- Anthropic is dedicating significant resources to applying its Frontier AI, Claude, within the life sciences, aiming to accelerate R&D by an order of magnitude and deliver beneficial impact.
- Claude is being developed as a "superhuman research assistant" to empower individual scientists by streamlining tasks across the entire research spectrum, from hypothesis generation to regulatory submissions.
- Recent advancements, particularly in Claude 4.5, include extensive scientific training and enhanced capabilities for long-horizon tasks and complex bioinformatics workflows, showcasing its potential to make scientific work more efficient and enjoyable.
Introducing Claude for Life Sciences
- Life Sciences as Primary Focus: Anthropic prioritizes the life sciences as the number one domain for applying its beneficial AI, aligning with its core mission.
- Empowering Individual Scientists: Claude's design emphasizes making scientists more productive and scientific work more engaging by serving as a brainstorming partner and delegating complex or repetitive tasks.
- Holistic R&D Coverage: The AI addresses a broad range of tasks across the entire scientific pipeline, from early-stage discovery (e.g., molecule design) through experimental execution (e.g., protocol debugging) to data analysis and regulatory processes.
- Ecosystem Integration: Claude is being integrated with essential daily scientific tools and platforms such as Benchling, 10X Genomics, PubMed, and BioRender to enhance its utility within existing workflows.
- Claude 4.5 Capabilities: The latest model incorporates extensive scientific training, leading to improved performance in various scientific domains (especially computational biology) and a significant leap in handling long-horizon tasks and bioinformatics pipelines.
- Agentic Code for Biology: "Claude Code," while generically named, is a powerful agentic system already proving highly useful for bioinformatics, drafting scientific papers, literature reviews, and project organization for biologists.
- Overcoming Research Bottlenecks: AI can serve as a distilled source of scientific knowledge, offering prompt, practical solutions to common roadblocks like assay optimization, helping scientists get "unstuck" and accelerate discovery.
- Integrated Safety and Responsibility: Anthropic embeds biosecurity and responsible scaling policies directly into its development process for life science applications, ensuring that advancements are made safely and ethically without conflicting with commercial aims.
Frontier AI — Advanced artificial intelligence models that push the boundaries of current capabilities, often large language models (LLMs).
Assay — An investigative laboratory procedure for qualitatively assessing or quantitatively measuring the presence, amount, or functional activity of a target entity.
Sample Matrix — The components of a sample other than the analyte (target substance) of interest; these can interfere with analysis.
Inhibition — A process where a substance or condition blocks or reduces the activity of a biological process, enzyme, or chemical reaction.
R&D — Acronym for Research and Development, the process of investigating new scientific or technical knowledge and applying it to create new products or services.
Regulatory Process — The procedures and requirements set by government agencies (e.g., FDA) for the approval, manufacturing, and marketing of products like therapeutics and medical technologies.
Bioinformatics — The application of computational tools and methods to analyze large biological datasets, such as DNA, RNA, and protein sequences.
Bio-Foundation Models — Large AI models trained extensively on biological data (e.g., DNA, protein sequences, expression data) to develop specialized capabilities for biological modalities.
Agentic Code — Code or an AI system designed to perform a series of actions autonomously, often involving tool calls or complex workflows, to achieve a given goal.
It took us three months, ultimately. Lots of people working day and night in the lab to fix the problem. I posed this problem to Claude. I said, hey, what should we do to get it stuck? And just in one minute, you know, one response, Claude actually just one shot at the answer. Hi, I'm Jonah Kuhl. I'm the head of Life Sciences, Focus on Partnerships and Deployment here at Anthropic. Hi, I'm Eric Codder Abrams. I'm the head of Biology and Life Sciences here at Anthropic. I'm focused on research and product development. And together, we're trying to teach Claude to be a biologist. All right, Eric, let's talk science. I'm really excited about this and also excited about the fact that Anthropic and, you know, we're leaning into this space. And, you know, maybe the place to start is just thinking about why the Life Sciences, why Claude and what Anthropic brings to, you know, what is already a really big ecosystem, but one that's moving really fast. Yeah, I think that's a really important question. So I'll start with why are we focused on the Life Sciences. And this goes right to the heart of our mission. I think it's something that a lot of people may not realize, but when we talk about the beneficial use cases of AI and all the amazing things that we can do in the world with the Frontier AI that we're developing, actually the number one place that we at Anthropic are excited about applying it is within biology in the Life Sciences, right? If you read our foundational material, and you talk to people in the hallway here, that's the primary area where we're really focused on delivering the beneficial impact. For me, that's been a super exciting thing to come in and plug into is all of that pent-up energy and excitement to apply everything that we have to the space. And then starting to get more specific in talking about why Claude and how is our approach as in Frontier AI, you know, maybe different from some of the other approaches that are out there. I think there are two things that come to mind. You know, Jonah, you and I have talked a lot about this, but the first is that we're interested in building tools that empower individual scientists and enhance the experience of being a scientist going about your life, you know, doing all the work that you're doing, right? So we want to give people the same experience that software engineers have had of, you know, having a brainstorming partner to work with and delegate a task to throughout the process. We want to bring that to biologists in the lab and on the computational side. And so our additional focus is really about building tools that makes scientists more productive. And also makes science more fun. Yeah, right, take away some of the groundwork that, you know, everyone would rather get out of and allow you to focus more on the creative high leverage side. That's the first part. And then the second one is we're really focused not just on the really exciting early stage discovery problems, right, molecule design and protein folding in these, you know, incredibly impactful problems that many people in the field have focused on. But we want to address the whole spectrum from early stage discovery all the way through development and translation. And so for us, that means, you know, breaking it down into the whole world of different tasks that exist in the space. Everything from, you know, drafting and reviewing protocols and debugging them to performing bioinformatics analyses and writing up your results in slides and papers and that sort of a thing, right? There's a whole world of tasks out there that are important. And we're taking a holistic view and addressing all of them. Yeah, I think it's a really interesting time to think about science and maybe AI more generally. And you know, there's this inclination towards, you know, what problem will AI solve for me? But you know, the way that I think we're thinking about it the way that you just really nicely described in the way that in machines of loving graces maybe also thinking through that slightly orthogonal point, which is like, how do we change how we do science? And that will then impact, you know, solving, you know, structures and molecules and tissues and imaging and starting to like think through that world. So, you know, with that in mind, maybe then to transition. So, you know, we've seen the power of cloth, I think both of us have experienced this and like the delight and the joy in doing science with cloth. And it's current capabilities. But then also now starting to thank you in the research group that you're leading and how we advance those capabilities. And so maybe for a minute, we can just chat a little bit about how kind of current capabilities and ecosystems and maybe then like extended contacts through MCPs and life sciences start to create this base case. And then your thoughts on how we extend that and even like push it further. And you know, what that starts to look like and takes shape. Yeah, totally. So, you know, we've talked about this a lot. I think that, you know, it's important to crawl, walk, run in the space, right? It's, there's a lot of things about doing biology and having AI be useful in science that are different from how the AI be useful. It is the AI space. So maybe you like to use an old running analogy, you know, you start fast, you pick it up in the middle and you sprint home. So very little crawling, but just, you know, sprinting and then sprinting faster. sprinting sprinting faster than flying to rocket exactly. Right. What we're going for here. But the very base level is we need Claude to be conversant with all of the tools that scientists are using every day. Right. And so there's a whole ecosystem of important tools and partners out there that we are integrating with. So we talk about benching on the experiment administration, lab notebook, side of things, TEDx genomics with cell nature, right? An incredibly important platform for analyzing stable cell experiments and then PubMed, for example, for being able to query the literature, right? And so these are just three incredibly important partners in a much larger ecosystem. And so that base level is we need to make sure that Claude can talk to all the major sources that scientists are using throughout their daily work. And I think the next level is we want to bring Claude to performing at the level of a superhuman research assistant that can assist you as a scientist throughout all stages of your project, right? From the early stage hypothesis generation, when things are more creative and you're reviewing the literature and your brainstorming to the experiment execution phase, where you're drafting protocols and you're debugging things in the lab and even actually running those experiments in the lab to the computational and data and all of those things, right? When you're running your bioinformatic scripts, you're doing machine learning on top of that or some statistics and you're presenting the results to colleagues or for yourself, right? And so here, this is where we've broken down tasks into all of those areas, right? And we're figuring out, all right, how do we evaluate how well our models are doing those tasks and how do we rapidly improve performance in all of those areas. So we're making a big investment in doing that right now. And I think it's also important to say that we're not just doing this generically, right? Like in some ways, you can't speak of life sciences as one monolithic thing, right? There's all these different subfields within it. And we have a particular sequencing in mind where we like to think of it as being consisting of this core of important tasks that are shared throughout many different fields. And then within that, there are different subdomains that are really important, right? And we want to address all of it. But we're really focused on starting with that core that's going to be useful throughout the whole journey. You know, one thing that I'm really ecstatic about with our current partners and folks that are involved early on here to build this foundation, especially with MCPs, is that you mentioned 10X genomics, you mentioned PubMed, you mentioned groups like Benchling, and then Sage, BioNetworks, BioRender, you know, kind of going back to that last point, it really demonstrates and hopefully puts the action, the fact that it's not just solving a problem. But you know, in that group, you've got the literature, you've got instrumentation, you've got analytical workflows, you've got the cherry on top with that like perfect image or you know, network diagram and BioRender. And I expect that over the weeks, months to come like that whole ecosystem is just going to grow exponentially. And with that, like the power for more and more scientists. And I think that's just like incredibly cool and exciting. Yeah, I think that's a great point. Is that that's the experience that, you know, a lot of us have had on the software side, right? For me, I've always been on the one hand a part of the software world, the other head a part of this bio world. And you know, things started on the software side, where you'd give Claude, you know, these little snippets of tasks, right? And over time, those tasks become longer horizon, Claude becomes more autonomous, you know, can more seamlessly integrate through the different tools there. And I think we're right at that takeoff point in the life sciences where we're just now with all of these connections that we're introducing, able to unlock that next stage where, you know, you don't have to just ask Claude to go perform an analysis and then you do some work. And then you come back and you make a biorender fair and you ask Claude to revise it, right? We could actually give Claude a whole, you know, meaningful chunk of the work that would take a human scientist a couple hours to do. And I think that that transition is the really exciting point in a field where it goes from being, you know, a useful kind of utility to actually a brainstorming partner. Right? I'm just kind of like embedded in the process and you know, collaborate. Yeah. So we recently released Sun at 4.5, really exciting super powerful model. And I think one of the things that we've seen and, you know, eager to hear your perspective on the research side is just like seeing how that model performs in the context of different areas of science. And so, you know, what have you seen maybe in the like, the evolution and the power of those models and maybe some of the early e-vals or benchmarks that we've been seeing in different tasks that are relevant to scientists? So I think there's two things about Sun at 4.5 that I'm really excited about that have enhanced my own work by a great deal. The first is that it's our first model that's undergone extensive scientific trading. So, Sun at 4.5, you know, is skilled in many different domains of science. And I think one of the exciting things is that there's a lot there that generalizes, right? And so, Sun at 4.5 being better at math, you know, has some effective uplifting different capabilities in bio, especially in the computational side. And so, I think it's just really exciting that it's our first, you know, scientifically, you know, really, really capable model. Yeah. And, you know, there were some new things on the training side that went into making that possible that were just leaning into it, accelerating with all future models here. And the second thing is its ability to do long-hares in tasks, right? Consisting of long strings of different tool calls. So, this is something that, you know, for anyone that's done these sorts of long bioinformatics pipelines and things like that, is absolutely critical. And we saw a major jump up in those capabilities with Sun at 4.5, which, you know, makes it, you know, uniquely able to start to do these, like, really long bioinformatics workflows. Yeah, I think, and the analysis workflows and also the thing about how it applies to the different surfaces of Claude. So, you know, a lot of people think about Claude and they think, you know, the chat interface. But of course, having for many scientists, maybe some that do, maybe some that don't realize this, the power of agentic code enables like Claude Code or other places where that longer contacts and all that power becomes really interesting for data analysis, for integration, for kind of like reasoning over different types of knowledge. And it's an incredible, like, starting point, right? Where that, if we can start to build. Yeah, really is. And I know this is one of the things that you and I have been the most excited about, that Claude Code is amazingly useful as it is today in biology. And most people don't realize that, right? It's called Claude Code. It's not called Claude biology, right? But, you know, underneath the hood there, there's a really powerful general purpose agent that I, in particular, you know, many people that we've talked to throughout the community, have started to use in bioinformatics, even in things like drafting papers, right? And in performing literature reviews and organizing your projects, right? And so I think that's definitely something that we're going to be putting a lot more energy into. Yeah, I mean, there's those moments and as a technologist and someone that loves to develop technologies and apply them to biology, which is an affinity that I know we both share. You know, there are those moments where you see technologies or kind of like experience technologies and you just like really feel them. And I still have that moment of like uplift, you're remembering the first time, like, playing with Claude Code and making tasks that are either kind of beyond my technical capabilities, retractable and manageable, or the task of just, you know, like workflow running and execution that are just time intensive and cumbersome trivial, right? I mean, it just like puts those tools in sign just hands in a way that is incredibly powerful. It really is. And, you know, you mentioning those moments where you just viscerally feel, you know, the new capabilities that are out there, that reminds me of that moment for me, that really woke me up at the first time. And this was actually back in the Sonic 3, 5 days that, wow, these LOMs and these frontier models are really relevant for what we're doing, the life sciences. So for me that moment was, at the time, I was just running a biotech company that I had founded and I had this idea that I wanted to try to see, to see, hey, if I had had access to Claude when I was starting this company five years ago, how much time wouldn't have stayed thus? And how much heartache in trying to navigate some of these really difficult R&D problems where you try to solve what it have saved us? And I'll never forget this because the very first huge tactical roadblock that we ran into when we founded this company, there was a problem we were developing in assay, trying to detect in this case COVID. And it wasn't working. We were getting inhibited by the sample matrix and we couldn't figure it out, right? And it took us three months, ultimately, and lots of people working day and night in the lab to fix the problem. And I post this problem to Claude. I said, hey, we're trying to develop this assay and we're seeing that the sample is inhibiting things and what should we do to get it stuck? And just in one minute, one response, Claude actually just one shot at the answer. And I said, hey, I think you should add this much of this chemical into the mix. And that was a really eye-opening moment, right? That here, in conversing with Claude, you're kind of talking to a distilled version of the totality of scientific knowledge, right? And at the time, it was imperfect, but it's rapidly getting better. There's always this tension and I think scientists want perfection, right? It's something that we all kind of strive for and want that specificity. But for a lot of the work that holds science back, protocol optimization, an imperfect but helpful answer is the sort of thing that we go to, the most trusted colleagues where they might say, like, I don't know if, but this looks familiar, right? I've seen this problem at some point, that there's kind of like sage professors, they're the super sharp student down the hall. And again, it's not looking for perfection, but it's looking to get unstuck. It's looking to be helpful. It's looking to just keep you moving and towards discovery, which is what we're all looking for, right? Yeah, totally. Totally. I think the other area that really jumped out for me early on was, this is relevant later in the translation phase, is in the regulatory process, right? So I've seen a lot of time writing regulatory submissions, going through those processes with FDA and Claude is really capable there. And I think there's a huge opportunity, both on the industry side and on the FDA side, to recognize that we have these tools that can speed up the process on both sides and facilitate consistent standards across the board. And I'm really excited about pursuing that. I know people across this whole industry are as well. Okay, so let's stick on this for a minute. Within biology within AI, maybe even if you take a step back from AI and think about just like engineering and technology, the life sciences biology is this frequent substrate where people get really excited about the idea of biology or how it's just like one step away from being immediately programmable. So some of those ideas and intuitions, I think we probably agree with. But I think in many cases, it's folks that are maybe more in love with the idea biology as opposed to like really know what the life sciences, what regulatory frameworks look like. Let's talk a little bit more about your experience or collective experience as scientists and kind of like bringing some of that detailed knowledge to our partnerships, to our research efforts, and what those moments of looked like for you in the past and how they're maybe like reading on to priorities or approaches. Yeah, I think this is a really important, it's also pretty fun topic to talk about. In some ways, it's one of the oldest tropes in the space of the computer scientists, the physicist, the mathematician that kind of walks into biology and have all these romantic notions and then spend their first year in the lab and come out kind of shell-shocked in some ways. Disillusion of all the things that are possible, right? I think that where you and I both are coming from and the way that we're doing things here is we know what life of the lab is like and we want to solve the real problems that are the bottlenecks for this field, right? I will say that that's my own background of coming more of the computer side side and the math side of things and have picked up bio over the years in being in the lab. And I'm not dissolution. I really believe that we have the opportunity to massively uplift the capabilities of biologists in doing incredible impactful research and that with the tools that we have now, we're finally at that moment where all these things are possible. So I remain the same optimism that I had when I first got into this. I think all the experience in the lab has been really clarifying to help point out, okay, there are real problems here that are not pretty and that require lots of grindy work to get in there and disentangle. But I think we're now set up to make a dent in that. But I'd love to hear what you think. Yeah, I totally agree, right? I mean, I share that optimism. I do think that there are many people that don't always understand like how difficult science is but also how important just persistence and the fact that research and I think this probably applies down the clinical pipeline too. It's because it's so difficult, because there's so much knowledge that needs to be incorporated and every like step and debugging in the complexity of biology, whether it's that protocol optimization or data analysis, it's really hard to hold all that expertise definitely in one person, probably not even in one group and infrequently in one institution. And the result of that, and I think where again, we provide a really powerful technology in Claude and a research assistant collaborator is it starts to like bring more of that fluidity, right? It lowers the bar for computational analysis for folks that may not have that computer science background. It brings some molecular biology and optimization skills for folks that haven't spent their whole life cloning and doing molecular biology. And then it also just like helps make discoveries transferable across fields, right? I mean, I was not trained as a neuroscientist. I used to love to go to neuroscience lectures, but we then have to like come back and either like ask a whole bunch of naive questions, but you know, seeing optogenetics for the first time, you know, discovered in neuroscience took way too long to get out to cell biology to other domains. And I think the power of Claude and Claude as a life scientist is it starts to like address some of those core problems in biology, but also just starts to create that fluidity and start to break down walls in some of the parts that makes science hard. I totally agree. And the other thing I want to mention when we're talking about our outlook and our research roadmap and things like that is we focussed a lot on the meat and potatoes, eat the vegetables, have all of these practical tasks, right? That are really exciting, but more sort of surface level. I also want to call out that we're seeing an increasing trend in the field focusing on these bio-foundation models. These models that have savant-like capabilities on biological modalities, right? The D&A sequences, and protein sequences and being multimodal and expression data and all sorts of things. And a trend that's really interesting to watch is seeing increasing number of papers come out over time that are demonstrating that these things that previously looked like you needed to be specialized bio-models for. Maybe you don't. And maybe actually with really large frontier scale models like Claude with the right type of training we can start to develop those capabilities. And so I think we're all as a field at the beginning of just sort of working through that, but I think it's a really, really exciting trend to follow and that we'll be pursuing pretty aggressively, right? I think having these savant-like capabilities in these bio-modalities is really powerful for these specific bio-foundation models, but to really make that accessible to people, you need to be able to interface it with language, right? And so I wanted to call that out as one interesting. Yeah, I mean, it's a great point that, you know, as the field progresses here, and by field here, I mean, both, you know, the field of AI, but also like many domains of the life sciences, and I think we're already seeing a whole bunch of really exciting, you know, partners that are the AI native startups in the biotech space that are kind of taking some of these tools as well as large pharma partners. And in the way that the different pieces come together, right? So bio-foundation models, general intelligence models, specific data sets, you know, it's gonna be fascinating, right? And I think a really exciting time. And maybe this also gets to the point of partnership, right? So starting to take those different pieces and like bring them together and how we think about partnerships, maybe some of the early learnings or opportunities or partnerships that have been in front of mind for you or kind of a philosophy of like building this ecosystem. Yeah, so the way that I think about it is we know what our North Star is. We want to enable the amazing world that Dario writes about machines of loving grace in which, you know, R&D throughout the life sciences is accelerated by at least an order of magnitude. We want to make that happen as soon as possible. And within that framing, I think about partnerships as we need to make sure that all the right pieces exist, right? Some of those pieces we're gonna do ourselves, right? A lot on the model training side, some on the product side as well. But other pieces, you know, it makes sense for us to just find the right partners and make sure that we're supporting as much as we can. And so when I think about the different types of partners, there's really important ecosystem partners, right? Like I would call it Benchling is one of those for us, where I think they have the majority of working, you know, bioscientists are using Benchling is how they engage every day with kind of managing their experiments and their data. And so that's really, you know, an important one for us to lean into. And I think there's a lot of exciting things that we'll be able to share soon that we're working on together. So that's one type of a partner. Another type is a partner that we want to, which we work with in which they're using what we're building to actually do science in a way that wasn't possible before, right? Whether it's doing more science per unit time, right? Getting more impactful results per unit time and they could otherwise or making a type of discovery that wasn't possible before it, right? And so there, you know, there's a few partnerships that we're pretty excited about. I think one that's worth mentioning is with the Arc Institute. I know that you're thinking a lot about this as well. So we'd love to hear your thoughts. Yeah, I mean, I think the affinity that we both had towards Anthropic because of the unique features of the models. You know, there's like deep thinking. I don't think it's an accident actually that so many scientists have gravitated towards using Claude just, you know, naturally. But then also Darius vision and I think the vision that we very much believe in which is, you know, our goal is to accelerate, right? It's 100 years of science that is possible in 10. That's bold, it's ambitious. But also I think the more you think about it and the more you think about what holds science back, it's achievable. And so I agree that, you know, within the life sciences and biology, I think the other thing that's unique is that it's an ecosystem that is incredibly continuous and fluid, right? The student that is in a lab and finishing their thesis one day is the founder of an AI native startup, you know, the next that is then like acquired or working with or advancing, you know, major pipelines at, you know, AI forward, pharma company is like Lily. And you know, that fluidity and thinking about kind of that entire partnership in that ecosystem I think is that that's the beneficial deployment. It's all of science and achieving that. The thing that I'm really excited about and maybe one feature you didn't touch on is our AI for science program and this is really looking to put tools and clawed into the hands of scientists that have a bold idea or a big project and they think that clawed can, you know, be useful to solving that. And I think it's a great way to, you know, power early stage discovery research. It's a great way for us to kind of lean into those partners and work with them closely and learn from them and like start to just, you know, keep drawing the aperture open and understanding like in these early days, you know, what is working really well? And, you know, frankly, equally important like what isn't working well? And, you know, I think we both believe in the power but also believe in the current imperfection. And so that opportunity to, you know, work with scientists accelerate their research judge success based on, you know, what their success is. Their discovery, their acceleration, their time. And so to see where we're doing pretty well and maybe some areas where we just need to be doing a lot better. Yeah, I'm really excited about that too. And I think that's such an important point, right? Like in this conversation, we've been emphasizing a lot of breaking the problem down into all these pieces that we're going to solve independently. But the most important part is when we put it all back together. Yeah. And scientists are actually using these things, you know, how's it going and what are we doing, right? And so I think the AI for science program is critical for us to get that feedback and be closing the loop with people that are using these things every day in the lab. And so I am super excited about that. One other point that I think is really important to make that speaks to why Anthropic and why the experience of doing this within Anthropic is so exciting. It's such a perfect fit. Is that as we're talking about accelerating and enhancing capabilities, right? The other side of that is safety. And the tremendous responsibility that we all have to making sure that we are improving the models' capabilities and releasing increasingly impactful products in a way that is responsible and aligned with our responsible scaling policy and best practices in the biosecurity community. It's something I care deeply about. I've worked in biosecurity for years. And I think that at most companies, there would be some tension between the impact in the commercial aims of making these models better and in biology, right? And the safety and responsibility side of slowing down what we need to and making sure that we're being careful and have all the right safeguards in place. But at Anthropic, we don't have that tension, right? That's our DIA as a company. I think that's so valuable here. It's also really familiar to everyone of the life sciences, right? For people that are developing therapeutics and medical technologies, right? On the one hand, you have your product development arm and your commercial goals. And on the other hand, you have a quality management system, which is a set of procedures and practices that govern everything you do in order to make sure that you're doing so safely, right? And so I think it's just such a natural fit. Our approach here to AI of making sure that developing really powerful AI goes well and is done safely. And what needs to happen in the life sciences, right? And so that's something that I'm personally really excited about that I also think is a big part of who we are as a partner of this field. Right, yeah, it's an assumption, right? What we have to do that, we owe it to ourselves, we owe it to scientists, we owe it to the world, to take those sorts of questions really seriously. Yeah, and I think the other thing that I think about a lot is like at our core, at our DNA, we're a research organization. I don't think you can say that about all other AI companies, you know, Frontier Labs, et cetera. But I think being a research organization allows us to engage with researchers, labs, other research organizations in a way that really creates kind of a shared sense of like, ownership in goals and working together, right? Like we want to advance the technologies and see them put to, you know, the full purpose and power that are really invested in seeing that forward. Yeah, I think we're really lucky that that's the case. And, you know, I feel that very viscerally, that so many people on our founding team and our leadership team and just throughout all levels and teams in the organization are scientists, right? Many by training, many by nature and disposition. And, you know, I think that that, you can feel that sort of, you know, and all the work that we do, it makes it, you know, so natural to just go out and get to work with other scientists and all these things. Yeah. Some, you know, it's a little bit like, you know, the monkeys are running the zoo, right? Where we have people that are so passionate about science, driving the ship. And I think that means that we get to have a lot of fun. Yeah. But I think it also, it's a lot of fun, but also that appreciation for core questions like safety and understanding what the power is. And also core, you know, questions about like, what are the right problems to solve? And, you know, an appreciation for what makes science hard, what's low science down, you know, if we need to bake 100 years of progress in 10, well, you know, what does that actually look like? Right? And, you know, you can draw back the veil of science and, you know, there are some of those things of just understanding the literature, right? Like, you can spend all day every day. As a matter of fact, I think a lot of scientists would probably love to spend all day every day reading the literature, but like even then, you'd get through, you know, some small, tiny fraction of what was published or pre-printed at any given moment. So it's just impossible to keep up, right? But Claude can keep up. Yeah. Yeah. OK, so let's talk a little bit maybe here at the end about the future of life science work. And, you know, we've talked about biophomatics and coding. We've talked about some, you know, clinical work and different work that has been like demonstrated by some early partners. And then maybe also just ways that we're thinking about, like building this up and continuing to develop new partnerships, push the models towards greater capabilities. Where do you go when you start to think about the future? Yeah, so when we start to think about the future, you know, I think first, we need to make sure that Claude has all of the foundational knowledge that any scientist in the bio world would have, right? It's that things like understanding protein structural biology, right? And being able to look at a molecule from organic chemistry and understand its structured function, things like that, right? And so once you establish that base, then I think there's some really exciting places that we can go after that. Where one of the ones that I really like to talk about and that I think is critical is Claude actually learning to execute experiments in the lab, right? I think in order to get to this world where we're all going, that needs to happen. And again, this is a problem that for so long, you know, we've been making a lot of progress, maybe not as much as some had hoped for it, right? In terms of this vision of automating the tedious work of life in the lab. But I believe that it's possible now and I think that that's a really important area where we have to drill in and focus on. And I think, you know, just pause for a moment as to what life will be like when we get there. It'll be incredible, right? We'll be able to go from, you know, talking to Claude about an experiment to designing an experimental plan with Claude, to having Claude draft the protocols and, right, you can go back and forth on them and then when you're ready, you could say, all right, now go run those experiments and I'll review the data right in the morning. And so I think that's critical for closing the loop and enabling that acceleration that we're talking about. And the other thing that I think is a really important theme for our future research is in biology as with any domain in science, we have the opportunity to learn directly from real data from nature, right? And so on the one hand, we do a lot of model training and learning on annotations that are created by humans and other data sets that are either curated or created by humans, right? But there is an opportunity here to really do sort of lab in the loop, active learning from high throughput bio measurements. And the other reason why bio's such a good fit for that is we really, you know, are every year on a scaling law of the number of experiments, right? Per unit that we can do in terms of the throughput of these systems. So those are two themes that I'm increasingly excited about where, you know, when you start thinking about how do we move beyond human capabilities in these tasks, right? At some point we're going to saturate learning from human experts. The answer is to get the data from the lab. Yeah, I think this is a great theme. And the other thing that maybe I would point to is I think there's still this huge overhang, if you will, in terms of like current capabilities and use. And one of the things that sticks out to me is like starting to get clawed in the classroom in basic training, like really, you know, kind of implemented in a deep way such that, you know, many scientists are using clawed and also that experience and the product you know, starts to have this very cohesive feel where clawed is that virtual assistant and that virtual scientist is helping not answer a problem but, you know, answer a scientist answer any problem. All right, Eric, this has been awesome. I mean, it's always fun to talk science. Some say we've got a lot of work to do. So thanks for taking the time and really looking forward to the future of clawed life sciences and pushing towards the frontier. Yeah, thank you, Yoda. This has been a lot of fun. And we're just getting started. Yeah.
TL;DR
- Anthropic is dedicating significant resources to applying its Frontier AI, Claude, within the life sciences, aiming to accelerate R&D by an order of magnitude and deliver beneficial impact.
- Claude is being developed as a "superhuman research assistant" to empower individual scientists by streamlining tasks across the entire research spectrum, from hypothesis generation to regulatory submissions.
- Recent advancements, particularly in Claude 4.5, include extensive scientific training and enhanced capabilities for long-horizon tasks and complex bioinformatics workflows, showcasing its potential to make scientific work more efficient and enjoyable.
Takeaways
- Life Sciences as Primary Focus: Anthropic prioritizes the life sciences as the number one domain for applying its beneficial AI, aligning with its core mission.
- Empowering Individual Scientists: Claude's design emphasizes making scientists more productive and scientific work more engaging by serving as a brainstorming partner and delegating complex or repetitive tasks.
- Holistic R&D Coverage: The AI addresses a broad range of tasks across the entire scientific pipeline, from early-stage discovery (e.g., molecule design) through experimental execution (e.g., protocol debugging) to data analysis and regulatory processes.
- Ecosystem Integration: Claude is being integrated with essential daily scientific tools and platforms such as Benchling, 10X Genomics, PubMed, and BioRender to enhance its utility within existing workflows.
- Claude 4.5 Capabilities: The latest model incorporates extensive scientific training, leading to improved performance in various scientific domains (especially computational biology) and a significant leap in handling long-horizon tasks and bioinformatics pipelines.
- Agentic Code for Biology: "Claude Code," while generically named, is a powerful agentic system already proving highly useful for bioinformatics, drafting scientific papers, literature reviews, and project organization for biologists.
- Overcoming Research Bottlenecks: AI can serve as a distilled source of scientific knowledge, offering prompt, practical solutions to common roadblocks like assay optimization, helping scientists get "unstuck" and accelerate discovery.
- Integrated Safety and Responsibility: Anthropic embeds biosecurity and responsible scaling policies directly into its development process for life science applications, ensuring that advancements are made safely and ethically without conflicting with commercial aims.
Vocabulary
Frontier AI — Advanced artificial intelligence models that push the boundaries of current capabilities, often large language models (LLMs).
Assay — An investigative laboratory procedure for qualitatively assessing or quantitatively measuring the presence, amount, or functional activity of a target entity.
Sample Matrix — The components of a sample other than the analyte (target substance) of interest; these can interfere with analysis.
Inhibition — A process where a substance or condition blocks or reduces the activity of a biological process, enzyme, or chemical reaction.
R&D — Acronym for Research and Development, the process of investigating new scientific or technical knowledge and applying it to create new products or services.
Regulatory Process — The procedures and requirements set by government agencies (e.g., FDA) for the approval, manufacturing, and marketing of products like therapeutics and medical technologies.
Bioinformatics — The application of computational tools and methods to analyze large biological datasets, such as DNA, RNA, and protein sequences.
Bio-Foundation Models — Large AI models trained extensively on biological data (e.g., DNA, protein sequences, expression data) to develop specialized capabilities for biological modalities.
Agentic Code — Code or an AI system designed to perform a series of actions autonomously, often involving tool calls or complex workflows, to achieve a given goal.
Transcript
It took us three months, ultimately. Lots of people working day and night in the lab to fix the problem. I posed this problem to Claude. I said, hey, what should we do to get it stuck? And just in one minute, you know, one response, Claude actually just one shot at the answer. Hi, I'm Jonah Kuhl. I'm the head of Life Sciences, Focus on Partnerships and Deployment here at Anthropic. Hi, I'm Eric Codder Abrams. I'm the head of Biology and Life Sciences here at Anthropic. I'm focused on research and product development. And together, we're trying to teach Claude to be a biologist. All right, Eric, let's talk science. I'm really excited about this and also excited about the fact that Anthropic and, you know, we're leaning into this space. And, you know, maybe the place to start is just thinking about why the Life Sciences, why Claude and what Anthropic brings to, you know, what is already a really big ecosystem, but one that's moving really fast. Yeah, I think that's a really important question. So I'll start with why are we focused on the Life Sciences. And this goes right to the heart of our mission. I think it's something that a lot of people may not realize, but when we talk about the beneficial use cases of AI and all the amazing things that we can do in the world with the Frontier AI that we're developing, actually the number one place that we at Anthropic are excited about applying it is within biology in the Life Sciences, right? If you read our foundational material, and you talk to people in the hallway here, that's the primary area where we're really focused on delivering the beneficial impact. For me, that's been a super exciting thing to come in and plug into is all of that pent-up energy and excitement to apply everything that we have to the space. And then starting to get more specific in talking about why Claude and how is our approach as in Frontier AI, you know, maybe different from some of the other approaches that are out there. I think there are two things that come to mind. You know, Jonah, you and I have talked a lot about this, but the first is that we're interested in building tools that empower individual scientists and enhance the experience of being a scientist going about your life, you know, doing all the work that you're doing, right? So we want to give people the same experience that software engineers have had of, you know, having a brainstorming partner to work with and delegate a task to throughout the process. We want to bring that to biologists in the lab and on the computational side. And so our additional focus is really about building tools that makes scientists more productive. And also makes science more fun. Yeah, right, take away some of the groundwork that, you know, everyone would rather get out of and allow you to focus more on the creative high leverage side. That's the first part. And then the second one is we're really focused not just on the really exciting early stage discovery problems, right, molecule design and protein folding in these, you know, incredibly impactful problems that many people in the field have focused on. But we want to address the whole spectrum from early stage discovery all the way through development and translation. And so for us, that means, you know, breaking it down into the whole world of different tasks that exist in the space. Everything from, you know, drafting and reviewing protocols and debugging them to performing bioinformatics analyses and writing up your results in slides and papers and that sort of a thing, right? There's a whole world of tasks out there that are important. And we're taking a holistic view and addressing all of them. Yeah, I think it's a really interesting time to think about science and maybe AI more generally. And you know, there's this inclination towards, you know, what problem will AI solve for me? But you know, the way that I think we're thinking about it the way that you just really nicely described in the way that in machines of loving graces maybe also thinking through that slightly orthogonal point, which is like, how do we change how we do science? And that will then impact, you know, solving, you know, structures and molecules and tissues and imaging and starting to like think through that world. So, you know, with that in mind, maybe then to transition. So, you know, we've seen the power of cloth, I think both of us have experienced this and like the delight and the joy in doing science with cloth. And it's current capabilities. But then also now starting to thank you in the research group that you're leading and how we advance those capabilities. And so maybe for a minute, we can just chat a little bit about how kind of current capabilities and ecosystems and maybe then like extended contacts through MCPs and life sciences start to create this base case. And then your thoughts on how we extend that and even like push it further. And you know, what that starts to look like and takes shape. Yeah, totally. So, you know, we've talked about this a lot. I think that, you know, it's important to crawl, walk, run in the space, right? It's, there's a lot of things about doing biology and having AI be useful in science that are different from how the AI be useful. It is the AI space. So maybe you like to use an old running analogy, you know, you start fast, you pick it up in the middle and you sprint home. So very little crawling, but just, you know, sprinting and then sprinting faster. sprinting sprinting faster than flying to rocket exactly. Right. What we're going for here. But the very base level is we need Claude to be conversant with all of the tools that scientists are using every day. Right. And so there's a whole ecosystem of important tools and partners out there that we are integrating with. So we talk about benching on the experiment administration, lab notebook, side of things, TEDx genomics with cell nature, right? An incredibly important platform for analyzing stable cell experiments and then PubMed, for example, for being able to query the literature, right? And so these are just three incredibly important partners in a much larger ecosystem. And so that base level is we need to make sure that Claude can talk to all the major sources that scientists are using throughout their daily work. And I think the next level is we want to bring Claude to performing at the level of a superhuman research assistant that can assist you as a scientist throughout all stages of your project, right? From the early stage hypothesis generation, when things are more creative and you're reviewing the literature and your brainstorming to the experiment execution phase, where you're drafting protocols and you're debugging things in the lab and even actually running those experiments in the lab to the computational and data and all of those things, right? When you're running your bioinformatic scripts, you're doing machine learning on top of that or some statistics and you're presenting the results to colleagues or for yourself, right? And so here, this is where we've broken down tasks into all of those areas, right? And we're figuring out, all right, how do we evaluate how well our models are doing those tasks and how do we rapidly improve performance in all of those areas. So we're making a big investment in doing that right now. And I think it's also important to say that we're not just doing this generically, right? Like in some ways, you can't speak of life sciences as one monolithic thing, right? There's all these different subfields within it. And we have a particular sequencing in mind where we like to think of it as being consisting of this core of important tasks that are shared throughout many different fields. And then within that, there are different subdomains that are really important, right? And we want to address all of it. But we're really focused on starting with that core that's going to be useful throughout the whole journey. You know, one thing that I'm really ecstatic about with our current partners and folks that are involved early on here to build this foundation, especially with MCPs, is that you mentioned 10X genomics, you mentioned PubMed, you mentioned groups like Benchling, and then Sage, BioNetworks, BioRender, you know, kind of going back to that last point, it really demonstrates and hopefully puts the action, the fact that it's not just solving a problem. But you know, in that group, you've got the literature, you've got instrumentation, you've got analytical workflows, you've got the cherry on top with that like perfect image or you know, network diagram and BioRender. And I expect that over the weeks, months to come like that whole ecosystem is just going to grow exponentially. And with that, like the power for more and more scientists. And I think that's just like incredibly cool and exciting. Yeah, I think that's a great point. Is that that's the experience that, you know, a lot of us have had on the software side, right? For me, I've always been on the one hand a part of the software world, the other head a part of this bio world. And you know, things started on the software side, where you'd give Claude, you know, these little snippets of tasks, right? And over time, those tasks become longer horizon, Claude becomes more autonomous, you know, can more seamlessly integrate through the different tools there. And I think we're right at that takeoff point in the life sciences where we're just now with all of these connections that we're introducing, able to unlock that next stage where, you know, you don't have to just ask Claude to go perform an analysis and then you do some work. And then you come back and you make a biorender fair and you ask Claude to revise it, right? We could actually give Claude a whole, you know, meaningful chunk of the work that would take a human scientist a couple hours to do. And I think that that transition is the really exciting point in a field where it goes from being, you know, a useful kind of utility to actually a brainstorming partner. Right? I'm just kind of like embedded in the process and you know, collaborate. Yeah. So we recently released Sun at 4.5, really exciting super powerful model. And I think one of the things that we've seen and, you know, eager to hear your perspective on the research side is just like seeing how that model performs in the context of different areas of science. And so, you know, what have you seen maybe in the like, the evolution and the power of those models and maybe some of the early e-vals or benchmarks that we've been seeing in different tasks that are relevant to scientists? So I think there's two things about Sun at 4.5 that I'm really excited about that have enhanced my own work by a great deal. The first is that it's our first model that's undergone extensive scientific trading. So, Sun at 4.5, you know, is skilled in many different domains of science. And I think one of the exciting things is that there's a lot there that generalizes, right? And so, Sun at 4.5 being better at math, you know, has some effective uplifting different capabilities in bio, especially in the computational side. And so, I think it's just really exciting that it's our first, you know, scientifically, you know, really, really capable model. Yeah. And, you know, there were some new things on the training side that went into making that possible that were just leaning into it, accelerating with all future models here. And the second thing is its ability to do long-hares in tasks, right? Consisting of long strings of different tool calls. So, this is something that, you know, for anyone that's done these sorts of long bioinformatics pipelines and things like that, is absolutely critical. And we saw a major jump up in those capabilities with Sun at 4.5, which, you know, makes it, you know, uniquely able to start to do these, like, really long bioinformatics workflows. Yeah, I think, and the analysis workflows and also the thing about how it applies to the different surfaces of Claude. So, you know, a lot of people think about Claude and they think, you know, the chat interface. But of course, having for many scientists, maybe some that do, maybe some that don't realize this, the power of agentic code enables like Claude Code or other places where that longer contacts and all that power becomes really interesting for data analysis, for integration, for kind of like reasoning over different types of knowledge. And it's an incredible, like, starting point, right? Where that, if we can start to build. Yeah, really is. And I know this is one of the things that you and I have been the most excited about, that Claude Code is amazingly useful as it is today in biology. And most people don't realize that, right? It's called Claude Code. It's not called Claude biology, right? But, you know, underneath the hood there, there's a really powerful general purpose agent that I, in particular, you know, many people that we've talked to throughout the community, have started to use in bioinformatics, even in things like drafting papers, right? And in performing literature reviews and organizing your projects, right? And so I think that's definitely something that we're going to be putting a lot more energy into. Yeah, I mean, there's those moments and as a technologist and someone that loves to develop technologies and apply them to biology, which is an affinity that I know we both share. You know, there are those moments where you see technologies or kind of like experience technologies and you just like really feel them. And I still have that moment of like uplift, you're remembering the first time, like, playing with Claude Code and making tasks that are either kind of beyond my technical capabilities, retractable and manageable, or the task of just, you know, like workflow running and execution that are just time intensive and cumbersome trivial, right? I mean, it just like puts those tools in sign just hands in a way that is incredibly powerful. It really is. And, you know, you mentioning those moments where you just viscerally feel, you know, the new capabilities that are out there, that reminds me of that moment for me, that really woke me up at the first time. And this was actually back in the Sonic 3, 5 days that, wow, these LOMs and these frontier models are really relevant for what we're doing, the life sciences. So for me that moment was, at the time, I was just running a biotech company that I had founded and I had this idea that I wanted to try to see, to see, hey, if I had had access to Claude when I was starting this company five years ago, how much time wouldn't have stayed thus? And how much heartache in trying to navigate some of these really difficult R&D problems where you try to solve what it have saved us? And I'll never forget this because the very first huge tactical roadblock that we ran into when we founded this company, there was a problem we were developing in assay, trying to detect in this case COVID. And it wasn't working. We were getting inhibited by the sample matrix and we couldn't figure it out, right? And it took us three months, ultimately, and lots of people working day and night in the lab to fix the problem. And I post this problem to Claude. I said, hey, we're trying to develop this assay and we're seeing that the sample is inhibiting things and what should we do to get it stuck? And just in one minute, one response, Claude actually just one shot at the answer. And I said, hey, I think you should add this much of this chemical into the mix. And that was a really eye-opening moment, right? That here, in conversing with Claude, you're kind of talking to a distilled version of the totality of scientific knowledge, right? And at the time, it was imperfect, but it's rapidly getting better. There's always this tension and I think scientists want perfection, right? It's something that we all kind of strive for and want that specificity. But for a lot of the work that holds science back, protocol optimization, an imperfect but helpful answer is the sort of thing that we go to, the most trusted colleagues where they might say, like, I don't know if, but this looks familiar, right? I've seen this problem at some point, that there's kind of like sage professors, they're the super sharp student down the hall. And again, it's not looking for perfection, but it's looking to get unstuck. It's looking to be helpful. It's looking to just keep you moving and towards discovery, which is what we're all looking for, right? Yeah, totally. Totally. I think the other area that really jumped out for me early on was, this is relevant later in the translation phase, is in the regulatory process, right? So I've seen a lot of time writing regulatory submissions, going through those processes with FDA and Claude is really capable there. And I think there's a huge opportunity, both on the industry side and on the FDA side, to recognize that we have these tools that can speed up the process on both sides and facilitate consistent standards across the board. And I'm really excited about pursuing that. I know people across this whole industry are as well. Okay, so let's stick on this for a minute. Within biology within AI, maybe even if you take a step back from AI and think about just like engineering and technology, the life sciences biology is this frequent substrate where people get really excited about the idea of biology or how it's just like one step away from being immediately programmable. So some of those ideas and intuitions, I think we probably agree with. But I think in many cases, it's folks that are maybe more in love with the idea biology as opposed to like really know what the life sciences, what regulatory frameworks look like. Let's talk a little bit more about your experience or collective experience as scientists and kind of like bringing some of that detailed knowledge to our partnerships, to our research efforts, and what those moments of looked like for you in the past and how they're maybe like reading on to priorities or approaches. Yeah, I think this is a really important, it's also pretty fun topic to talk about. In some ways, it's one of the oldest tropes in the space of the computer scientists, the physicist, the mathematician that kind of walks into biology and have all these romantic notions and then spend their first year in the lab and come out kind of shell-shocked in some ways. Disillusion of all the things that are possible, right? I think that where you and I both are coming from and the way that we're doing things here is we know what life of the lab is like and we want to solve the real problems that are the bottlenecks for this field, right? I will say that that's my own background of coming more of the computer side side and the math side of things and have picked up bio over the years in being in the lab. And I'm not dissolution. I really believe that we have the opportunity to massively uplift the capabilities of biologists in doing incredible impactful research and that with the tools that we have now, we're finally at that moment where all these things are possible. So I remain the same optimism that I had when I first got into this. I think all the experience in the lab has been really clarifying to help point out, okay, there are real problems here that are not pretty and that require lots of grindy work to get in there and disentangle. But I think we're now set up to make a dent in that. But I'd love to hear what you think. Yeah, I totally agree, right? I mean, I share that optimism. I do think that there are many people that don't always understand like how difficult science is but also how important just persistence and the fact that research and I think this probably applies down the clinical pipeline too. It's because it's so difficult, because there's so much knowledge that needs to be incorporated and every like step and debugging in the complexity of biology, whether it's that protocol optimization or data analysis, it's really hard to hold all that expertise definitely in one person, probably not even in one group and infrequently in one institution. And the result of that, and I think where again, we provide a really powerful technology in Claude and a research assistant collaborator is it starts to like bring more of that fluidity, right? It lowers the bar for computational analysis for folks that may not have that computer science background. It brings some molecular biology and optimization skills for folks that haven't spent their whole life cloning and doing molecular biology. And then it also just like helps make discoveries transferable across fields, right? I mean, I was not trained as a neuroscientist. I used to love to go to neuroscience lectures, but we then have to like come back and either like ask a whole bunch of naive questions, but you know, seeing optogenetics for the first time, you know, discovered in neuroscience took way too long to get out to cell biology to other domains. And I think the power of Claude and Claude as a life scientist is it starts to like address some of those core problems in biology, but also just starts to create that fluidity and start to break down walls in some of the parts that makes science hard. I totally agree. And the other thing I want to mention when we're talking about our outlook and our research roadmap and things like that is we focussed a lot on the meat and potatoes, eat the vegetables, have all of these practical tasks, right? That are really exciting, but more sort of surface level. I also want to call out that we're seeing an increasing trend in the field focusing on these bio-foundation models. These models that have savant-like capabilities on biological modalities, right? The D&A sequences, and protein sequences and being multimodal and expression data and all sorts of things. And a trend that's really interesting to watch is seeing increasing number of papers come out over time that are demonstrating that these things that previously looked like you needed to be specialized bio-models for. Maybe you don't. And maybe actually with really large frontier scale models like Claude with the right type of training we can start to develop those capabilities. And so I think we're all as a field at the beginning of just sort of working through that, but I think it's a really, really exciting trend to follow and that we'll be pursuing pretty aggressively, right? I think having these savant-like capabilities in these bio-modalities is really powerful for these specific bio-foundation models, but to really make that accessible to people, you need to be able to interface it with language, right? And so I wanted to call that out as one interesting. Yeah, I mean, it's a great point that, you know, as the field progresses here, and by field here, I mean, both, you know, the field of AI, but also like many domains of the life sciences, and I think we're already seeing a whole bunch of really exciting, you know, partners that are the AI native startups in the biotech space that are kind of taking some of these tools as well as large pharma partners. And in the way that the different pieces come together, right? So bio-foundation models, general intelligence models, specific data sets, you know, it's gonna be fascinating, right? And I think a really exciting time. And maybe this also gets to the point of partnership, right? So starting to take those different pieces and like bring them together and how we think about partnerships, maybe some of the early learnings or opportunities or partnerships that have been in front of mind for you or kind of a philosophy of like building this ecosystem. Yeah, so the way that I think about it is we know what our North Star is. We want to enable the amazing world that Dario writes about machines of loving grace in which, you know, R&D throughout the life sciences is accelerated by at least an order of magnitude. We want to make that happen as soon as possible. And within that framing, I think about partnerships as we need to make sure that all the right pieces exist, right? Some of those pieces we're gonna do ourselves, right? A lot on the model training side, some on the product side as well. But other pieces, you know, it makes sense for us to just find the right partners and make sure that we're supporting as much as we can. And so when I think about the different types of partners, there's really important ecosystem partners, right? Like I would call it Benchling is one of those for us, where I think they have the majority of working, you know, bioscientists are using Benchling is how they engage every day with kind of managing their experiments and their data. And so that's really, you know, an important one for us to lean into. And I think there's a lot of exciting things that we'll be able to share soon that we're working on together. So that's one type of a partner. Another type is a partner that we want to, which we work with in which they're using what we're building to actually do science in a way that wasn't possible before, right? Whether it's doing more science per unit time, right? Getting more impactful results per unit time and they could otherwise or making a type of discovery that wasn't possible before it, right? And so there, you know, there's a few partnerships that we're pretty excited about. I think one that's worth mentioning is with the Arc Institute. I know that you're thinking a lot about this as well. So we'd love to hear your thoughts. Yeah, I mean, I think the affinity that we both had towards Anthropic because of the unique features of the models. You know, there's like deep thinking. I don't think it's an accident actually that so many scientists have gravitated towards using Claude just, you know, naturally. But then also Darius vision and I think the vision that we very much believe in which is, you know, our goal is to accelerate, right? It's 100 years of science that is possible in 10. That's bold, it's ambitious. But also I think the more you think about it and the more you think about what holds science back, it's achievable. And so I agree that, you know, within the life sciences and biology, I think the other thing that's unique is that it's an ecosystem that is incredibly continuous and fluid, right? The student that is in a lab and finishing their thesis one day is the founder of an AI native startup, you know, the next that is then like acquired or working with or advancing, you know, major pipelines at, you know, AI forward, pharma company is like Lily. And you know, that fluidity and thinking about kind of that entire partnership in that ecosystem I think is that that's the beneficial deployment. It's all of science and achieving that. The thing that I'm really excited about and maybe one feature you didn't touch on is our AI for science program and this is really looking to put tools and clawed into the hands of scientists that have a bold idea or a big project and they think that clawed can, you know, be useful to solving that. And I think it's a great way to, you know, power early stage discovery research. It's a great way for us to kind of lean into those partners and work with them closely and learn from them and like start to just, you know, keep drawing the aperture open and understanding like in these early days, you know, what is working really well? And, you know, frankly, equally important like what isn't working well? And, you know, I think we both believe in the power but also believe in the current imperfection. And so that opportunity to, you know, work with scientists accelerate their research judge success based on, you know, what their success is. Their discovery, their acceleration, their time. And so to see where we're doing pretty well and maybe some areas where we just need to be doing a lot better. Yeah, I'm really excited about that too. And I think that's such an important point, right? Like in this conversation, we've been emphasizing a lot of breaking the problem down into all these pieces that we're going to solve independently. But the most important part is when we put it all back together. Yeah. And scientists are actually using these things, you know, how's it going and what are we doing, right? And so I think the AI for science program is critical for us to get that feedback and be closing the loop with people that are using these things every day in the lab. And so I am super excited about that. One other point that I think is really important to make that speaks to why Anthropic and why the experience of doing this within Anthropic is so exciting. It's such a perfect fit. Is that as we're talking about accelerating and enhancing capabilities, right? The other side of that is safety. And the tremendous responsibility that we all have to making sure that we are improving the models' capabilities and releasing increasingly impactful products in a way that is responsible and aligned with our responsible scaling policy and best practices in the biosecurity community. It's something I care deeply about. I've worked in biosecurity for years. And I think that at most companies, there would be some tension between the impact in the commercial aims of making these models better and in biology, right? And the safety and responsibility side of slowing down what we need to and making sure that we're being careful and have all the right safeguards in place. But at Anthropic, we don't have that tension, right? That's our DIA as a company. I think that's so valuable here. It's also really familiar to everyone of the life sciences, right? For people that are developing therapeutics and medical technologies, right? On the one hand, you have your product development arm and your commercial goals. And on the other hand, you have a quality management system, which is a set of procedures and practices that govern everything you do in order to make sure that you're doing so safely, right? And so I think it's just such a natural fit. Our approach here to AI of making sure that developing really powerful AI goes well and is done safely. And what needs to happen in the life sciences, right? And so that's something that I'm personally really excited about that I also think is a big part of who we are as a partner of this field. Right, yeah, it's an assumption, right? What we have to do that, we owe it to ourselves, we owe it to scientists, we owe it to the world, to take those sorts of questions really seriously. Yeah, and I think the other thing that I think about a lot is like at our core, at our DNA, we're a research organization. I don't think you can say that about all other AI companies, you know, Frontier Labs, et cetera. But I think being a research organization allows us to engage with researchers, labs, other research organizations in a way that really creates kind of a shared sense of like, ownership in goals and working together, right? Like we want to advance the technologies and see them put to, you know, the full purpose and power that are really invested in seeing that forward. Yeah, I think we're really lucky that that's the case. And, you know, I feel that very viscerally, that so many people on our founding team and our leadership team and just throughout all levels and teams in the organization are scientists, right? Many by training, many by nature and disposition. And, you know, I think that that, you can feel that sort of, you know, and all the work that we do, it makes it, you know, so natural to just go out and get to work with other scientists and all these things. Yeah. Some, you know, it's a little bit like, you know, the monkeys are running the zoo, right? Where we have people that are so passionate about science, driving the ship. And I think that means that we get to have a lot of fun. Yeah. But I think it also, it's a lot of fun, but also that appreciation for core questions like safety and understanding what the power is. And also core, you know, questions about like, what are the right problems to solve? And, you know, an appreciation for what makes science hard, what's low science down, you know, if we need to bake 100 years of progress in 10, well, you know, what does that actually look like? Right? And, you know, you can draw back the veil of science and, you know, there are some of those things of just understanding the literature, right? Like, you can spend all day every day. As a matter of fact, I think a lot of scientists would probably love to spend all day every day reading the literature, but like even then, you'd get through, you know, some small, tiny fraction of what was published or pre-printed at any given moment. So it's just impossible to keep up, right? But Claude can keep up. Yeah. Yeah. OK, so let's talk a little bit maybe here at the end about the future of life science work. And, you know, we've talked about biophomatics and coding. We've talked about some, you know, clinical work and different work that has been like demonstrated by some early partners. And then maybe also just ways that we're thinking about, like building this up and continuing to develop new partnerships, push the models towards greater capabilities. Where do you go when you start to think about the future? Yeah, so when we start to think about the future, you know, I think first, we need to make sure that Claude has all of the foundational knowledge that any scientist in the bio world would have, right? It's that things like understanding protein structural biology, right? And being able to look at a molecule from organic chemistry and understand its structured function, things like that, right? And so once you establish that base, then I think there's some really exciting places that we can go after that. Where one of the ones that I really like to talk about and that I think is critical is Claude actually learning to execute experiments in the lab, right? I think in order to get to this world where we're all going, that needs to happen. And again, this is a problem that for so long, you know, we've been making a lot of progress, maybe not as much as some had hoped for it, right? In terms of this vision of automating the tedious work of life in the lab. But I believe that it's possible now and I think that that's a really important area where we have to drill in and focus on. And I think, you know, just pause for a moment as to what life will be like when we get there. It'll be incredible, right? We'll be able to go from, you know, talking to Claude about an experiment to designing an experimental plan with Claude, to having Claude draft the protocols and, right, you can go back and forth on them and then when you're ready, you could say, all right, now go run those experiments and I'll review the data right in the morning. And so I think that's critical for closing the loop and enabling that acceleration that we're talking about. And the other thing that I think is a really important theme for our future research is in biology as with any domain in science, we have the opportunity to learn directly from real data from nature, right? And so on the one hand, we do a lot of model training and learning on annotations that are created by humans and other data sets that are either curated or created by humans, right? But there is an opportunity here to really do sort of lab in the loop, active learning from high throughput bio measurements. And the other reason why bio's such a good fit for that is we really, you know, are every year on a scaling law of the number of experiments, right? Per unit that we can do in terms of the throughput of these systems. So those are two themes that I'm increasingly excited about where, you know, when you start thinking about how do we move beyond human capabilities in these tasks, right? At some point we're going to saturate learning from human experts. The answer is to get the data from the lab. Yeah, I think this is a great theme. And the other thing that maybe I would point to is I think there's still this huge overhang, if you will, in terms of like current capabilities and use. And one of the things that sticks out to me is like starting to get clawed in the classroom in basic training, like really, you know, kind of implemented in a deep way such that, you know, many scientists are using clawed and also that experience and the product you know, starts to have this very cohesive feel where clawed is that virtual assistant and that virtual scientist is helping not answer a problem but, you know, answer a scientist answer any problem. All right, Eric, this has been awesome. I mean, it's always fun to talk science. Some say we've got a lot of work to do. So thanks for taking the time and really looking forward to the future of clawed life sciences and pushing towards the frontier. Yeah, thank you, Yoda. This has been a lot of fun. And we're just getting started. Yeah.