The Briefing: AI for Science

[music] Please welcome [music] head of go to market for healthc care and life sciences at Anthropic, Zubar Jandali.

Hello everyone. Good morning and welcome to the briefing for the leaders in this room and everyone joining us on the live stream. Thank you for being here today. My name is Zuber Jandali. I lead our commercial work in healthcare and life sciences. And I've been at Anthropic since we shipped the first Claude and I've seen every release since. There has never been a moment that is more exciting to me than this one right here. And you're about to see why.

Six months ago on this stage, we made a claim that Claude could help with the work of life sciences R&D. Today, we're going to build on that claim with a new one. The claim is this. Claude can run the work, not help with it, not accelerate it, even run it. It's a bold thing to say out loud. So, let me tell you why we believe it.

We've all seen this happen in software development. Coding has irreversibly changed. Now software development is a loop. Write, run, fix, run again. Once AI could sit inside that loop, the loop could keep turning for longer and longer stretches of time before an engineer had to step back in. Two years ago, a few minutes. today, hours, soon, days.

The scientific method is a loop, too. The original loop. Design the experiment, run it, analyze the data, ask the next question. The experiment happens at the bench. The analysis happens at a keyboard. And that is where the loop stalls. Every R&D or in this room knows the shape of it. The experiment that takes three days to run but three or more weeks to analyze. Those weeks those weeks aren't science. They're the toil you endure to get to the science. It's our belief that that toil is collapsing. And as it does, your scientists will get back more of what they trained for. Time at the question.

We've designed today's program around this transition. First, Daria is going to sit down with the scientist who turned GLP1 into a medicine about compressing timelines in biology. Then, we're going to introduce you to what we've built and what it looks like in the hands of a scientist. And finally, we'll wrap with three leading lights from the industry sharing with us how AI has transformed their organizations. Two years ago, our CEO wrote this that AI enabled biology and medicine could compress 50 to 100 years of progress into 5 to 10. At the time, it read as ambition. The morning ahead is our case that it has started.

Now, few people alive have f have carried the arc of a scientific idea all the way through. Please welcome the scientist behind the GLP1 medicines laser award winner and chief former chief scientific off adviser at Novon Nordisk Lier Kudson in conversation with Dario Amod our co-founder and CEO for a conversation about compressing timelines in biology moderated by stats senior writer for medicine Matt Herper.

>> [music] >> So, hello everyone and welcome. I'm Matt Herper. Uh, I'm a journalist at Stat. We're an award-winning medical news site. Um and we're here for a conversation between two people I think we can fairly call revolutionaries about what might be a revolution. Um lot of nuden I mean you've you've heard about her role uh in creating GLP1s. She is a visionary who saw the impact these medicines could have not only in diabetes but obesity. a field by the way and a theme that I think you we want to pay attention to that you joined by accident um and set off one of the biggest medical revolutions I've seen in a 25-y year career. She has won a lot of important awards including the Lasker Prize which is her lapel um the Breakthrough Prize and not least the stat biomedical innovation award. Um, introducing Daario here is a little like introducing Santa at the North Pole, but uh, I want to point out that he did key foundational work in the development of artificial intelligence, but also before that worked in the biohysics of electrophysiology of human circuits and that he has said previously that he thinks this what we're talking about here is the most important thing that that AI can do. So Dario, thank you for having us and uh I can't wait to hear what you guys say to each other.

I I just want to start off lot you've been I we've talked a bit about AI. You've seen revolutions in medicine. You've also you know as well as anyone how the existing system works and what slows it down. Are we on the brink of a big change? Does this change everything?

>> Yeah, I I trul and and thank you for for having me here. Uh it's exciting times. I I really do believe we are on at a real uh inflection point right where there's so much data becoming available every day. There's someone sharing a new idea on how things can be compressed or move forward faster. So I think it's a it's a real uh inflection point, right? And I I've experienced both inflection points as well as hype uh in my my world, right? where the real goal with some tide was to actually create a health impact. Uh and we did that first with cardiovascular deaths being reduced, right? But the hype was the weight loss that really got more people to actually use the medicines. But the real uh goal is the health.

Daario, I'm mean I'm fascinated by this prediction that I mean you it said the progress of 50 to 100 years over five to 10, but that means you're making the progress of a decade in a year. How is that possible? What kind of evidence can we expect for it over the next six to 12 months? Is this actually do you do you still feel this is happening? You you laid it out in your essay two years ago.

Yeah. So, first of all, thank you both for coming. Um uh but >> we're so thrilled. >> Uh um so, you know, one thing I would distinguish between and and I say this uh a little bit a little bit in the essay is uh I you know, you know, I do say like, you know, 10 years from now, I think, you know, we we may be making progress at a rate of 10 years per year. Um I I don't think that today we can make progress at a rate of 10 years per year. um for a number of reasons. One, the technology is still getting better. Uh it's it's on a fast exponential. It's much better than it was uh before. Every new model release is is just better at everything from you know computational biology to you know thinking about proteins and proteins and DNA. But you know we still have some some ways to go on the exponential and and the second is I think just the inertia in the system both the inertia of getting used to uh kind of using these tools and operating in the new way and figuring out you know how do they help with academic biology research how do they help with you know with you know new target discovery how do they help with uh running clinical trials faster and I think the longer response to the regulatory system which is you know it's it's it's going to take a decade for all this new new abilities to be developed with AI and for the whole system which is used to operate in a whole bunch of different ways many parts of which don't yet even believe that AI is going to revolutionize it to to learn to operate with AI but I think once we get it going particularly in all the parts of all particularly in all the parts of the pipeline I absolutely believe that we can make 10 years of progress every year and I think the way it works is, you know, and I thought about it as I was writing Machines of Loving Grace and thought about my own my, you know, my my own history, you know, at least trying to trying to do research in in in in in biology that there are a small number of really essential discoveries. You know, G GLP is one, right? But you know but think of crisper think of you know advances in microscopy for doing contoics and for understanding you know systems systems neuroscience like there are a small number of these discoveries and some of them feel like they could have happened decades earlier than they did right like you know with crisper it's like oh yeah, you know there was someone was studying genetic engineering and then they happened to go to a seminar on you know the bacterial innate immune system and you know no one would have thought that there was a connection but as the AI models get smarter and smarter, I I have to think that they're going to be better at discovering these things early and they're going to discover 10 times or hundred times as many of them. And then it's all going to be about taking those new discoveries which have such broad implications across everything and and kind of using them and translating them both in terms of the science operationally and in terms in terms of the regulatory system to make everything truly go faster. It's it's going to be, you know, it's it's it's it's going to be the work of the work of years, but I think it can be done.

A lot of we I mean, we know the difficulties in this system. I mean, the tough center for the study of drug development estimates it's $2.6 billion to bring a new molecule to market, including cost of failure and capital. I think that's actually even low. It's 10 years to bring a drug to market. How fast do you think the system can start to absor absorb AI and where will that happen?

>> Well, it's already happening, right? There are already areas of the whole drug discovery and development process that already has been revolutionized. And then as the models, as Dary saying, it's just getting so much bigger now. You can can break it all down to little uh areas that you need to work with, but you can you can compress all of them, right? So, I think it's hard to put a number on, right? But some of the work that I've experienced, right? So, so four years to make uh the right molecule. I we still need to make molecules and prove that there are the right ones, but it could probably go to one year. Uh and another area is recruiting patients for clinical trials. You know, some of the stuff we had to do sometimes recruit 20,000 people for a fiveyear study and it takes two years to recruit the people yet people have never been more connected than they are now. So, it must be possible to do that faster. I also think the regulatory process could really be revolutionalized by AI. So there so there's there's just I could go on for a long time thinking about I think so many areas could really become completely uh not completely compressed but really marketly compressed and then there are areas that will be more difficult like uh I'm very passionate about a new target discovery you know how do we find the next JLP1 or the next new biology and that might seem like the most difficult project uh right now but then uh every day there's something new being shared oh now we do this faster this faster and we can make sense of all of this and then add up on top of all those uh wonderful agents you can have doing stuff for you while you're sleeping right so I I think we'll we really will see a a large compression but we still need the clinical trial data so it's it's probably hard to imagine that you can go below five years

>> I mean that's a there's a fundamental that that presents another fundamental problem for anthropic and for AI I compared to other one of the big use cases obviously for AI changing something has been code but you write code and put it in to uh run it and you find out if it runs. How does having an industry with this cycle time where very often you find out your drug didn't work 10 years later when you were convinced the whole time it would. Right? I've watched people go through this process for 25 years. It's an immensely difficult thing. How do you deal with that? How much does the cycle time affect what what AI can do here?

>> Yeah, I I I think that's the thing that's going to control the pace. Um, that's going to be the limiting factor. I think there are there are a lot of things we can do to speed it up or to work around it, but I think it's always going to be the limiting factor. Right? That's why I didn't say, you know, we'll make we'll make a thousand years of progress in in in in in 10 years and and why I'm very specifically not saying we're going to get all this crazy stuff by 2028 or something. we can't we can't get something out out the other end of the pipe. There's just there's I just absolutely no way to do it. So there's this inertia. But I think there's lots of things we can do to work around it or to shorten the cycle time. First of all, I think on the fundamental scientific discovery side, I mean, you know, just imagine you're trying to develop the next crisper or the next GLP1 or the next, you know, Ed Bdon's crazy expansion microscopy thing like there the cycle time is pretty fast, right? you're it may not be your compiles in seconds like code but you can make your cycle time hours depending on the type of experiment you're doing. So if you can speed that up with a relatively short cycle time, then you have all these additional tools. And what those tools allow you to do is lower the cycle time on everything else, right? Where you know, you've you've you know, you've done a better job optimizing the drug. You have better kind of measurement ability. You've found a spade of new targets. So, it allows you to both shorten the cycle time because you can you can be more um you you just you've turned it more from an art into a science, right? We've been gradually doing that much more slowly than we'd like over the over the previous decades, but we can accelerate that process. And note that there are some parts of the slow and expensive process that will speed up once we have things that work better, right? If if we have to do less clinical trials because things work more of the time. If when things work they have a stronger effect then we need to recruit less patients. So this fiveyear thing that you talked about will be shorter. So every part of this long cycle time we can chop. Now maybe we can only chop it from you know five to 10 years to five years or three years or whatever. It's still going to be the controlling thing. And so we're going to have to do a lot of stuff in parallel. That's the only way. But the the biggest issue for drug discovery, the cycle time will help a lot, but the biggest issue is the failure rate, right? And and that's your that's your machines of love and grace piece. And I've seen technologist I talking with Andy Grove about all the ways drug discovery could change in the times since then. You know, the technology industry has Moore's law where things get cheaper and cheaper and faster and faster. And the drug industry has rooms law, which is Mo's law backward. So, do you think that you why is AI different from all these other technologies? We've made genomes go from three billion to $300 and we still spend more on drug development. Why is this different?

>> Yeah. So, I I mean I remember when Andy Grove was talking about these things, there was there was a writer, I forget who it was, who coined like the Andy Grove fallacy of like if you try and think Derek Low >> Derek Low. Okay. Um uh uh if you uh if you just think of biology as kind of this engineering system like we just need to you know make things more rational, make things more make more sense, it you know it just kind of doesn't work right because this is this isn't a design system. It's this like super messy evolved system. Um so uh I I basically agree with that. I think that's right. But the the you know the way we as humans have have only made progress is you know we've used our human brains which are capable of comprehending complexity to to kind of wrestle with that to make sense of it to make progress against it and I think what I'm saying is a AI here is not going to be another engineering technology that organizes our data better or you know tries to unblock one part of the process when there's a world of complexity. It's it's going to be a general purpose technology that helps us to make sense of that complexity in its full complexity better. We don't know for that's a story. We don't know for sure if that's going to work out, but I think we're seeing signs that it's all we're seeing the beginnings of it. It's already starting to. That is the hope. That is what we should shoot for. I I am optimistic. I can't be confident because we don't know the future, but I'm optimistic.

>> I just wonder if you have a thought there, L.

>> Yeah. So I think we'll get better at uh improving the probability of success of new medicines because we'll we'll understand them better. We have a better foundation for why we picked those targets. We can understand the MOA. We can make smarter clinical trials. So I think we will see less failures. But of course we're going to have to listen to Derek Love probably for another 10 years and he's he's very insightful, right? uh um but there will be criticism until uh some examples have moved forward and people are going to say oh this is not fully AI designed but that's again that's not the point the point is that there's so many things you need to know in order to make uh choose the right medicines and develop the right medicines and all the areas of it can be improved upon so you will improve your probability of success for whether a new medicine actually comes out successful but again also many learnings right and a very important learning from Jill P1 is that actually a pleotropic effect uh is a good one right the what's really so fantastic about JLP1 is that you have all these benefits on multiple different organs yet the whole field of drug discovery is still you looking for the genetics with the highest uh window or the mouse model with the highest window they should look at also at you know finding these broad signals and we can help.

>> So you've also said in addition to this being the area that you have the most hope that it's one of the ones you worry about risk most. I mean the obvious worry is could all these cool technologies help people make boweapons or synthetic organisms that run wild or every science fiction novel we've read that actually could happen. How do you guard against that Daria?

Yeah, I mean, you know, I think, you know, Anthropic as a company has has thought about the the risks of AI and, you know, different different fields different fields quite a lot. Um, you know, we're currently, you know, confronting one of those those sets of issues with kind of the cyber risks of AI, right? We've kind of entered uh inflection point or a critical window, you know, as we go along the exponential different things in terms of both benefits and economically useful applications and and, you know, potentially concerning applications turn on at different times. And so we we've had the benefit of seeing the window kind of turn on for cyber. It is not yet turned on for um for for biology. Uh now I think those two examples are are are very different right with with cyber it's like you have a you know you you have you know you have the ability to find exploits and then it's the same model can also patch the exploits and you can find the exploits in a few seconds you can patch the exploits in a few seconds. Biology as we all know is is rather different from that as we know as we we learned during uh during co covid-19 that there's not necessarily the same symmetry but there are some lessons we can learn for example the safeguards that we put on the models um to make sure that uh you know they they can output uh you know beneficial content but can't output worrying content and the difficulty of distinguishing between the two. So we've spent a lot of time putting effort into, you know, what is a helpful query, what is a dangerous query. You see that in cyber where it's like, okay, the model can find bugs. Is that helpful or is it dangerous? It it can be it can be sum of both.

>> How do you balance putting the model putting safeguards on the model versus putting the model in the hands of people who can kind of be the white hat hacker and figure out what you can do with it that's harmful?

>> No. No. Exactly. And and you know I I think this is another thing not just in cyber not just in biology that's going to be the work of years where you have a model that depending on what you do and what you say to it and how you interact with it can do can do lots of wondrous things can create enormous economic value but then there's a tiny slice of things that are actually dangerous and within that a tinier slice of things that are dangerous and you couldn't do without AI or where AI actually is the limiting factor right because we should we should think about that and and so having accurate threat models. I think you both need to have these safeguards and you need to have trusted access programs, right? So, you know, within pharmaceutical companies, people handle dangerous biological material all the time, right? And they have their they have their own protocols for it. So, can we piggyback on those protocols where we say, "Okay, the the elements of society that already handle these potentially dangerous things, can can we just say, okay, you guys, you know, you guys already know how to do this. you're cleared for handling this but you know maybe we don't want to we don't want to give it to you know kind of you know to to to just anyone or just anyone without verifying something. So this idea of kind of trusted access, existing elements of society that manage these risks, safeguards on the general models, it's it's kind of a almost a multi-layer cake of how to make sure we get as close as we can to 100% of the benefits while blocking as close as we can to 100% of the real counterfactual actual dangerous harms, which is a very narrow slice, but that we need to make sure that, you know, we put a proper buffer around to manage appropriately.

L you I mean you've dealt with the problem of developing a drug in tens of thousands of people and giving it to millions and with the worries of side effects that come and you've both seen cases outside of your work where there were drugs that had real side effects that required they be withdrawn and also some of the worries about GLP1s which have mostly not panned out. How do you think about these risk problems?

>> Yeah. So I I think just so in pharmaceutical industry many people know that we're very much used to balancing risk versus benefit and I think it's uh just having a strong focus on it like entropic absolutely has uh and it's it's even I get put in your bylaws that how you are thinking about uh the benefit versus uh the risk and I think just being focused on that having a a really strong view on how to evaluate that and also having maybe an indep so we have independent data monitoring committees uh in in large clinical trials maybe that's also something that might already be in place but really having someone who cannot um who's not motivated by money uh to actually look at that this is being dealt with properly

>> who would that person be

>> so you know I I think I think within uh entropic we have an organization called the Long-Term Benefit Trust. It it governs actually the entire company. It appoints a majority of the board seats. And, you know, we have we have the head of the the Clinton Health Access Initiative. We have a former um uh Supreme Court justice on on California's um uh uh Supreme Court. And we're always looking to add others who have different uh you know, different different experience. I think o over time, you know, it's certainly going to include folks who have you know, experience in uh in in in in biology and medicine. And the one rule for people on the LTBT is they don't have any stock in the company. Um and and so I think exactly this kind of this kind of this kind of independent governance makes sense. Now that's for the the the company as a unit. That's for the corporate structure overall. Um it may also be the case that for specific applications like like bio and cyber it may make sense to have independent monitoring committees either for associated with specific companies or or kind of across the industry. And you know, as as we're seeing, the government probably probably also has a role to play.

So, Dar, I asked Claude what I should ask you. And this was the first question, and I like it because it's tougher than mine. Why should pharma trust AI predictions when your models hallucinate? Where's the validation data for claims that AI accelerates timelines? I thought it would actually

>> I I you know, Claude Claude has been my among my toughest interviewers over over [laughter] the over the over the years. Um, you know, so I I would say on on hallucinations, actually hallucinations have gotten better and better over time. Um, you don't hear as much about hallucinations as you used to. They still happen, but you know, uh, the situation with hallucinations, it's a it's a little bit to make another analogy like the situation with cars that drive themselves where I don't think we will ever have an AI model that never hallucinates. I think just the the probabilistic way in which these models reason, which I suspect is the same as the probabilistic way in which humans reason, is is it it is it is prone to a duality between creativity and you know uh uh basically hallucination, right? In order to be creative, you're often straddling the boundary between making things up and having good ideas. And so I think it's never going to fully go away. the models will just get better at distinguishing between the two, at, you know, at having better filters. Um, again, similar to self-driving cars, like we're just like we're never going to have a human that can perfectly drive a car, like, you know, we're we're never going to have a perfect self, you know, self-driving machine, perfect self-driving algorithm, right? It doesn't know how to deal with like a duck jumps across the road. It doesn't know how to deal with like, you know, there's more, you know, there's like 5 in of ice on the road when previously there's only been 3 in. So, I think hallucinations are going to go down and down and we're just going to have to get used to the idea that as with humans, oh yeah, like this here is a is a is a brilliant person. They have lots of ideas. Many of them are wrong. Sometimes they'll have misconceptions. You know, sometimes, you know, like humans, I imagine the AI models may get, you know, dogmatically advanced. What if your AI model is a Nobel Prize winner, but it's Carrie Mullis and it's talking to a glowing green.

>> I I man, you know, [laughter] I uh there's the list of Nobel Prize winners with that.

>> The line is falling and it just says everybody vitamin >> very long, right? It's a very long list. Um I think that's maybe another illustration of this balance between between creativity and hallucination, right? like a Carrie Mullis. It's like that's, you know, he came up with the brilliant brilliant idea ideas and and yet his hallucination rate is pretty high, isn't it? Um, [laughter] it's really quite high and there were some chemicals involved, I believe.

>> Yeah. Yeah.

>> What question do you hear from skeptics in the industry? You're getting to talk to everybody.

Yeah.

>> That we haven't asked him yet.

>> Um, yeah. You know, most the question I most often get is when I'm out speaking, right, is like, "Yeah, but AI doesn't work for my field, right?" Yeah, lucky for you. But I'm not sure it's true, right? Uh so I for my field I would say, you know, I'm a scientist, right? I would say AI is not going to replace scientists. I'm not worried about that at all. But scientists who are not using AI are going to get replaced and you can expand that to every possible field and every organization. So people should be skeptical absolutely but they should lean in because that is how we can create more solutions uh for the world.

>> You told me a story of a researcher you saw and you were impressed by how many agency had arguing about the work he was doing and and my question for this for a lot of researchers is look as a journalist I can use AI to speed up my work. I can also spend um my entire day arguing with certain LLMs, especially there are particular LLMs for which that's true. Um how do you make sure you're using this productively as a researcher?

>> Yeah. Uh well I guess you got to have a a really strong focus on what is actually is that you're trying to do and then you have to be informed about the what the different models are and the ways of using it and to stay in anthropic language. You know many people understand what a what chat dubity is. They've gotten that. Even my mother who's 92 understands that. But then most people some people stop there. They don't understand the difference between a chat and a code and cloud design and why should they have agents and how should they have their brain bank connected to the agents so the agents keep learning right uh so I think that would be my my take you know more people just need to understand that in order to get the benefits but they should be mindful of what the goal was as I was saying earlier you know for jilopes the goal is the health uh and not get distracted by the hype of the weight loss, but keep being focused on the multiple health benefits.

>> So, we're running low on time now. Um, we're expecting to see a lot of change very fast, but we're also, as we mentioned, in the pharmaceutical industry where things can happen very slowly. I want to ask each of you for two things. What is it you'd most like to see for AI and life sciences over the next year as a positive? And what is it you're hoping you don't see? Um, and I wanted to start with Daria.

>> Yeah. So, um, I think what I'd like to most see, I think Latte mentioned it, is is, uh, uh, uh, some success of, of AI with discovering new targets because I think I think that is the bottleneck. I think the models are just knocking on the door of where they can help a lot with that. Um, and it's on a fast exponential. So you know exponentials really catch you off catch you off guard. Uh uh so you know the the field rising to meet the moment and scientists having the foresight to say well the AI model I had three months ago couldn't help me at all with this but the one I just got today uh you know actually is helping me a lot. um just the attention and the foresight to keep pace with the technology and and and just keep revisiting and understanding how how fast it's improving. So that's what I hope I see. What I hope I don't see is also something we've alluded to, which is which is the kind of uh uh reflexive uh uh skepticism. So I I have the benefit I know, you know, less about uh uh uh less about biology than than some people on this stage, but I have seen how AI gets applied to many many different fields. And there's a story that's the same everywhere. I've I've seen it run through when you know when when AI first beat the world go champion to how well it's performing on code to what we've seen with mythos and cyber over the last uh few months to how AI is doing on mathematics to the quality of AI is writing to to site something that like biology is not as not as easily verifiable um and and the pattern is always the same within any area or it's true also for sub areas of biology the the models are useless, useless, useless. They're dismissed and and then they get to a point where they can help the ordinary practitioner and then those who are the the the most skilled, the most advanced still dismiss them. They're like, "Ah, this is this may help the median person, but it won't actually advance the field." And then the exponential does its thing and and they actually get to the point where they can help you a lot. And it it happens so fast and people get set in their ways because they've seen 12 generations of AI models that don't help them at all. And and then then suddenly overnight the thing you know again we're we're we're you know the last six months the thing that's fresh in my mind is cyber. It's more verifiable than biology. It won't be an exact exact analogy but but I I think it's going to happen. And the thing I don't want is the scientists, the pharmaceutical companies, the regulatory system to be incredibly slow to recognize it because then we'll we'll, you know, we'll get the benefits years later than we would otherwise.

>> So I'll start with the negatives. You know, I have witnessed an absolute revolution in science. It's become a global sport. Uh and it creates wonderful variability of opinions which which means we can solve greater questions together. So I hope I fear the world will stop collaborating because US will have their models, China have their models, Europe not so much yet, right? But I that's my fear, right? Uh and and my hope is that really within the next year that institutions uh and companies really see the potential and actually help their people to understand the full implications of how they can be helped with AI. So I want more uh bilingual people uh in all teams everywhere and of course I don't mean people who speak two languages I mean people who are completely fluent in some scientific topic as well as in digital and AI and and then one person in each team can do wonders you cannot just say to people use AI right you need to actually they need to be there and embedded in team so if everyone does that and please do everyone who's listening because then we can truly accomplish great things together for the benefit of the world.

>> Well, I hope maybe this helps some people become a little more fluent. Uh, thank you very much. We're at time. [applause] [music] [music]

I took two things away from that conversation. One, we can't neatly reverse engineer our way to the vision that Dario set forth in machines of love and grace. And two, that shouldn't inhibit us from simply starting as L called out. And as we've seen AI continue to grow exponentially, organizations that start to track the exponential are beginning to understand that it matters a lot less where you start than the fact that you simply do start. And an anthropic, the person who's been tasked with starting our journey in life sciences is a gentleman named Eric Carter Abrams. I'd like to welcome onto the stage and please join me in doing so, our life sciences leader, Eric Carter Abrams.

[music] [music]

It's great to be here with you all today. So returning to the image of the exponential and how it affects different disciplines, we've seen it happen in coding where we went very quickly from AI participating in autocompleting work to working perhaps at the level of a junior engineer and then rapidly progressing from there to be increasingly autonomous and capable thanks to the underlying progress in our models and products. Now, it's going to take longer for this same change to happen in the life sciences because to give us some credit for a moment, it's a much harder problem. In the life sciences, our feedback loops take longer and they involve running real experiments in the physical world and there's so much uncertainty and noise in all biological data. But our message is that though it's harder in the life sciences and it's going to take a little longer, the same change is absolutely coming.

For our life science efforts, we have two primary objectives that we're pursuing. The first is accelerating scientific discovery as an end in itself, pure pursuit of basic research. And the second is accelerating alleviating the burden of disease and aging. Everything that we're doing is constructed to build a full stack approach to go pursue these objectives as fast as we can. And the approach that we're taking includes many parts. It starts at the foundational layer with our foundation models claude. It includes the product layer of optimizing the product for scientists because even with the most capable models in the world, we still need to have the right product features to make the model intelligence integrated into workflows and accessible to scientists. And then on top of that, as we'll talk more about today, there's all the work that we're doing with all of our partners and customers and internally to directly pursue these objectives that we have.

So we'll start at the model layer. Can AI actually tackle scientific problems? What we've demonstrated over the course of the last six months or so is rapid progress in the underlying foundation model capabilities. Here we're showing our Claude Opus series going from Opus 4.5, which we released about 6 months ago to Opus 4.8, our most recent Opus model. And we're seeing rapid progress in several benchmarks here. We've chosen a few from organic chemistry, biioinformatics, and structural biology. And the models over this time frame have gone from being not all that useful to performing at a level that is on par or greater than the average PhD level scientist in these fields.

So what else do we need to do to solve the problems that scientists face? As I said before, it's not enough to build great models. We also need to build great products. So we the way that we think about it is on top of our foundation models we have a series of products that are designed for different users. For developers we have claude code for knowledge work we have but for scientists we believe that we need something else. There's so many things that are unique about scientific use cases that aren't well captured by the use cases and the workflows of these other fields. In science, there are all these problems that we have that really have little to do with the model capabilities and everything to do with things like connecting to tens of different databases and perfecting every last iteration of the figures that go into manuscripts and performing literature reviews and converting between file formats. There's so much work to do that should be captured in the product layer. And so for scientists, I'm very excited to announce that today we're launching our newest product, Cloud Science.

[applause]

Cloud Science is intended to be your AI workbench that drives all of the work that you do across the whole scientific workflow. I'll go through some of the key features now that make it well suited to scientific use cases. The first is that Cloud Science has a rich set of scientific artifacts that it supports that are fully reproducible. So every time that you create a figure, you have the code history associated with it. So you know the exact set of analyses that got you there. In addition, it supports a wide array of different types of artifacts. Science is a very visual affair. You need to deal with protein structures and small molecules and m multiple sequence alignments, figures in in your papers, whole manuscript drafts, and you need to be able to iterate with Claude in real time and explore these different types of artifacts together.

Next, Cloud Science manages your compute and scales on demand. It's increasingly becoming the case that many scientific workflows are leaning more and more on these high performance scientific computing jobs. For example, in biology running folding models and molecular design models. So we need cloud science to first of all run wherever your data lives. If it's on your local laptop or your cluster and it needs to be able to set up, execute and manage all of the computing jobs that you have. If you want it to run on your cluster, it will run the jobs on your cluster. If you don't have the computer, you prefer that Claude handles it, it will spin up its own compute and GPUs and get the jobs done. And the next feature that I'll talk about is it comes ready for each domain on do on day one. And in biology, this means connecting to tons of different databases and specialized tools so that you don't have to go in and manually configure that. But it is rapidly reconfigurable so that if there are additional tools that you want to connect to, you can set those up very quickly as well.

So we have our underlying frontier models and on top of that we have our products cloud code for developers, cloud co-work for knowledge work and now cloud science for scientists. And next I will welcome up Alec Terashansky, the engineer who has led the development of Cloud Science to walk through a demo.

[music] [music]

I lead product development for Claude science. But earlier in my career, I spent years as a computational biologist, so I've experienced the toil in the scientific process firsthand. Old databases with poorly documented schemas, pipelines that break when a dependency updates, Jupyter notebooks scattered throughout your messy file system with cells that were executed out of order, all those horrible hours spent making figures in Maplot, Lib and Illustrator. Entire fields are underexplored because the research cycle is too slow and too tedious. So how can we accelerate science progress overall and for biology in particular? That question led us to build cloud science, an AI workbench for every stage of scientific research. We've been running it internally on real problems for months and the results have completely transformed what we think is possible. To demonstrate its capabilities, I'm going to walk you through one workflow. The example a real drug program end to end. The disease is fenyl kitinura PKU. One broken enzyme and an amino acid builds up until it damages the brain. There's an approved small molecule drug for it, but it doesn't work in the most common severe mutation. And in fact, that's why other groups are pursuing new therapies for it. Month one of a program like this is a lot of manual work, literature review, structure prep, doing the genetics, scoping the screen, building the business case. 3 to six weeks before a single experiment can be run. So to start, I just gave Claude one sentence. Find me a stabilizer for PH variance. That's the enzyme. and get me up to speed on PKU as the indication so we can put a program together. Claude went ahead and built a plan. So this plan it built has three phases. First do the landscaping analysis. You can see here is going to execute this phase with three parallel sub aents. Clot science is natively multi- aent and uses sub aents to execute the work. First, it'll study the variant biology. Then, it'll look at the structure and assess whether there where the pockets are in the enzyme. And then, it'll build the investment case. Phase two, Claude will build a library. It'll assemble a set of compounds. It'll fan them out across across a bunch of GPUs to assess their binding affinity to the enzyme that we're interested in. And then, it'll score the results. And then finally it'll produce the deliverables ultimately leading to a go no-go verdict. Claude even in cloud science will tell you it's confidence in the scope and feasibility in the plan. Now normally with plans like this where it can take an hour or two to even conduct the computational screen or do the landscaping analysis you would iterate with claude to refine the plan. And in fact that is you know that is encouraged to make sure that it's going to do what you want it to do. here. For the sake of the demo, I just approved it and then I'll show you what it came back with. So, it spun up three sub aents. You can see them here. Variant biology, structure and pockets, investment case. I'll drill into them in a second. The first thing it did though is it confirmed that R48W is the variant that matters. It is in fact the most common severe mutation. And then it did three things that I didn't need to ask it to do. So drilling into the structure and pocket sub agent, you can see here, this is

The brief that the parent agent gave it. You are the structure and pocket sub agent for a PA stabilizer discovery campaign. Here's what you have access to. Here's the steps you should, you know, you should complete. And here's the output schema that you should return.

Now, because the work is being done through sub agents here in this product in in particular, it's extremely important to make sure that you guys have as much transparency and visibility into what's going on at any given time as possible.

So, at the very bottom of this transcript, it produced two figures. The first thing it did is it mapped the mutation onto the crystal structure. So, this is the enzyme. The active site is here and then the mutation site is over here. And you can see it's over 20 anstroms apart. So immediately we can tell that this is likely not a problem with a broken active site. This is more likely a folding problem. So a stabilizer is the right call.

Then it checked to see if we could do the obvious thing. Is there a pocket where the mutation is? So it checked found it was essentially a smooth surface. Zero drugability on a scale of 0 to one. So that dead dead end is closed before we even had it before we even tried it.

And then it built us the business case. It told us who we have to beat sepia taran a drug approved in 2025. It also gave us the regulatory precedent. So I came in asking for a stabilizer and it told me why that's right, where we put it or where we don't put it and who we have to beat. All this was done before I ran a single computational screen.

I want to take a step back a bit and talk about the product primitives that make this possible. First, Claude science ships with capabilities in many different domains. Whether it's proteomics, structural biology, chemistry, genomics, literature review, altogether more than 60 plus databases and scientific resources that it has access to. These these are through the skills that come built into the product and are available for you to use as well as the connectors. Cloud science comes ready for your domain out of the box. And in the event that it doesn't, it's missing capabilities. The product is extraordinarily customizable. You can add a skill simply by chatting with Claude. You can write one from scratch. You can upload it from a zip file or you can import it from your favorite GitHub repository. for MCPs. Any local or remote MCP that you have access to, you can add them.

The second pillar of the product that I want to talk about are the artifacts. Every output that cloud produces comes with its full history attached. So going back to the figure I showed you initially, if you click here and look into the provenence of this artifact, you'll see that the artifact has the code that produced it, including the input artifacts it depends on, the full execution log of all the cells executed in that session leading up to the production of the artifact, the conversation around the artifact, the exact environmental snapshot that produced this artifact, What this means is that you can come back to the product 6 months, a year, two years later. Every artifact is reproducible by construction. And because Claude knows how every artifact was made, this supports a lot of incredibly powerful interaction patterns.

Here, I think the label is hard to see. So, I'm going to just click and say this label is hard to see. And I want to point out one thing here. I could say this without referring to what I'm labeling specifically because Claude sees every annotation that you make with its vision capabilities as well as its ability to read text of course. So you can send this off. I pre-ran this. So I'll show you over here. First message that I sent. These labels are hard to see. You can see the first thing it does is I'll look at how this figure was generated. So I can fix the label legibility. It looks at the code provenence. It makes the precise edit and then it saves a new version of the figure. And you can see at the top over here v2. And that's the other thing. Every artifact is versioned. Versions are immutable and each version has its own provenence attached to it. Later on I thought that this dotted line needed to be more vis more more visible. So I asked for it. It was a little bit off center. So I asked it again, all four versions are permanently on the record and artifacts are checked.

Going back to the structure and pocket sub agent, underneath every agent is a built-in reviewer that is assessing the accuracy of every claim that the agent is making and every artifact it produces. You can see here the agent wrote a brief. This is a markdown document. The reviewer caught a mistake, injected a notice into the agent thread, the agent corrected it, both versions on the record. You can see the diff here. To give you a sense of the power here, in my own personal research project where I continued my own PhD, at this point, Claude has produced thousands and thousands of artifacts all leading up to one output, a manuscript. a manuscript that is by construction fully reproducible end to end. I honestly think that that is the future of science.

Okay, going back to the campaign. So, what's next? We built we did the landscaping analysis, the next thing Claude did is it compiled a focus library of 2200 compounds and then distributed them across 80 GPUs. And that leads us to the third pillar of the product, compute. Cloud science is portable. It runs. We wanted to be as much a drop-in replacement for Jupyter notebooks as possible to make sure that computational biologists can use this in without worrying about it not running where their data is. So it can run on your laptop, it can run on a cloud VM, it can run on a Linux workstation. and it can connect to any SSH host that you have access to whether it's your lab cluster, any EC2 machine you provisioned on AWS in the cloud. We also have built-in integrations with modal as a cloud provider and also we can access model endpoints through Nvidia.

So Claude collected all 80 jobs, they completed, two of them failed, no matter. And then it did the scoring. So 2200 went in, 21,00 or so were folded and assessed for binding affinity with the compounds of interest with the enzyme. 723 passed the thresholds and then four survived a check with an second independent model. And that leads us to the deliverables. So, as you can see, Cloud Science is a very visual product. We have built-in artifact renderers for images, but not just images. As you can see elsewhere here, markdown documents, HTML dashboards, chemistry jars, structure viewers, and more. And we're only going to keep building more because we know that the tale of data types in this space is heavy.

Ultimately, Claude produced a dashboard. Let me zoom out a little bit here. So on the left here are the compounds we've screened. Here's the protein with some of the compounds overlaid. I'll turn them all off. Here at the very bottom of the list, you can't see it because the legend is covering it, but we can always just annotate and say the legend is covering it. And you can iterate with Claude on HTML documents as well. But at the bottom of the list is sepia teran. This is the approved drug and it's ranked dead last for binding affinity to the pocket. At the top of the list, our top four candidates. You can even see the stabilizing arms of the compound. Claude also produced a ranked list of genes with all the metadata attached of compounds. I mean the top four are the ones that survived the check with the second model and then the 700 trailing behind them. And most importantly, it led to the go no-go memo. Claude gave us a conditional go. It told us the first decisive experiment we would need to do to fund a larger experimental campaign and then it gave us the kill criteria for that experiment. At what point do we decide to look for another approach? And that's the workflow. One sentence in and a full campaign and a go no-go verdict out. We went all the way from the bench to the boardroom in a single session.

Now, why stop there? Why do it just for one disease? In parallel, I had asked Claude to run a month one stabiliz stabilizer assessment for a 100 rare monogenic diseases. For each one, tell me which variant matters, what the structure says, and whether the screen is worth running. It did this with a 100 parallel sub aents. Each one doing the landscaping analysis and assessing whether or not this compound is worth computationally screening. At the end of it, we found out of a 100 rare diseases, 32 were worth a fall a computational screen. Altogether, the total time for this was under an hour. total time for the last session including all the the entire computational screen under two hours. So why stop there even? Why a hundred? Why not a thousand? Why not 5,000? Why not 10,000? At this point, scale is no longer an issue because of how good Claude is at actually executing on the campaign and on any computational biology problem in general. So, I said at the start, entire fields sit underexplored because the cycle is too slow and expensive. This is what it looks like when it isn't. To be honest, it's a dream to be able to build this product because if I had this when I was in the lab, the scale at which I could have worked and the ideas I could have afforded to try would have been completely different. And I think that's about to be true for everyone in this room.

to share more about what we what we've been hearing from our early access partners. Please help me in welcoming Eric back to the stage. [music] Everything that Alec just showed you is powered by an open ecosystem of connectors to so many different partners throughout the scientific world, including key partners like Benchling, Bolt, Latch Bio. We've designed this to be easy to add connectors to all the tools that scientists need to use every day.

So now I want to go through a few examples of what some of our early access customers have been doing with cloud science. The first example I'll start with it begins with a very familiar experience for for those of us who have done genomics research. There is a scientist at UCSF who was working on an RNA sequencing data analysis and before using claude science over the course of a year something wasn't quite right with the data and eventually a year later his team figured out that there was a virus contaminating the sample that explained the unusual result. when they provided the data into claude science on the first go over the course of a few minutes cla noticed the virus contaminant and it's these sort of experiences that are critical for building confidence in using AI so thoroughly in scientific workflows

the next example is from a scientist at manifold bio doing drug development and here they call out that with cloud science they're able to go all the way from raw data to publication quality figures in a single session going through all of the analyses, all of the visual iterations on the figures and the whole time having the full history of the changes available.

Our last example hits on what what I consider to be one of the most important themes of claude science. And this is a professor at the Whitehead Institute who shares how claude science is able to take scientists who have primarily experimental biology backgrounds and make them able to perform entire workflows that include the computational part as well. And so in all of these examples, the important themes here are accelerating workflows by an order of magnitude in some cases and enabling small teams of scientists to do what previously would have taken, you know, much larger teams including a larger array of different skill sets.

So in the life science world and taking drug development as an example there's a lot else going on besides R&D. So in this presentation we focused mostly on R&D but of course there is critical work happening in clinical and regulatory in commercial in manufacturing and operations and our cloud offerings from the models and products that we have are designed to address all of it.

So next I'll go through a demo to help illustrate the different use of our products for different parts of this value chain of a scientist that's doing biostats work using claude code our product for software developers. Let's see it in action. So here we have a scientist that's responsible for biostats at a hypothetical therapeutics company and they're tasked with the very difficult job of taking a legacy codebase that's written in SAS and porting it into Python. So to do this they fire up cloud code and they give the prompt in that they need to perform this migration and you see cloud code gets to work. The first thing that cloud code does is it makes a detailed plan for all of the steps that it's going to follow to perform this job. The first step of the plan is to build a dashboard so that you can follow along and see where the agent is over the whole process. After that, it goes through to actually performing the migrations and then critically producing the validation data and all the documentation that's required by regulations.

So we see it start and you see first Claude has quickly whipped up this dashboard so that you can follow along and very quickly it gets into the work of going module by module and actually converting this codebase from SAS to Python. So, as as Claude is working, it's going through the codebase. And a key feature of how Claude code works is that it's set up to flag you and stop if it needs your input on something. And we put a lot of work into making sure that the models have the right judgment and knowing when to grab your attention. So, in this case, you can see through the dashboard side by side the legacy SAS along with the migrated Python. So, you can follow along in real time as the agent is performing this job. But as it's going through the course of this analysis, it's going to come upon an issue and flag that it needs our input on something. So we can see up there that it's flagging that there's something that that it requires our attention for. And in this case, when we click into it to go see what's going on, we can see here that it's detected that the migration is trying to use a package that isn't yet approved, right? And so here we go in and we resolve this particular flag. And this is a good representation of of the way that that we find works best with using these longunning agents. You need to set them up and give them ambitious tasks but make sure that they are programmed to get your attention when they need it and have the right judgment. So here we go through we resolve that conflict. Over the course of the migration, there are one or two others that that will come up. And when we resolve those, from there, it quickly goes through and it completes the migration. And it not only migrates the code, but it produces all of the documentation that's required to perform validation. And at that point, it's complete. So what we've seen here in this a little bit sped up demo is a team of agents through cloud code performing a migration from SAS to Python of a production biostats codebase that otherwise would have taken a team of software engineers and clinical scientists perhaps a few months and can be completed here within a single session over the course of a few hours.

Let's keep moving now looking across the broader life science industry people are building with claude in all of these different areas today from discovery and pre-clinical manufacturing and regulatory commercial clinical interfacing with lab equipment itself. And there's already customers doing enterprisewide deployments that hit on all of these areas. Today, we're very proud to be working with many of the leading organizations throughout the life sciences that have been at the forefront of doing the most aggressive and forward-looking deployments of Claude and AI generally into all facets of their operations. And you can hear you can see on the screen here you know some of their testimonials about this experience.

So coming back to the bigger picture now our objectives the reason that we're doing all of this is to participate in accelerating basic science and reducing the burden of disease and aging. So in doing that our approach is to start with the model training. Everything rests on the models having the right underlying capabilities and to add on top of that the product layer. As we have announced today with our cloud science product, we believe that we need the right products to make the model capabilities integrated into workflows and really solve the problems that scientists are facing. But we've also been asking ourselves, what else should we be doing besides training models and building products? And I'm excited to announce that one answer to that question is that we've decided to start running some drug programs ourselves. And we've chosen to do this by running drug programs in the pre-clinical stage, so the earlier stages of discovery, and choosing indications for neglected disease. So these are areas that are outside the scope of what the traditional pharma and biotech landscape might consider attractive targets, but nonetheless, you know, have real burden associated with them. And we're doing this because we believe first and foremost that to build the right models and products and tools that accelerate the whole industry, we need to live it along with all of you. We believe in the power of tight feedback loops. And there's no substitute for having our own experiences alongside you all in the trenches trying to develop drugs. And the second reason is that we're fortunate due to our public benefit mission to be able to go after these neglected diseases that otherwise wouldn't be addressed. So we're very excited about this new direction and you'll be hearing more from us about this soon.

With that, I will welcome to the stage our head of life science partnerships, Jonah Cool, to moderate a discussion with some of our largest customers. [music] It's a pleasure to be here. Thank you all for joining us. So in the partnerships and deployment group, we are the beneficiaries like all of you of the models and capabilities that you've been hearing about and we get to ask two key questions. One, where are the frontiers of science that we believe we can have a positive impact on? And two, who are the partners that we believe that we can invest and spend our time with to bring that possibility to reality. Now, as a scientist, I want to see these capabilities and models in as many scientists hands as possible across academia, biotech, pharma, and the wider ecosystem. But I also know that as scientists get new technologies, that is actually where many experiments begin. We've all probably experienced a new technology whether it be crym next generation sequencing or perhaps even a computational collaborator that un or unears and opens up our mind to what is possible and how we can pursue that science. But it's also this big question and even an experimental question of how we integrate that into our work, how we integrate that into our organizations and how we start to rethink the questions that are possible. And so it's this topic of how that we're going to jump off and focus on the panel. And it's really a privilege to have an incredible group that come at the pharma industry from very different perspectives but have a similarly expansive view. With that I would like to welcome our panelists to the stage. First Chris Bourner, board chair and CEO of Bristol Myers Squib. Aviv rev the executive vice president and head of research and early development at Janentech and finally Vase uh Nara Simhun the who is both a board member of Anthropic and also the CEO of Novartis. Please welcome Chris, Aiv, and Boss. [music] [applause] >> Great. Thank you all for joining us. >> Great. >> Uh we're going to jump right in. Avivve, I'd like to start with you. >> Yes. and the earliest stages of discovery and thinking a lot about this uh notion of compressing biology that we've heard Lota and Daario talk about and you have thought a lot about as well. Biology we as we know um really is complex in its nature and so sometimes it's less of you know what question you're answering and more the space that you're searching for it in. Help us maybe understand where this notion of lab in the loop and how you're thinking about early stage discovery and integrating AI at Janentech and and where that's going.

So I think it's worthy understanding why biology in general is difficult drug R&D is worse than the general difficulties of biology and it comes from certain kind of inner properties of biology chemistry and and that world the first is it's not just complex it's actually just huge like huge huge all the numbers are big no matter what they are and it's huge enough that it's bigger than our experimental capacity ever it gets bigger than you know number of cells on the planet definitely people on the planet, stars in the universe, atoms in the universe. Choose your number. It's bigger than that. So that's that's problem number one. It's not the only one. Second problem is it's multiscale. So you have to think at many many many scales. We actually saw that in the beautiful presentations from Eric and others. You have to think at the level of atoms and then molecules and then cells and then tissues and then patients and then populations of patients. So that's a lot of scales. And then on top of that, we have a lot of measurement limitations. So we don't measure one thing for all of its properties. We kind of get a lot of separate views of any one entity. Say if you look at a cell, you can look at it with many different measures in many different ways. And then on top of all of those things, that piece that people often call complexity is within this vastness having to wade through it. You can't automate that because it's not exactly the same way. Even when you ask a really similar question to the one you asked before, it's not like perfectly the same. So you can't just automate it in the standard way. you have to somehow operate through it which is what humans have done.

Now you look at AI. So when you have really huge spaces kind of nominally but actually the real dimensionality is lower. Well AI is great at that. When you look at things that are multiscale and as you move from one scale to another you have some nonlinear transformation. AI has proven itself being really great at that. When you have two views of the same thing language and and video and audio whatever take your pick. AI is great at that. And that last piece which is kind of waving through a world that is not exactly specified although there there's kind of a playbook but it's not precise. Well, AI agents are actually great at that. So that's kind of a great great great great and great. And so now really the promise is that you put all of these things together somehow any changes. And then it's important to remember that AI is not magic. It really needs certain conditions to be true. It it needs um surely it needs algorithms and models and it needs GPUs but that's not enough. It needs lot of data and it needs really the ability to understand whether what it did means anything and that means that it needs iterations. So that really leads to this idea of a lab in the loop or a clinic in the loop. By the way, the faster you are in the clinic, the same idea applies that your data are the basis really for learning your models. Your models hold the answer, not the data, but the models hold the answer. And you have to shift to that world view. But the model will really guide you to the next step because the space is so big. We're not going to measure all of it. But we are going to get to a general model in this way. So it will allow you to iterate, repeat, etc., etc. In some way, it's exactly what biologists and scientists have done all along, except it operates at the scale of biology that before was >> was beyond us. And it actually it does actually move the needle. >> It's not the same as it used to be. >> And I think Daario touched on this earlier, but like there is a very unique shape of generative biology that I think matches this space in biology. Yeah. So well uh we heard a lot of great great great and promise where is you know where is it? >> Difficult. Yeah. >> Where are we falling over here? >> The first is first of all let's start with the fact that it's not enough on its own today. Yeah, >> biology and chemistry and so on, they have a huge long tail. So people love focusing on examples that are in the areas where data abound and model perform. But actually for those things to work, yes, the AI might show you a great starting point. We have a nice example like that in oncology. The outcome was what I would call an alien. No human would have thought of that. The humans say that the same, but that was just the starting point. when you needed to figure out the mechanism of action. You actually needed the kind of experimental science and thinking that AI cannot help you with right now. There's no data like that. The experiments are all small scale. They're very bespoke, very specific, very imaginative. And that's great. You put these two things together, you get your answer. So that's one layer of of of real difficulty that needs to be solved. Many of these problems AI performs in an interesting way, but it doesn't take you all the way through. I think it can one day help you get much closer but what it does let you it lets you work in a comprehensive way that's why I say the answer is in the model not in the data before it was only in the data and people have to change how they operate within it and that that shift is also quite difficult >> so on the tone of comprehensiveness and also you know the the breadth of tasks Chris I'd like to transition to you you've been outspoken and talked a about how at BMS you don't want AI simply to accelerate the current processes but actually transform how you operate do work you've also now rolled out claude um across nope over half and so tell us a little bit about both your vision for that and then how how it's going >> well look I mean at a macro level our view is that this technology is ultimately going to transform every piece of the value chain in the slides that you've seen in the previous discussion um and While we're still in the early innings of that journey, uh we think we're already starting to see promise across each of the areas that we're we're utilizing this technology. We're placing at the company really three big bets. The first bet is that AI and machine learning can really help us identify those hidden patterns in biology that ultimately will enable us to um drug previously undruggable targets and hopefully bring be able to bring the next new medicine to patients. We're making good progress. Um, we already today all of our small molecules and a large percentage of our large molecules go through an AI screening and validation process before they ever get into the wet lab. So, we feel good about where we are there. There's a lot more work to do. The second bet is that AI will fundamentally change drug development. It will reduce the times. It will reduce the cost and hopefully improve the probability of success. We've set the target internally that we can reduce cycle times by 30%. We're well on our way to that. will likely beat that target. Um, but good progress there. And third, we do believe this technology will create productivity tailwinds really across the organization. And we've got thousands of use cases uh that we can point to if you double click on that. We rolled out the first chat GPT models in early January, February of 2023, very early on. Today, over 30,000 employees have access to a whole suite of AI tools. And so we've generated thousands and thousands of use cases that we believe ultimately will show a 5 to 10% at at a minimum increase in productivity across various parts of the organization. So great promise we're in the early part of the journey but it's a journey this industry ultimately needs to take because Vosa you can speak to this as well. The business model of this industry will change over time and we think AI and will enable that and if we're successful it should make us better at delivering on our mission which is bring more medicines to patients faster and ultimately change patient outcomes. On that note, Vos, you are a physician. Um, have developed drugs, now the CEO of Nardis. We've spoken about the lab, we've talked about the operations. I I'd love to hear your perspective on the vision and where you see this work impacting patients. >> Yeah. So, thanks. It's great great to be here also with two great colleagues. You know, I think it's worth putting it in perspective. And when you look at the 120 years of this sector, we've really only discovered around 800 to 1,000 medicines. And actually, when you look at actual mechanisms of action that are notable, it's actually quite small. I mean, this is an industry where we spend 150 to 200 billion a year in drug R&D amongst the bigger companies. And yet, we've only found a handful of medicines, which I think actually shows you how hard this is. Unpacking billions of years of evolution. And so the the tools now that that you guys are talking about here today can hopefully get us to to another level. But I think it's important to break down how that's going to work. So when you think about it from a development time time standpoint, I think of three categories of latency that drive drug development timelines. Information latency, operational latency, and biological latency. And information and operational latency are actually about 40% of this time. And I think with the tools you you saw today, we can bring information latency down almost to to zero. I mean the information will be at scientists fingertips. Operational latency we can probably reduce significantly. This is organizing trials, organizing experiments, getting all of the work to happen in a large organization. But I think the biological latency we're stuck with. I mean that is the actually have to run the experiment in a animal model and a cellular model or in humans and that's about 60% of this timeline. So what does that translate to? That means that I think as Chris right rightfully said you can get this down from 12 years from when we actually have a candidate to the end of this journey down to seven to eight years which if you compound over this entire industry is is massive. So I think that's where we'll get the speed. And then when you think about probability of success, again I think I think of four components. In general, drugs fail in phase one because of safety. Can we get a lot better at predicting safety with these models? I hope so. We're certainly trying. I think all of us are trying to to do better on that front. Second is the biohysical properties of the molecule. There I think these [clears throat] models will have a huge benefit. we can leverage all of the historical knowledge to design a better drug that hopefully is better behaved, more manufacturable. Third is the patient selection and actually getting the right indication there remains to be seen. Hopefully we can really get there. The hard one again is the underlying biology of is this a good drug target for this disease. And that comes back to the first topic uh first thing I said that's really hard and in the end after 10 years we might learn that some of these targets aren't the right targets. So take that all together I think we could probably move from 8% to 16%. And if we were actually able to across all of our pipelines go from 12 years to 7 years and 8% POS is 16% the impact on public health is massive. And it's important to understand that those those sound like small moves, but compounded over the the size of our industry pipelines would would be big. And then hopefully some of that means that more diseases get treated, undruggable targets get drugged, and we have, you know, a much bigger public health impact. >> Yeah. That those kind of compounding effects even if it's not perfection, right? And yesterday chatting with Lotus, she gave this great example of the the four years and several thousand compounds that it took to test and get the stability right of GLP-1 and simlutide and again 25% you know if you go to a thousand compounds in a year let alone maybe even more aggressive you know that's an incredible acceleration and >> all the steps of this process are very difficult >> very difficult >> and every step that you move the needle on is great and there's not one that if you just solve that one everything else would be solved. So even if you had the perfect oracle for targets, you would still have to drug them. You would still have to find the right patients. You would still have to execute a clinical trial. That's why it's kind of nice that you can take an all of the above approach and any benefit that you get is a real benefit. As it compounds and also as models can let you work more end to end rather than look at each step separately, then they also can be more um foreshadowing challenges. think about safety way earlier than you would normally think because you have so much that's already happening on the virtual side and on the targets. We look actually as an industry at a very small sliver of them because of this difficulty to manage the comprehensiveness and because of the inability to really track experimentally what you call the biological latency. But if some of this biological latency as we have seen can start becoming virtualized you can run for example not just a small molecule virtual screen for in vitro you can also run small molecule virtual screens phenotypically for cells >> which is possible now and when you do that then biology becomes a lot more accessible in a much broader sense than it has been historically >> but I think it's what you're hearing really across the collective is that there's a lot of possibility and opport and opportunity here with this techn technology, but we also need to make sure we don't set expectations for what we're going to be able to accomplish that we simply can't deliver on. So, when you hear the sort of we're going to cure cancer in our lifetime, >> we're going to make a lot of progress on cancer in our lifetime, but but let's we don't want to get over our skis. >> So, so maybe on this topic of both what didn't work, what is challenging, we just heard that we are starting to pursue some of our preclinical work internally. The big goal there is both to focus on neglected disease but also for ourselves to learn learn these lessons and learn alongside you and really have um seen in the game. You all have have lived this experience. So you know phoning some friends here as we get into this you know what what has not worked. What are the lessons um that you would kind of impart and you know set those expectations at the right time? >> Yeah. Well maybe I'll start the you know from our standpoint I mentioned that we have thousands of use cases. When we started rolling out this technology, our intention was to let a thousand flowers bloom. Let people take this technology and do what they would like and and try to improve their ability to be productive and and efficient. And that worked. Um, but what we found was that the use cases tended to be quite narrow and were very difficult to scale. And that sort of makes sense, right? Because you're developing these tools for a very individualized case. you're putting it on top of processes, but you're not changing the process itself. Um, and so we really were struggling with scaling. I'll give you one sort of somewhat mundane example, but we knew two years ago that AI could predict and forecast our business and many parts of our businesses anyway, way better than the battalions of people that we had actually doing the forecasting. But we we couldn't scale it. And the reason we couldn't scale is when you we talk talk about changing the forecast and planning process of a company of our size. That's not a business analytics exercise. That's a finance, commercial, manufacturing. The list goes on. And you simply couldn't get the scale we needed. So what we realized very quickly is that we needed to supplement this sort of bottoms up innovation with a very robust and rigorous top- down approach. So we created this AI accelerator in the company. We had my team say where within the key vertical in the organization can you get real productivity improvements? Where can you change the processes leveraging this technology? We put small teams that had six to eight weeks to prove out the concept and then we pushed to scale that and so we have I think 40 or 50 projects in the incubator now. Uh we've gotten 30 ongoing efforts that are pretty far down the path. So this ability to take bottoms up and top down approaches to these problems so that you can overcome the organizational and people issues associated with fully leveraging this technology is super important. >> So I I I'll give you a perspective more from the research side. So before development just to just to complement what what Chris described is actually what we would call reshaping of a process. Like the AI problem is actually not an AI problem at all. You kind of know to solve it it demonstrates success. How do you take the whole company's process to change? But on the research side, we didn't come from a million flowers bloom. We came from a very particular hypothesis that is now years old. So we started in 2020 on this that is like a big idea. This lab in the loop that you can really change how you do the drug discovery part, the target and the drug discovery part. And that required several shifts in like scientific work. So for example, we needed certain styles of data to exist. So we invested a lot in data generation capacity that would be the right fit for what AI needs and then something unexpected happened. It wasn't a problem to set up the data generation capacity but once the data were generated you know these are kind of cool data people forgot that the goal was to end up with an AI model. They started looking at the data >> the and it wasn't just the experimental people the computational people too because there's like real results there. We of course build them for our systems and so on. So we had to to pivot for that and we called something foundational data sets for foundation models. It's an initiative so that people remember that the goal is to train the model and at first the reaction was like but why are you making me do this? the data are there. I want this for my problem, for my system. And within not a long period of time, a few months, people were like, why is the foundation model not trained yet? They were so realizing how much more is in a model than in data. And similarly, and this might actually be useful for you. Once you start on projects, people want that molecule to succeed. >> And now their goal is not to learn how to do the process in general and and rightfully so because we want to change the lives of patients in the end. It's not to make the model general. It's not to learn lessons. It's to actually get a project into patients. And when that happens, they're like, why should I spend synthesis money just to make the model better? Even though that would be good for all projects, I need to move my project. And so shifting that also required some thinking and work. How do you still move all your projects, but but also generalize from it? And that's very natural when your goal is to impact patients, but you still want to think big. So you have to you have to build something in there and and like you you know it's kind of in all molecules now there's something with AI. Sometimes a big thing sometimes a smaller thing depending on the project the challenges and so on but there's a something and that's a big shift from a few years ago. Maybe maybe the only thing I'll add in terms of advice, I think all the math shows when you look at pipelines that to Viv's point before phase 2A, you need to set high bars and be willing to let things fail. And I think one of the challenges we're going to have with AI is the ability to generate molecules is going to increase dramatically. So decision-making criteria is going to be really important. So setting really good experiments whether it's in pre-clinical or or in early clinical to really kill early based on hard data. And then the other part of the math is once you find an active drug and it actually is hitting the target and has a pharmacodnamic effect you need to find its use case in late stage and that that's where you don't want to give up too soon that and the examples are many in our industry where we eventually after many years actually find the right use case in late stage for a drug. I I have to comment on that in the decision- making because I agree so much for us the entries into the research portfolio more than doubled in two years. Doubled is a lot more than like 10%. It's a it was you know 130% or so. That's huge. And we didn't add any more biologists during that process. It's just and it's not just the AI it's also the data and data generation capacity that we had. And so actually how you manage the process of vetting and entering became also a thing we had to change because it was just not the old process was not made for this with this type of approach. And now if you multiply it across the whole industry and tools like the ones that we saw today will surely get there. then the volume just in the whole system all of a sudden rises and there is also the risk that people will do the same thing again again again again again again again again again again again again again again again again again again again again again again again again which is not actually a very desirable outcome so I think there's a lot of room for thinking about how we make our decisions yeah and science >> the the vision and perspective that you all have of leading some of the largest organizations in the world >> and the challenges of of scale here are very real ones and and with that scale of course also comes difficulties and a change of shape of how work is getting implemented, the power, the potential. And there's risks associated with that. And perhaps one of the risks that we've we've chatted a little bit about is this creep towards mediocrity. >> And especially if we're talking about, you know, again, all phases of science from discovery to drug development to patients. What we do not want more of is more mediocre discoveries or mediocre drugs. We we want to maintain that level of excellence. This this in my mind is probably the biggest thing I worry about. Um there are all sorts of things that you worry about malfeasants and the like and those are very they're very real and I don't want to downplay them but but in my mind my worry is that employees and society in general goes on autopilot because when you have tools that can create content they can edit content summarize take action on your behalf it's very easy to phone it in and effectively say well you know what I'm going to stop checking my homework or interrogating the output it and in our sector that's existential. Um I was very fortunate when I was at Genentech and started my career the person who hired me said we we hire the best and the brightest because the problems that we have to deal with in our industry are the most complex problems that we have to deal with. You've heard that in spades on this discussion. Um and the moment we stop bringing our intelligence and our creativity and our discipline and the approach that we take towards science to the table, that's a big problem. And so, you know, I think you don't solve that with governance or guard rails. You solve that with a culture of how you engage this technology. That's much harder to do. And you got to be super thoughtful about how you approach it. >> I will take a slight I I worry about mediocrity a lot, but in a slightly different style than Chris's, although I agree with everything he said. I worry

about the narcissistic version of the model. So, like our narcissism meaning the model reflects back to us ourselves, and it makes us more and more embroiled in our previous collective scientific self and makes it harder and harder, in a sense, to break free.

So, people do all the right things in my, in my nightmare. They still check their homework. They still read everything that's written there, but that almost all almost infects their brain, and they become even more and more ingrained in that. And the reason I worry about that is that this is something that is actually not new. It's something that happens in scientific communities all the time. And you see that happen to fields that, as we call them, are well-established. Everyone talks in this way and gets kind of embroiled in their own field. It happens to large and illustrious companies over their many, many, many decades and more than decades, centuries of existence. They have like their way of doing things, and then some other abstract comes and disrupts the whole thing because they're able to break free from that.

So, there is a risk that the model kind of takes over our thinking, emulates it beyond perfection, has the broader, bigger brain because of this, and everyone gets sucked into that world and is unable to break free. And I think there are ways to mitigate for that. You can mitigate for that with the modeling itself. But we have to think about it. We have to be extraordinarily careful for the evals not to make that happen to our models. And it is not a trivial problem to solve. There are real examples now of people getting excited about, you know, you can get code to beat any, any other approach. But a lot of it is actually how it's evaluated. And those evals are actually where it gets an edge, but they're not the right evals for novelty.

And then you can also mitigate for that, especially in our field, by using the real world, which is the lab, to let you kind of escape free from where you're already at. And unless we do this, I actually would be very worried that we'll keep kind of regurgitating the same thing again and again and again and again, instilling this discernment. And that's narcissism. It's like watching yourself, like a narcissist in the pond and liking it so much and just getting fixed into that. So, you know, the ancient Greeks already told us everything.

Maybe, maybe the only thing I'd add is on, on the from a public policy standpoint. I think clearly we need to both have regulation, and I think Anthropic is a leader on this, and Dario has written, I think eloquently recently about the need for regulation on AI systems to mitigate a lot of the threats. And at the same time, I think we're going to have to modernize our drug regulatory system to fully leverage these technologies. If we can actually better predict pre-clinical safety, maybe we can do more streamlined animal models. If we can better model control arms and control groups, maybe we can do truly do synthetic control arms, which has always been a challenge with, except in specific instances, with the FDA. Maybe we can better model dosing and not have to do as much dose range finding. I mean, all of these things are possible, but we also need to bring the drug regulator along with us to get the full benefits of speed and POS we've been, we've been talking about.

Okay. So, uh, we're going to end with a final question for everyone. Uh, it's the same question that L and and Dario were asked. So, when we come back next year, and hopefully we're all sitting here or catching up, what, what is the one thing that you want to be held to, and also one thing that you, you hope to be proven wrong about? Moss, why don't you start this time?

Yeah, I think to be held to that we can actually demonstrate that these technologies are starting to have an impact on, on the speed of our, our drug R&D process because I think we've made a lot of bold proclamations, and now I think we need to actually show for patients that we actually are delivering real, real results. Um, and I hope I'm proven wrong that it's not going to take a crisis to actually get good regulation to happen because I think it would be a shame that a crisis is what pushes us to to get appropriate AI regulation in place.

Um, I think for many of the pieces, we already have had them in place now for a while, including, you know, real deployment and use. I want to see more of the end to end, like that the model looks further down the road, and the pieces are more connected together, rather than here's my solution for target discovery, and here's my solution for a small molecule, and for this step in the process, and for that step in the process, and for the other step. I think putting them together would be a good, a good aspiration. And what I'm, I hope I'll be proven wrong. I think it is reasonable to assume that people's stress levels are rising, and increasingly so, um, in companies, in, in other scientific communities, like in academia, in society as a whole. I really hope because I really want to be proven wrong on on that rising stress level, that we'll start seeing people's joy in the process, their real enjoyment from the fact that something new and exciting writing is in their hands, that they can use a new thing in new ways, not just in old ways, and that, and to see the arc of that, of that positivity, because because I think that's extremely important for getting a good outcome in the end.

From my standpoint, I just build on what, what Voss said. I think that there's a lot of hard work for us to put around how do we actually start to demonstrate the return on the significant investments that we're making. Increasingly, investors are asking us not what you're investing in or or begging us to make more investments. Oh, there's some of that, but now they're saying, "Okay, what's the return?" And that's a tricky, that's a tricky thing to get at. We had a month ago or so, one of our phase one programs went into phase two. We started to see this kind of gnarly side effect that fortunately we had the data curated. We had AI tools sitting on top of it. And within a week, we were able to very clearly articulate a very small change that had been made in manufacturing that probably saved the program and certainly saved the program six to twelve months of time. How do you define the return on that investment? That's tricky, but we've got to figure it out. And if we don't figure it out, the thing I worry about, and I hope I'm proven wrong on, is at some point organizations, certainly large organizations, are going to become put under pressure to ration the use of this technology. Um, and in fact, one of the reasons we did the Anthropic partnership was we've now embedded Anthropic across all of our work processes, and we hopefully will be able to now apply the right tools to the right problem so that we're not using the most advanced tools to summarize people's emails. Um, because that's not the way to get a good return on this investment. And I think we've got to get more sophisticated around that concept generally.

So, as a scientist, I always feel compelled, and it's, it's most critical to be both rigorous and, uh, really honest with these new technologies and to think about how we're deploying them. Uh, when I ran a research group, I used to always remind the group that the only bad experiment was the one that we didn't carefully analyze. And I'm really grateful to the panel for the time because I think we heard how AI and Claude is touching all parts of science and drug development, as well as the business. And also that there's still plenty of room to improve, right? We're, we're at early stages, and the impacts are real, but, uh, the upside is still ahead of us. And so, very grateful for, for the time. A little bit of a performance review for Claude. So, you know, we'll, we'll make sure we keep working. And finally, thank, uh, all the audience and again for that discussion and partnership, and also the work going forward, both on model capabilities, but also doing these experiments and how we get them to patients. Thank you all. Great. Thank you.

[applause and music]

[music]

Please welcome back to the stage, Zubar Jandali.

[music]

All right, bringing us home. Six months ago, we told you that Claude could help with the digital work of life sciences R&D. This morning, I said something different, that Claude could run that work. We told you that the morning ahead would, would, would include us laying out the case for it. And here's how we spent it. Eric and D, sorry, Latt and Dario walked us through their unvarnished views on how compression may or may not happen, the challenges and opportunity ahead. Eric and Alec walked you through the basis of our claim. Models that meaningfully improve in biology with every release, and a workbench that runs the analysis your scientists actually run. Pipelines, figures, every step reproducible and traceable. And Voss, Aviv, and Chris just walked us through how they're charting the path of AI transformation at BMS, Genentech, and Artis.

In the end of the day, this is just the beginning. To build an intelligence platform where scientific discovery truly happens is going to be a collaboration with the industry and our partners represented here. So, we'd love to invite you into this journey, and there are a couple of ways to start. Number one, as far as cloud science goes, try it. The QR code on the screen will give you access to it and put it right in your hands. And secondly, join us for a hackathon with the Gladstone Institutes from July 6th to the 12th. Our first ever that'll feature Claude science.

In the end, this is about time at the question, giving your scientists back what they trained for, and you deciding which questions they'll spend it on. We'd like to invite you to ask the questions as well. The demos are at that door, are at the door. So, pick the question, give it the work, and see how far it runs. To everyone who joined us here today, both in person and virtually, thank you.

[applause]

[music]

[music]

[music]

[music]

[music]

[music]

Hey, hey, hey, hey, hey, hey.

[music]

[music]

Hey, hey, hey, hey, hey, hey.