Tomorrow's Medicine, Ep. 16: Insights by ClawBio Founder - Dr. Manuel Corpas

With Dr. Manuel Corpas, Genomics Researcher & ClawBio Founder. Hosted by Arseniy Arsentyev. Published July 22, 2026.

https://aarsentyev.com/podcast/ep16-manuel-corpas

Machine-transcribed and lightly edited: filler words and false starts are removed, and the wording is otherwise as spoken. Where the two differ, the recording is the record. Times are minutes into the recording. Not medical advice.

[0:00] Dr. Manuel Corpas: We are at agentic speed, AI speed. One week is one month, one month is one year. I've been to talks with Anthropic. My agent is not a server, my agent is a colleague. So the bottleneck is not intelligence anymore. I think what is our goal is to go through this new period of enlightenment.

[0:22] Arseniy Arsentyev: It's the 16th episode of Tomorrow's Medicine. And today I'm talking to the person who invented the term agentic genomics. It is Dr. Manuel Corpas. He is the founder of ClawBio and the person on the front line of agentic engineering in biotechnology and bioinformatics. I would not pause any further. And let's proceed to our conversation. Could you outline to us how you started? And what were bioinformatics feel like when you were in college?

[0:57] Dr. Manuel Corpas: Oh, my God. It makes me feel very old now. So it was 2000 when I started doing bioinformatics. So I finished my MSc and my undergrad in biology. And I really wanted to go into computer science. So I found at the time an MSc in Bioinformatics from the University of Manchester here in the UK. It was perfect because on the one hand, I would make use of my biology background, but then I would get into the more sort of computational side of things. So year 2000, it was right in the middle of the Human Genome Project.

[1:42] And there was a clear acceleration and a sort of revolutionizing of how biology was being carried out. But it was the transition between the moment when it was mostly lab-based, where you would use, you perform your analysis in one gene, into then the more high throughput, where now you start talking about many genes, even the whole sort of genome. And so that was the time when really biology started to become a data-rich, data-intensive science. And so I then continued doing a PhD. And so my PhD was part of what is now called a database called InterPro. So you may or may not have heard, is one of the European Bioinformatics Institute databases whose objective is to basically classify protein sequences according to their motifs, to their domains.

[2:45] And so by looking at how the conservation of these specific features are across the whole spectrum of existing sequences developed, then you can evolutionarily classify them. So at that point was already when I started getting into machine learning, artificial intelligence, support vector machines. So technologies which now have been superseded dramatically with the development of neural network or, well, neural networks were already developed then. But then when they really exploded was around 2009 when we started with a backpropagation algorithm and the deep networks, which allowed the abstraction of much more complex concepts into the mathematics of neural networks. So after my PhD, I was hired by the Wellcome Sanger Institute, which is the flagship genomics institute on this side of the Atlantic.

[3:53] And I was employed to develop a database called DECIPHER, whose objective was the analysis of patients with rare genomics diseases worldwide. So this effort still exists. I think now I have like around, I don't know, last time I looked a year ago was like 60,000 patients with rare genomic disorders who subjected, who have been consented for public research. And this infrastructure and this postdoc work sort of changed my focus from the protein sequences into genomics. So I've been doing genomics since 2008.

[4:38] And at that time, I was quite piqued by the fact that it was incredibly difficult, even for clinicians who have been working at that time with this kind of data to really help them understand all of the intricacies around the interpretation of data coming from their patient's genome. And it was also the time when sequencing started to become accessible. And in fact, in 2010, I actually sequenced my whole exome. So I was one of the first people who sequenced their genome. And that led to the sequencing of the genomes of my family, which then we published online.

[5:24] Many people thought it was crazy, but we just decided to sort of put it outside in the public domain so that anyone, everyone who wanted to use those data for the development of tools, simply for research purposes. And I gave the condition that anyone can use it. But if you want it, I'd be very happy if you found any kind of resource about my family, which could help us inform the understanding and really push the boundaries in terms of how much your genome can tell about, future life decisions and life choices. And since that time, I've already been obsessed with that same question, which is understanding how the genome can help inform life decisions for a better quality of life. And I don't think really we have made a lot of progress.

[6:21] I think that's one of the main bottlenecks in terms of medicine of the future. The fact that we don't have enough data or at the levels required to be able to sort of generate the sufficient training that would be required for sophisticated models to be able to really represent the richness of biological complexity. I'm hoping that this may change in the near future as more data and more biobanks are made accessible for research. But it's not, the fact that clinical data, medical data is so unique, so sensitive, is already a big part that is sort of, I guess, hampering the faster advances that we have seen in other fields such as fintech or law or even crypto.

[7:49] So I felt when I started getting into agentic genomics, which was like last year, it was sort of colonized by crypto people. So anyway, so I've been a senior lecturer in genomics for about five years now at the University of Westminster here in London. And that has really conditioned my exposure to being really, I guess, up to date in terms of the AI developments. I think London is really a thriving place for AI. There's always something happening. So I've had, and I also have a critical mass, with many universities, many research institutes, Google DeepMind is here and many other startups. So that has really conditioned where I am today.

[8:46] Arseniy Arsentyev: And you were working on DECIPHER before this agentic boom in AI, and you are now the person on the frontier of agentic science. How do you think if these years ago you had the tools that you have now, would it speed up the process consistently?

[9:04] Dr. Manuel Corpas: Yeah. But the thing is that I don't think there's a real understanding of the type of usages that agentic capabilities are able to provide. And so right now I feel like in heaven. I am really, I have all these data sets that I've been accumulating for a long time. Like the ones from my own family and so on, which was quite rich in terms of the amount of data that we had. All of the other projects that I've been working on. But I don't think most people who are doing genomics really have a deep grasp of where this can take us. And so I've been frantic, especially since the beginning of, I would say, I guess for the past two or three years. I knew.

[10:00] So one of the things I can say is that I started a program on AI and digital health at my own university, directing a master's degree on that subject. And so that, together with my fellowship at the Alan Turing Institute, which is a national capability sort of institution for AI here in the UK. So that sort of helped me get up to speed in terms of making sure that I wasn't just going to be a passenger. I knew I was going to drive the process forward. I really want to, particularly on my domain expertise, which is the application of these technologies, particularly around the genomics and health equity aspects.

[10:56] And so I would say, the thing is that, to be honest, for me, what was really changing, apart from the fact that I have the capacity to develop things at a speed, which is maybe 10 times faster than it was before. So I sort of have built a, probably an army of agents around me that that helped me with the automation of the many tasks that I've been sort of confronted on a daily basis. And so, for me, making sure that I kept up to speed, hasn't mean that I actually kept with what the developments have been for the field. Because every week there's a new model, there's a new harness, there's a new tool, there's a new, and it's absolutely impossible. And I think everyone feels that way.

[12:04] And in a way, that sort of being on my own is what actually motivated the development of ClawBio, which is the open source bio domain library of skills for agentic AI. Right? And what a lot of people don't see about what I've been doing is that it hasn't, it's not actually as much as of course, it is about the technical developments and the capacity and the capabilities and the fact that we have a set of tools, which are, and a platform that potentially is useful for many people. But I think it's also this community aspect, the fact that I can't keep up to date with everything.

[12:55] But if we have now a community of people who are experts, who care about, how the sort of agentic revolution is going to be invented and developed moving forwards within the bioinformatics slash omics realm, that's the best possible way of really surviving the tsunami we are all in.

[13:24] Arseniy Arsentyev: And I feel it's, first of all, thank you for the comprehensive overview. And I feel it's a great moment to clarify the terms, because maybe someone in the audience is not familiar. How would you describe agentic genomics? People probably understand for now what agentic means. And some people understand what genomics means. Sure, not all. And what are this combination of words, in your opinion, should convey to people? And the second sub-question, when was this time, this light bubble, when you understood that agentic genomics might be a thing and you started pursuing it?

[14:07] Dr. Manuel Corpas: So, agentic genomics is the use of autonomous AI agents, powered by a large language model, and operating within a domain-constrained skill library, so that they can discover, plan, execute, and iteratively refine multi-step genomic analysis, where the agent exercises runtime decision-making over tool selection, parameterization, and error handling, as well as output evaluation. So, that would be the sort of formal definition. But if I put it in simpler terms, is this idea that a harness, that you may have, Claude Code or Codex, or any of the current sort of layers on top of the LLM that allows you to autonomously and sort of, in using natural language, typing English, basically, on the command line, is going to be able to interpret your intent.

[15:21] And it's going to basically solve a task without you having to actually do any other programming. That's really what this is. If I was to put it in a simpler term even, sometimes it's a little bit like Alexa or Siri, right? You tell something to the robot in English, and then it does something and comes back to you. Except that Alexa is kind of poor because it's only, it's just one back and forth, whereas the agent loops itself. So, it's going to decompose your intent into a number of tasks that is able to sort of ascertain, based on the context and the training that it has. And then it can execute complex workflows in an automated way. It can sort of trace back steps if there's an error.

[16:24] And so, it can do a lot of the things that a developer could do before, but now in an automated way.

[16:34] Arseniy Arsentyev: And when was the time that you understood that it's possible that current LLMs are on the level to fulfill the tasks you need? Was it after GPT was released?

[16:46] Dr. Manuel Corpas: No. It happened really when I got into Claude Code, which was last Christmas.

[16:53] Arseniy Arsentyev: A year ago.

[16:54] Dr. Manuel Corpas: Okay. It happened a number of inflection points. The last one was really last Christmas when suddenly, and it wasn't just me. I think it was for many of the top leading AI voices where there was a little bit of time, holiday time, and then people started tinkering about this thing that allowed it to sort of work with powerful models. So, you had to have what we call the harness, which is the layer on top that goes on the LLM, which allows the sort of loop aspects and then the sort of internal mechanisms for the robot to act in a series of workflow actions.

[17:44] So, you needed to have the harness, in this case, it could be Claude Code, but then you also needed to have the model to be powerful enough to allow the sort of reasoning behind for it to happen so that it could execute these complex tasks and understand the intent in a sophisticated way. And that's only happened. And the earliest people who would have worked on that would have been around November, which was, in this case, for instance, I can think of Peter Steinberger, who developed, I think it was called Clawd.

[18:32] It had a number of names and it ended up being OpenClaw, which was basically the pipeline to help you interact with the LLM through different channels. So, you could now then talk to your agents via Telegram, via WhatsApp. And so, suddenly it became something that you could be running 24-7 and it would, you could add its own soul. So, it would have its own personality. And so, at that point, around, especially from November, December, January, that was when, in my opinion, when we've had the last sort of big inflection point. And that was when Anthropic stopped being a sort of underdog. And then, suddenly, that, everyone then realized, wow, Opus, the things it can do, and it can do it automatically. And it's going to do this super complex reasoning.

[19:47] So, that was the time when, and it's kind of really incredible. It's not even happening in a year. And it's transformed the way I work. And it's going to transform, it's transforming the way any digital work is happening. And it was very clear by then that this thing, could and should be applied to bio. And so, I went to, I presented this for the first time in late February. I presented ClawBio as a library for reusable skills. And by skills, like self-contained context that agents can discover and reuse for performing these automated tasks. So, I presented it at the end of Feb at the London Bioinformatics Meetup. And I was absolutely surprised because the London Bioinformatics Meetup has people from the top scene in the UK.

[20:55] In London, you have people from the top pharmaceutical companies, from the top universities, global leaders. And nobody knew what Claude Code was. And it was like, oh my God. And so, at that point, then the next week I joined the AI agent hackathon, Agentic AI, Imperial College London. So, it's the biggest hackathon on Agentic AI in Europe. And then I presented ClawBio. Yeah, there were hundreds of groups for the hackathon. I presented Claw and it won the gold medal. It actually won the hackathon. I've never won anything in my life.

[21:45] And so I thought, wow, I'm onto something. And since then, the growth of ClawBio has been exponential. And I get these people now contacting me from, I don't even know who they are. I get all of these people contributing code because there's nothing like that. I think that the main set of advantage that ClawBio has is that, although I'm sure there are other companies who are doing agentic stuff, there's nothing open source. And so here we are. I'm this unlikely person who somehow seemed to be a couple of months ahead of where the curve would start inflecting. I'm not saying I'll be that way in a few months, but I can tell you that I've been to talks with Anthropic and with other big, big frontier labs.

[22:53] And definitely what we're doing right now in terms of agentic genomics is certainly a few months ahead of what they are talking about. Maybe they might be talking internally, but publicly, I believe that this is ahead of the game right now. And so for me, I need to capitalize on this moment because we are at agentic speed, AI speed, one week is one month, one month is one year. And so I've been trying to cope somehow with this new way of doing things. And I think today has been a culmination with this first paper ever on a top journal on agentic genomics. So I think, and I'm hoping that this new paper will be defining for the field.

[23:44] Arseniy Arsentyev: I think that one week is 10 years now. I'm a computer scientist by training and the stuff that I can achieve now is just mind blowing. And speaking about why you were on this wave, I think it's a combination of the right skills because you were working in genomics. It's not that you learn genomics in one week. And the second is that you were the right place, as you said, I know London, it's a great hub where many science and biotech projects are happening.

[24:19] I'm just grateful that agentic genomics is existent because I know that many bioinformaticians were crazy about routine tasks that they should perform. And many even top people were doing some stuff that is probably automated. And they were doing it. And right now we can offload this to the agents. My next question for you is, why does it matter? Why does genomics matter? We can pose even these questions, what scientific research done with the help of agentic genomicists and the systems can bring to the world?

[25:07] It's a deep question, actually.

[25:09] Dr. Manuel Corpas: I think what I always say, my own personal strategy, is that the questions are the same. But what's changing is how you do it. And I always say that the way to get into this is not by trying to find new problems. You're already working with the problems. It's just that now the thing is, how do you do what you were doing before, but now in an agentic way? And that's what I've been doing again and again. And first I started with pharmacogenomics. So, because pharmacogenomics is a easy tractable problem. So you have a mutation, you're going to have a phenotype. Full stop.

[26:04] So there's a black or white, an easy way to validate it because you have the CPIC guidelines, which is an international society for pharmacogenomics that tells you very clear criteria. And so those are the low hanging fruit where you already know what the correct answer is. And so then if you test your particular problem, then you can easily know whether the agent workflow has worked correctly or not, because you can benchmark it. But where the real sort of challenge is when there are no sort of clear benchmarks, say for instance, variant calling, or even, there's a benchmark for variant calling using, for instance, Genome in a Bottle. But you have very few genomes. Anyway, more on the interpretation aspect, you are trying to sort of interpret the actual effects of any given variant.

[27:17] It's not a black and white answer. And it's very convoluted, the type of criteria that there are, and so on. And it's very difficult to really have a sort of publicly and openly accessible standard gold truth, which then you can benchmark against. So I think that's the sort of, where the future is going to go. And it's going to be jagged. So for some things, agents are going to do really well, particularly if you have well-benchmarked standards of truth. For those part of the genomic world or omics or bio work where it's not sort of clear what the right answer is, that's where there's going to be a lot less development.

[28:16] And so the level of specificity that is required for this incredibly heterogeneous and diverse set of skills required for any kind of bioinformatics pipeline is never going to be captured in any of the generalist models. They'll be really great for reasoning. But when you get into the real deep sort of expertise domain, that's never going to be captured. And that's where I see the opportunity and where I think that we are right now in, at the right time for people to be able to join and to hopefully build on those very sort of specific, thin, but deep expertise.

[29:17] Arseniy Arsentyev: And how many people are working on ClawBio right now? And how many people are using ClawBio? Can we get some numbers?

[29:26] Dr. Manuel Corpas: Yeah. So it's completely open source. It costs me money. I think right now, my last count, there have been 43 people who have contributed code. And it's been downloaded around, cloned, should I say, around 5,000 times a week.

[29:50] Arseniy Arsentyev: Oh, wow. It's a great number.

[29:54] Dr. Manuel Corpas: Actually, considering that it started in March, really, when I made it public. So it's only been like four months. This is, again, agentic speed and exponential growth.

[30:10] Arseniy Arsentyev: Was it growth right from the start when you published?

[30:13] Dr. Manuel Corpas: Sorry?

[30:14] Arseniy Arsentyev: Was this type of growth the same right when you published or was there like a plateau?

[30:20] Dr. Manuel Corpas: Yeah. It's been a little bit like ups and downs, and it follows if I am giving a big hackathon or if I put something on LinkedIn that suddenly goes viral. I wrote this book on agentic genomics.

[30:34] Arseniy Arsentyev: I've seen it.

[30:34] Dr. Manuel Corpas: And then suddenly, so it has some peaks. But I found it quite stable, and it's always around 4,000 to 5,000 sort of downloads. And I've never built anything like that before.

[30:55] Arseniy Arsentyev: And you've said you were inspired by OpenClaw founder, and I've seen on the LinkedIn, I believe, that you met him at some presentation at Imperial that he did. Did he gave you some advice and was it some interaction?

[31:13] Dr. Manuel Corpas: Yeah. He was famous and busy. So Peter Steinberger was based in Vienna and London, somewhere in between. Now he's also a lot of the time in San Francisco because he's working for OpenAI. But I've been at a number of events where he's been around and I've been listening to him, but I've never had a proper conversation with him. He was around and: oh, Peter, can I have a picture with you?

[31:45] Arseniy Arsentyev: It's like me attending some Bitcoin conference. Yeah. And so this agentic speed and it brings us to the second part that I wanted to talk about, which is the wrong scaling. Cause right now agents are scaling, anything that people do. And health, bioinformatics, genomics data is very inequitable. It's a huge topic on your research that about 86% of data is of European ancestry. And then my kind of next thought that this agent systems even exacerbate such an inequity, how can we deal with it? And what are the biases that you currently seeing?

[32:40] Dr. Manuel Corpas: Agents and AI in general, they amplify everything. Yeah. So it starts from a base, which you have to imagine. And this is something, it goes a lot deeper than the actual data that we are dealing with. You think about the scientific enterprise, in the past, let's say since the industrial revolution, where has it really happened until very recently? It's been Europe and the US. And so there is this traction, that has gone for decades, even centuries where the people start, what are the best known diseases? Well, the diseases that affect the global West. What are the sort of best understanding of the biological mechanisms? Well, around the things that have, where most of the funding, most of the energy, most of the companies, the industry have been developed.

[33:48] And so this is incredibly entrenched in all the biomedical enterprise. And this is something that I don't think has been properly put across. I have a number of figures. You think about, for instance, clinical trials: in ClinicalTrials.gov, there's around half a million clinical trials. And you basically then compare the number of clinical trials associated to specific disease. And then you classify them where these diseases are predominantly from the global North or the global South. And the disparity is breathtaking. Then if you look at the, even just the sheer, scientific literature, there is the body of knowledge, you can look at something called PubMed, which contains all of the biomedical research that have been published since the fifties or something.

[34:51] And again, the body of knowledge is incredibly well developed for things that are from these particular countries. And then the further you go from those countries, then you have diseases like malaria, and until very recently COVID. Basically the connection between the language that is being used in those publications for type two diabetes, for sort of industrialized nations' diseases, is incredibly well connected. And so you have this set of situation where let's say, you develop something for COVID, and then anything around, topics that would be related to COVID, such as for instance, immunology, or let's say, DNA recombination, whatever they get this, there's quite a lot of spread, the things that have been developed for type two diabetes or for heart disease, it can be easily spread to other diseases of the Western world.

[36:07] And then you analyze the language of diseases that are affecting the global south, such as for instance, parasite diseases. And you find that the language is quite disconnected. The diseases that affect 1.5 billion people, they're still at the same sort of levels they were a couple of decades ago. So this level of granularity in which we are able to sort of identify how knowledge is being produced hasn't been properly analyzed. And that by definition, which is the sort of body of data from which AI is then built, then it's completely amplified because that's what AI does. AI amplifies everything. And so the way to really fix that is that health equity has to be architected from the design itself.

[37:27] You cannot really expect that by doing more research on a particular sort of neglected disease, you're necessarily going to sort of have all of the impacts that you could have. There's a lot more structural type of ways in which research is done, which is mostly driven by financial sort of needs. Of course, we live in a capitalist world and you have to have return for investment. But what we are not aware of is this cycle of negative reinforcement, where you have a cascade of neglect that keeps those diseases that affect the poor, the global south, still not advancing at the same level.

[38:29] And I have data where you can see if you have diseases that are well connected in the knowledge space, the degree of hallucination is a lot smaller than diseases that are not so well represented. And in fact, we show that we can predict it: we benchmark it with nine models, Gemini, Claude Opus, Claude Sonnet, GPT-5.4, and some others. And they all independently have a similar amount of hallucination that correlates inversely with the knowledge representation for those diseases, especially of the global south.

[39:18] So, and then what we also find is that you can encode a lot of those inequities in the skills. And if you're able to sort of architect and design, and know specifically that you encode these differences that are specific for those diseases by design in the skill that then agents are going to be exposed to, then you have a tremendous guard rail. And we find that those differences that we find, especially in different populations, from Europeans into Africans or Asians, we see that those differences that you sometimes find due to the ancestral ethnic background, they disappear if you are able to encode them by design as part of the sort of skills and the design of the context that you give the agents. So I think there's room of hope.

[40:26] We have on the one hand, better ways and better tools with which you can analyze and understand the structural inequities. And we can understand that from angles that were not possible before. But then at the same time, we know that if we are able to pick what those gaps are, there are ways in which you can use those sort of strengths that AI has for amplifying into then closing that gap.

[40:59] Arseniy Arsentyev: Do you think LLMs would be, LLMs and these big AI companies would on itself be able to fix this health equity problem? For example, when training the models, we can use the same models to evaluate the strength of evidence and understand the statistical distributions and understand that probably information is misrepresented? Or would it be necessary for real people to take part in it? What's your sense on that?

[41:34] Dr. Manuel Corpas: I think they're going to be in the same situation as we are. So unless you identify where the gaps are, we are completely blind.

[41:43] Arseniy Arsentyev: I think people should identify where the gaps are.

[41:48] Dr. Manuel Corpas: People like myself, we have a set of tools now, which give us an amount of capability that didn't exist before. And we can use those new capabilities specifically for the purpose of identifying health inequities. And this is what I've been doing now for five years, six years. And it's incredible. And so, you are able to find new sources of inequities that we were not even aware of before. And so, then, by identifying those gaps, then you can leverage the power of LLMs to hopefully engineer capabilities that will help us make those skills aware of those gaps and therefore at least close the gap. Now, you cannot expect that the companies are necessarily, the frontier labs are necessarily going to address this.

[42:52] I know that Anthropic just launched this sort of rare diseases challenge, but there are very specific use cases. And I think they are playing with the same data that is available to all. And so, the thing is that if you have biased data, then you're going to have biased models. And so, there's no way around it, unless you have a very specific targeted way where you can really, not just the models, but the way you apply those models, make them aware of the specific gaps and hopefully close the gap.

[43:28] Arseniy Arsentyev: Okay. And you have a growing, ever-growing community for ClawBio. You see some ideas that people suggest. You have ideas on your own. What are three main threads that you see this product evolving and what are you working on right now?

[43:48] Dr. Manuel Corpas: I'm just going to give you just one. So, I've said it already in some other conversations. Having skills and having a diversity of different sort of capabilities that could then be reusable by the community is great. But I think where the bottleneck is in having a set of robust skills that you can trust. So, this sort of trustworthiness slash validation aspect of can you really use this? Let's say at a, imagine that you're a hospital or you are a pharmaceutical company, the stakes are huge. So, how can we really be in a situation where we feel confident that this agentic thing is going to perform in a way where the error or the hallucination is not 99% correct? 99% correct in the clinical aspect and the clinical realm is inadmissible.

[45:00] You cannot have that. You need to have a millionth of a percent sort of error for it to be clinically applicable. And many orders of magnitude even than the current sort of established standard because when you're introducing a new technology, you don't expect that technology to be as good as it was before. You expect it to have three, four orders of magnitude better accuracy. So, understanding how we can trust those agentic enabled workflows. Understanding not only whether they are accurate, but understanding the different breakdown checkpoints because accuracy, you can have an accuracy and that accuracy you've achieved through a number of routes. So if I think, for instance, of genome interpretation.

[46:07] So, you want to classify a particular genetic variant whether it's pathogenic or not. So, of a benign variant, a variant that doesn't do anything, you can reach that route through many different paths or a variant of uncertain significance. You don't know whether it's uncertain because you lack evidence or because there's not enough sort of ingredients for it to be sort of clear or maybe something failed. So it's not just about understanding the accuracy itself. It's about having the granularity of all the different components that then give you the final number of accuracy. So that decomposition of how you arrive to the specific result in a way not just that is validated, but also auditable, traceable, reproducible.

[47:08] And it can be inherited. That's where I think the real future is. And that's where I'm going to spend my energy for a long time moving forward.

[47:22] Arseniy Arsentyev: Do you think it would be a long time? What's your time scale for reaching some, maybe not the grand milestone, but the closer milestones? Is it a few months in which...

[47:33] Dr. Manuel Corpas: Next month. Yeah, it has to be next month. Next month, I will have something. I'm not saying I'm going to solve all of the problems. But again, I think it's this question of picking... And this is where the real strategic thinking comes. The number of problems you can work with is incredibly diverse. So pick the problem where most of the people in genomics, they're interested in. Better interpretation. Everyone wants to do better interpretation. Hospitals, researchers. And it's a hard problem. So that's the problem I pick. And that's the problem where I'm hoping to have a solution the next month.

[48:28] Arseniy Arsentyev: Great. In 2026, it may be a personal question, but do you run your agents 24-7?

[48:39] Dr. Manuel Corpas: Yes. And I've had a number of surprises. I'm completely broke.

[48:46] Arseniy Arsentyev: A pipeline broke something, you mean? No, no. Poor.

[48:51] Dr. Manuel Corpas: I've spent a huge amount of my money.

[48:54] Arseniy Arsentyev: Oh, okay. I see.

[48:56] Dr. Manuel Corpas: For instance, I ran out of subscription credits. I did the other day. So I had something, an agent working for about three days nonstop. And then I thought he was going to do something really great. And then I had to discard it completely. But then it's part of the game. And this is, I guess, part of my secret ingredient for being successful in this: I was very quick to spend my dollars, get the subscriptions. And I spent a lot of money in credits, trying this, trying that.

[49:45] And I already have assumed that maybe 80% of all the money you spend is going to go to nothing. But then I get all this experience. Then I start understanding how the different harnesses work, how the different models charge. Sometimes I was spending like $60 a day just on all of these things. And I had to stop. Sometimes I've had the situation where I had agents spiraling. And suddenly I see like £500 going out of my account. So it was having to stop everything quickly.

[50:29] And it's all experience. And I think the fact that I've been able to spend that amount of money and...

[50:38] Arseniy Arsentyev: I was about to tell you, in finance, we have time value of money. Dollar today is valued more than the dollar tomorrow. And in AI agents, it's completely reversed. Like what impact $20 Claude subscription gave you a year ago is incompatible with this $20 that it is giving you right now. Yeah, it's crazy. So I got token-conscious, if we can say so. And I'm truly thinking maybe when Fable came out, I'll test it. And I've postponed it on some of my projects to wait for the newer model to allocate intelligence better. Because when I started using Claude Code, it was also a year ago, similar to yours experience.

[51:37] Let's be honest, it was so much dumber. And it couldn't do many things. And I've lost also pretty much of credits on it. And yeah, it's just a weird paradox in AI. That you have not increasing value in time, but decreasing.

[51:57] Dr. Manuel Corpas: Yeah. Many of the paradoxes that we are now confronted with. But if you want to be six standard deviations ahead of the curve in terms of adoption, you're going to have to spend your money. You're going to have to spend your time.

[52:14] Arseniy Arsentyev: 100%. And I'm seeing that you're building media presence. Is it your intent to just get more people to know agentic genomics and your project ClawBio? Or were you doing it before you started working on it? What's your stance on media presence? I'm personally interested because I'm also trying to put my work out on the internet. I'd just love to hear your story.

[52:48] Dr. Manuel Corpas: So I think my goal is to be the person who defines how agentic genomic interpretation gets trusted. And I guess unless you really are able to ship products, things that add value, you're out. And so I used to be one of those who, again, in the social media years, I have my blog. I was one of the first bloggers in my field. I also have my YouTube channel. I have to be on everything. Then I've come to realize that now I don't care so much to just have mere presence or put pictures or say, oh, look, I attended this, I attended that. But now I am more of someone who says, here is something, here's a powerful deliverable, like this article.

[54:04] Here is this tool or here is this new cool functionality that I've been able to do that wasn't possible before. So I've changed a lot in terms of how I publish things. So I just don't put things that say, oh, hey, I'm showing off this particular, oh, I'm here, oh, look, I'm so famous. So I'm more of someone who pushes things when I have something important to some learning or some value contribution that I have to share the world. So every week I have something big to say, I've published this book or I've published this article or I've published this or I'm organizing this event.

[54:59] So, and that somehow retrofit itself into a community where especially my LinkedIn account has been exploded since in the past few months, from, I get, I don't know how many new contacts every day in my, as followers. So that has happened before. And so I guess it's like, it kind of fits itself. And then when you get your particular name, your particular brand, your particular niche, then people are really interested. Especially if you're producing things that people perceive value.

[55:46] Arseniy Arsentyev: And now we're at the place when we can put all things together. Bioinformatics went from this wet lab science and basic computer programming to big data, to this boom of databases, huge data sets, huge computation, into now self-sufficient, maybe in the future fully agentic system. And our challenge is to make sure that it provides necessary health equity, that it is validatable, that it is trustworthy. And we're about to speak on the topic of my show, which is tomorrow's medicine, and bringing this all together, bringing your work on agentic genomics. What do you think will happen in a decade, in 10 years? And what do you think will happen to our medical and health system? How do you think this would practically change, not only scientifically, but practically change how medicine practitioners, providers, clients, patients interact?

[57:10] And do you see a bright future or a great one for a healthcare system?

[57:17] Dr. Manuel Corpas: I see a great future as long as there are some of us who fight for the truth. And I see a future where capability and knowledge is democratized. So the bottleneck is not intelligence anymore. It's your ability to ask questions. I see a future where you do your research in conjunction with the AI agent as a partner, which I'm already doing that anyways. So it's not, my agent is not a server. My agent is a colleague. I keep asking things. We go back and forth. And so you have this sort of symbiosis between the human point of view, which has this sort of nuance of the real world that I don't think that AI will ever capture.

[58:21] AI will be able to mimic human intelligence and go much beyond the processing that is able to hold a billion token context windows and things like that. So they'll be able to do a huge amount of thinking that we will not be able to do. But how we apply that to the human reality, to the things that humans really care, as a way to sort of develop us, not as servants, but as sort of enhanced, enlightened beings. That's where I want it to go. And I want it to be something which is not used as a weapon by institutions, by governments, something that is able to reduce the barriers and allow us to do things that were unthinkable before.

[59:37] And I'm already doing things that were unthinkable before, in terms of my outputs, in terms of the kind of thinking that I can do. I don't have to worry about the math so much about what algorithm to pick or choose. I know that AI is going to help me choose the best optimal algorithm for this particular task. But what it doesn't know is what I really care about. And so they'll be able to be very empathetic. But I still think that the sort of human relationship aspects and the, I guess, sort of physicality that we are able to have, the experience, the organic side of things, they will never be able to sort of reproduce.

[1:00:28] And so I'm not interested in superintelligence. I think it excludes. And what happens if you have Down syndrome or people who have disabilities? I don't think superintelligence is a worthy goal. I don't think that superabundance is a worthy goal. I think it's selfish. I think what is a worthy goal is to go through this new period of enlightenment, super enlightenment, where we're able to achieve a new growth and understanding that makes us more human, that makes us more complete, that is able to help us understand our nature, our physical nature, our spiritual nature, our emotional nature better than ever before. And that's the goal I want. And that's the goal I'm going to fight for.

[1:01:21] Arseniy Arsentyev: Thank you very much, Dr. Corpas, for appearing. It was a nice episode. I took a lot from it. And it was great to talk to the person who is on the frontier. We'll definitely follow you, follow what's the future of ClawBio, what are its developments. And once again, thank you for participating.

[1:01:41] Dr. Manuel Corpas: My pleasure.

[1:01:42] Arseniy Arsentyev: Everyone, it was 16th episode of Tomorrow's Medicine. You know what to do. Subscribe to Dr. Corpas. Subscribe to myself. Leave a comment. Share with your friends. And see you next time.