audio driver still doesn't work. Um so that is like what it's like to bring a new instrument on platform in biology and and physical sciences now is that we're in the like driver on a floppy disk and the manual to try to get to work. So again one of the things that we hope a unified platform enables is instrument onboarding time eventually goes to zero where you have the the spec from the manufacturer the model reads it the right APIs get abstracted. We're working with some instrument vendors to make this process easier um but I think a lot of the the way that we think about modulating a system is conditioned on how we do it now. And so again we're hoping that a unified platform makes onboarding instrument two years from now you know a 30-minute exercise versus a 30-day exercise. Again it's a hard thing to do could be wrong we might not be able to do it um but like that is the that's the future that that we're pointing to where currently it we actually can swap out existing instruments very quickly. So if we need to replace a Hamilton you know with a different liquid handler that swap actually happens very quickly already.

1:16:36 · And so we do have some reasonable belief um that onboarding instruments will get faster better more reliable over time. And again, like the the we don't want to be doing 2026 science in 2036. And so we hope that some of these instruments get deprecated or the way that we're measuring things changes. Otherwise, like lots of assumptions we and everyone else made about the rate of progress in the next decade will have been wrong.

1:17:00 · They were wrong.

1:17:01 · And we're already benefiting from you know, the instrument vendors, right?

1:17:05 · Like I wish the problem we have is is what you're describing that we'd run out of science to do with the instruments.

1:17:10 · That would up the ante for the instrument vendors. The instruments we have now are as powerful as a beamline would have been 10 years ago. We're taking measurements today that 10 years ago would have requested you to ask the federal government for a time slot at 2:00 in the morning somewhere out there, you know, to waste a couple of nights of sleep taking measurements at a really bright neutron or x-ray source. And today the vendors make instruments like those that we can put next to you know, the the quantum dot or next to the protein expression.

1:17:40 · So Yeah.

1:17:41 · I'm I I wish you know, that that's an end state that is you know, a desirable but very very unlikely. And I'm sure there's going to be new science to be asking of the instruments we have.

1:17:50 · We've had guests who have had both of these themes. Uh first of all, the none of the devices you buy are set up to do high throughput AI science. And also that there are new scientific devices which come up every day which just like open up something which was impossible like 5, 10 years ago.

1:18:06 · Like inline NMR, there's lots of sort of characterization, miniaturization, and also of more resolution, more bright sources that are just transformational and they marry really well with the kind of automated high throughput science we're doing.

1:18:22 · So we're moving into this facility in Camber in A and I of Massachusetts and it's just a 3D rendering. It's a 100,000 square foot space and we will move towards AMRs, autonomous mobile robots as some of the transport. And so you can see some of that there.

1:18:38 · We'll put it in the show notes. Um okay, so kind of switching topics a little bit. Um so you were talking about your scientific pile of 10 trillion tokens.

1:18:45 · Mhm.

1:18:45 · When I hear 10 trillion, my first thought was, "Man, that sounds like a lot." Things like this is three 3,000 human genomes, which would cost roughly 3 million to sequence.

1:18:54 · Yeah.

1:18:54 · It is uh roughly 1/2000 of the several of these large foundation models like Evo and, you know, nucleotide transformer and so on. So in some sense, it is a lot of data. In other sense, it's not a lot of data. And there's certain Not all tokens are the same. So I'm curious like what went into creating this?

1:19:11 · Mhm.

1:19:12 · What were your thought processes? And then how much actual useful information is in 10,000 tokens? Or 10 trillion tokens?

1:19:18 · yeah.

1:19:18 · So it's tokens in the same way that we think about counting post trade tokens from the internet or from post trade runs. So these are again like RL is The best way to think about RL is a data generation mechanism. It's a way to steer the model towards more and more more and more valuable tokens. Like better tokens. And so these are the result of running that process across many different scientific RL environments at Lyra, where the tokens are a mix of English, tool calls, and um experimental feedback. Um so uh they're, you know, quasi-English tokens as we've been talking about um tokenized by um the tokenizer. Um so that That's where they came from.

1:19:58 · So you're not tokenizing like We're So we're not nucleotizing sequences in general.

1:20:03 · Implicitly, because if the model's asked a question about DNA, like there are DNA tokens in there. It's not like we downloaded DBGap or um the the PDB or um Swiss-Prot or something like that and tokenized at the sequence level. These are reasoning tokens model-generated that are experimentally verified.

1:20:21 · On top of this, you also still have your AlphaFold, your nucleotide transformer, you have all your sequencing data which goes into this. So 10,000 tokens is 10 trillion.

1:20:30 · So 10 trillion. I'm like looking at 10 T on my laptop and The reason why we think that level of data is important, pre-training corpuses are usually somewhere between 15 and 30 trillion tokens. And so, that's the scale at which you see these like emergent things happen. And so, once you're in sort of the trillion token regime, we feel confident that that's enough for the model to start to master and see emergent capabilities.

1:20:56 · So, are you starting from scratch with your model or you have some open source Yeah, well Again, in the interest of being ambitiously over scoped but not pathologically so. We have not decided to take on pre-training as well just because the black magic that you have to do is is insane and we've been gifted, you know, something like a billion dollars worth of compute in the form of open weight models. Yes. So, we start with an open weight model that has been pre-trained and the assumption that we are making is that the model has been pre-trained on the internet and a large fraction of the scientific literature. Therefore, it's a good scientific prior over what is known and therefore a good base base camp to to build upon.

1:21:37 · So, it's it's 10 trillion on top of the trillions that, yeah. Been And we use like we use Nematron quite a bit because we have a partnership with Nvidia. And I think there's like 30 trillion tokens that go into the pre and post training for that model.

1:21:50 · Have you all So, in the process of these reasoning tokens, you are also creating what are arguably probably just rather useful data sets themselves. Have you thought about independently releasing some of those data sets open source and even in the absence of the reasoning model which may still be quite valuable to the community but doesn't actually deteriorate reach your moat at all.

1:22:09 · So, one of the things that we've developed along the way is a test suite of something like a thousand unique scientific RL environments where you can drop in a frontier model, you can drop in your own model, we drop in our models. So, almost surely we're going to open source a subset of that. Um some of it based on data that we've generated, some of it that we have curated um for the community to use. So, there will be some open source version of the, you know, benchmark that we've uh assembled um as doing part of that. And there will be, you know, probably some data training data um that goes along with that.

1:22:42 · Cool.

1:22:43 · Do you have benchmarks internally that are that actually operate the lab?

1:22:47 · Uh like a like a essentially like a benchmark for how well does uh Maybe not the way of saying it. Do you have experimental automated experimental controls?

1:22:55 · Yes, I mean we have I think for every of the we've we've put together from from the beginning of the company sort of these multidisciplinary teams to work on a specific sort of closed-ended problems. And the modus operandi has always been to benchmark training something naively from zero, calling frontier the frontier models that everybody would go sort of use right out of the box, and our own internal So, with everything we've done, we do have a an internal benchmark. Now, the the domains are very specific, right? Right? And they're not as general and all-encompassing as as the benchmark that Andy was describing because they are the things we really care about and and the products that we want to deliver and sort of the places where we want to make a difference.

1:23:36 · Uh but in all those places, we've typically seen that uh the the scientifically pre-trained model that Andy is describing with access to tool calling typically demolishes, of course, anything else that we compare it to.

1:23:50 · mean it's it's worth like thinking about like what we're trying to do, how that is additive with like LLMs.

1:23:58 · If you think about like an experimentally verified reasoning trace, um how many of those do you think exist on the internet or in the pre-training corpus?

1:24:06 · Order of zero?

1:24:07 · Order of zero, yeah. It certainly rounds down to zero versus the next order of magnitude. So, like we have just seen an incredible lift from showing the model that Yeah, even if we're at like a parameter disadvantage relative to the frontier models, just showing it an experimentally verified reasoning reasoning trace. You you see just immediate lift when when we do that.

Lila & the Flagship ecosystem

1:24:27 · Lila is a Flagship company. Flagship is like basically one of the biotech incubators in the world. They've had something like what? I think 30 successful IPOs or I don't know you but your parent Yeah.

1:24:41 · Um, you know, we're all including you yourself. We're just part of generate generate biomedicines um and just had a successful IPO very recently. So, um, you know, Lila is very good at biotech. Um, I would say from history it's very much single asset, you know, traditional biotech.

1:24:59 · Flagship is very good at biotech.

1:25:00 · Flagship. Yes, yes. What did I just say?

1:25:02 · Lila.

1:25:02 · Lila. Yes, yes. Well, yeah, Flagship is very good at biotech cuz you know, very historically been very focused on single assets. I guess in the last few years with generate with um with I guess expedition, Velo, there are some branching out into more platforming things.

1:25:17 · Yep.

1:25:18 · I'm curious about one, how does Lila fit into the broader Flagship ecosystem?

1:25:25 · Was there a specific reason why Lila is now? Like why the sort of pivot from single asset into scientific reasoning and like what is the broader interaction?

1:25:38 · Like in particular you mentioned that you know, you had a drug which was or you had a CAR-T um drug which was at the level of IND. So, you know, you clearly have the ecosystem to make that into something.

1:25:49 · So, I'm curious like well, you know, maybe like where is this going?

1:25:53 · Yeah, great question. Let me do a little Flagship framing and then I'll sort of talk about um so, we all started as the same pluripotent stem cell but there's differentiation [laughter] that we all take. The the traditional path for a Flagship company is there's so, the history of generate is I was an early advisor to generate, a consultant over 2018. Um, there was this idea to use machine learning for protein engineering.

1:26:16 · Um, me and a couple other folks at Flagship um and some other external folks who came in. Dartmouth professor named Gabor Gregorian was part of this.

1:26:24 · Got seed money from Flagship to then go and spin that out. We worked on building the technology and then usually the deal is that Flagship is the sole investor during a series A. And then the series B is normally the first point at which external capital comes into that. To your point, they often end up being asset-based companies. Generate has a phase three trial for a monoclonal antibody to treat asthma, you know, phase one behind that to treat COPD. I think the recognition from some folks at Flagship, especially our CEO Jeff Builtzen, had created, been involved in creating a lot of these companies and he saw, he's like hiring the same team over and over and over again.

1:27:01 · You uh, you need the ML team, you need the the the platform team. And so I think he saw shared DNA between all these companies and like let's have one company that can essentially support all these different things. Year one of Lyra was essentially like when O1 dropped.

1:27:18 · And so we had all these pieces in place and it just became clear that we could create a platform to support a new kind of scientific model. In the early days, we didn't know like how how do you monetize that? What's the commercial strategy? We've gotten a lot of clarity over that um, uh, over the years, but sort of the core conviction that we had two years ago was the bitter lesson is correct.

1:27:39 · Science could be an infinite token generator if Operationally, the way that we're different from a normal Flagship is outside investment came in before the series A. Again, the lead of the series A were was not Flagship. So like we we do have that lineage, we do come from Boston.

1:27:57 · We we do have a lot of the shared learning that company that has created 110 startups. I think so they they normally so Generate was FL 56 57. It was actually a merge. In the early days, Lyda was 96 97. And so the Flagship has this enormous this you know long history of creating companies. So we have that network and we have the learning of leaders who have created that many companies. But we we're such a weird creature that we essentially went down a very different path very very early.

1:28:29 · So why is it that when I hear you know you have a very promising car-t therapy like what you said you had a dizzy like why not just you know partner with that out partner that out or maybe this is on the horizon or something but The short answer is that is that we we we are engaging in commercial partnerships around car-t therapies for sure.

1:28:47 · Some of them are further development to increase some of the or change some of the properties so like you know by specifics and and things like that going after novel indications but we've we've used that one car-t to essentially launch several partnership programs allowed it.

1:29:05 · Okay so it's just sort of like the proof of principle but it itself was not you know quite exactly what you what a drug needed to be or something.

1:29:12 · Well so just to like to be clear like we could go and try and license or partner that specific thing. We found that it was better to take that and secure several partnerships around further development of it.

1:29:26 · You're getting a basically you're you're doing some sort of code development thing where is the this is the virtual startup idea where company starts a virtual startup around one of these indications and they essentially pay us revenue to further development and again we have these milestones and things around it. Yeah.

1:29:47 · So like long term since since Flagship is specifically you know bio and is never really branched into materials how does that sort of weight Flagship or Lyda's strategy does that play into it at all or is like at this point you've kind of launched and sort of used that lot of the resource wise so if you differentiated into something different so quickly right I think part of, you know, the the breadth of the mission clearly was beyond biotech from from day one. Um and the people we needed to hire came from different networks. The instruments where we had to buy came from different vendors than the flagship vendors would have usually been. So, I think that was part of sort of, you know, the reasons why it feels somewhat different. Um but it's also core to the mission, right?

1:30:33 · It's We cannot get these to work uh on a narrow field. By definition, we want to be as broad as we can possibly be because that's where the where the emerging behaviors are going to come from.

1:30:45 · And I think if you looked at the composition of people who work at Lila now, it would look like categorically different than what you would expect like a median biotech company to look like. So, we we hire out of or compete for and sometimes win against people who are considering Frontier Lab offers.

1:31:01 · Um we have a heavy software engineering and tech presence. Um the amount that we spend on GPUs would be atypical for a biotech, I will say. Um I think that, you know, if you if we called ourselves a biopharma, we probably would have a top three GPU cluster in the world. It's true that that's part of our DNA, but we've been intentional about trying to um make decisions that put us on what we think is the most promising trajectory for us. So, this isn't just like kids rebelling against their parents or something. We think that like the thesis is right and it points towards a very like valuable, but also important company for um not just biotech, but for materials and chemistry.

What's harder: materials or biology?

1:31:39 · Okay, that brings me to what I think is my last question. Um what's harder, materials or bio- biology?

1:31:45 · [laughter] They're actually very difficult. It's funny, right? They're they're they're I feel like we're about to do the Spider-Man meme in like [laughter] When I was around for the first uh merry-go-round of AI for the small molecule drug discovery, I mean, the atom wise stride and generate, you know, the in citros. So, I think the the hardest is the thing that is a small molecule. It has all the difficulties of chemistry, of knowledge reasoning or synthesis.

1:32:13 · And then it has all the difficulties of the reasoning about biology and adverse effects and immune response.

1:32:19 · Yes, but the counterpoint being that we have so many tricks in our, you know, toolkit which you can borrow from biology, right? So, it's harder, but you also have Well, I think materials are harder. Um, so they have the benefit of like great simulators like that we don't have in bio.

1:32:32 · Well, I think materials are harder. Um, so they have the benefit of like great simulators like that we don't have in bio.

1:32:38 · Yeah.

1:32:39 · Um Like in material science you don't have the like mature high-throughput um automation that you have in biology. Um, for me materials as a subject is interesting because there's not a unifying principle like the central dogma. Like materials means lots of different things. Like it means like I I actually still don't quite understand the unifying principle when we say material science like what exactly um that means and then the commercial dynamics are completely different. Like like again with CAR-T we know if we wanted to like how to monetize that directly. With material there's a supply chain, there devices, the testing that you do in the lab is only partially predictive of like the lifetime of how that material will be used. Um, and the math is harder.

1:33:19 · I mean like in terms of supply chains still matter for both. Um, you know, maybe you replace clinical trials with some, you know, product validation and verification. Like it's qualification as the term is. Um, so like there are direct analogies and there are there hard parts for both of them.

1:33:34 · Well, the economics are very different.

1:33:36 · Like if you pass a clinical trial, you make money. You're going to like that thing is valuable. And kind of how much it costs to make it is very rarely the blocking element.

1:33:46 · much easier to underwrite an asset in biology than it is in materials. See.

1:33:52 · Do you guys know the name of a company that makes a superconductor? You know, this always come up. Are you guys doing Yeah, we care about magnets, we care about superconductors. They're really cool science. Do you folks know the name of a company that makes super No one knows. Like these things are super important.

1:34:06 · that they're used in MRIs.

1:34:08 · Exactly.

1:34:08 · That's the only commercial application I know.

1:34:09 · But it turns out, right? Like these things when you succeed, you kind of are when you make a cool material that does something, you're kind of a nameless company that makes this thing and is successful and has good cash flows, but but but you don't get to break sort of you know. Everybody knows a big pharma, but other than you know, you've you've got your 3M's, right? This sort of Yeah, and like most of the big material companies are behind closed doors. Like most of commercial engagements look like getting them to tell you what the important problem is. And there's like less of an open innovation ecosystem.

1:34:42 · There's a couple things in materials that are obviously recognized to be valuable, but like it's just I think very different than than life science.

1:34:48 · And maybe one one of the last things that we haven't touched upon a lot of and and I want to flag out. I think in chemistry and especially materials, uh government sponsored research uh is a big driver. So in the in the same way that you know, the government doesn't feel they need to do drug discovery other than than funding NIH for early stage open science, hypothesis-driven science.

1:35:07 · Um you know, the government and national security drive materials innovations in ways that that are unique. And and you see this in the way we engage with the British government. We have partnerships. We work with the US government. We have awards. We participate in sort of developing materials and technologies, um which is a different part of the of the ecosystem that drives innovation. That's also different.

1:35:32 · Yeah, definitely. Are you guys working in Mission Genesis?

1:35:36 · We were one of the named partners. Um we've had an ongoing exist a relationship with a lot of the national labs, and so we have been working on that.

1:35:43 · With this send 25 Yeah.

1:35:45 · Genesis lighthouse proposals last week.

1:35:48 · I was joking. I was joking. I was joking. He left academia thinking Great Writing was behind him only to have to write to [laughter] Yeah.

Bottlenecks, MFU & closing thoughts

1:35:57 · They speak about 20,000.

1:35:59 · Yeah.

1:35:59 · The question that we like to ask all of our guests is if you could remove a bottleneck in your domain by and you can define domain by fiat, Mhm.

1:36:12 · what would that bottleneck be?

1:36:15 · To me I'm going to go to old timing Rafa that was doing physics based simulation.

1:36:18 · I would say the sim to real. I mean sim to real for the people that come from sort of the physics based world. I mean the sim to real having like an Can you explain what that means?

1:36:27 · What I mean? So these people have typically meant it in the context of robotics where your virtual simulations in 3D spaces kind of allow you to train robots that will move in physical spaces, but there's a gap and they call it the sim to real gap. For us in in physics based simulations is that you know, we do molecular simulations of gooey stuff. We do electronic structure simulations of hard stuff and they're okay, but they're not predictive enough. So and this is the reason why you know, if I if it wasn't for that, maybe we wouldn't have had to make a self-driving lab for materials because we would have been able to just predict. So I think we know there's physics, but it doesn't quite go the way to being predictive, meaning that the models that we train on physics cannot possibly close the gap either because they're still missing these rather they're trained on on approximations that are just not good enough. So I think the thing we've been chasing for a decade in AI for materials has been sort of if we train on computational data, can we answer real world experimental questions?

1:37:31 · And that would have been the place where if I if I get to go also go back in time in addition to taking the bottleneck out, it would be the the accuracy of the underlying simulation that we've been training on all the time.

1:37:42 · So this is sort of like Heather Kulik said, there is no AlphaFold for materials.

1:37:47 · Well, the funny thing is AlphaFold was trained on experiments, so it's a different I mean, that's kind of funny.

1:37:51 · Yeah.

1:37:51 · She and and I we both come from doing physics-based simulations, and the fact that she called out something that had no simulations in it whatsoever is kind of a meeting the same underlying issue, which is like all these, you know, Meta has produced tens of millions hundreds of millions of training data points, but they're all virtual simulations that just don't carry enough water for the thing we actually want to do.

1:38:13 · This is going to be like a a boring and obvious one, um but like um there's a a metric that you use to track how efficient your training runs are. Um it's called mean flop utilization or MFU. So, the GPU comes with an advertised like peak flop throughput, which is under the best situation, doing a calculation that you don't actually care about, how many floating-point operations can you do per unit of time.

1:38:37 · Um MFU is always a very small fraction of peak theoretical flops. Um and for reinforcement learning, it's always somewhere like around 5 to like 6%. So, said differently, that means that we're getting like 5% of the actual GPU computing power that we're paying for.

1:38:54 · So, if I could by fiat wave a wand and make our stack perform at like 100% mean flop utilization, um I would do that because we would one get to the answer faster, but then also be able to buy fewer GPUs and redeploy that capital to the to the lab or something like that.

1:39:10 · interesting though because cuz your rollouts, aren't they constrained by the lab?

1:39:14 · They are, but the when we train a big model, like all of that data So, there's RL training pipelines are very complicated. So, like one one way to think about how you would do this at scale is just to have the model doing rollouts left and right waiting for enough trajectories to pile up and then back propagating that into the model. A different way to do that would be to factorize that, have a bunch of expert models that are trained in parallel that are either generating data or um being uh trained themselves, and then you distill that back into the central model. And second way is the most efficient the more efficient way to do that. So, cuz all those things are happening at different time scale and so it's that big when you have the 10 trillion tokens and you want to push them through the model as efficiently as possible, you're still going to be doing some reinforcement learning on top of that. So, like if you if we could get all the flops that we're paying for, I would I would buy fiat declare that.

1:40:07 · Cool.

1:40:08 · Well, yeah, before we end is there anything you want to leave the audience with? Um Let me say like why we're here. So, we have an office in San Francisco now. It's 181 Fremont Street in downtown San Francisco. There's currently 20 20-ish 30-ish people who sit there, but we are looking to expand that aggressively.

1:40:30 · We're looking to pull from sort of all areas of the stack. So, both like post trading obviously aggressively hiring for that. Folks who've been working in like domain AI like life sciences and material sciences we're also hiring for that. No wet lab here currently, so it's all it's all computational work. If any of this stuff that people have heard about today sounds interesting, feel free to shoot either me or Raphael a message if that sounds interesting.

1:40:51 · Thank you for being here. It's been really really fascinating conversation. Appreciate it. Thank you for having us.

1:40:56 · Yeah.

1:40:58 · [music]