Eli Lilly & Company Built an AI Supercomputer to Ask Bigger Questions

0:00Dr. Thomas Fuchs: Today, AI, it's like soccer. Everybody's talking about it. Everybody has opinion. And nobody is actually playing it in the world class league, right? Very few people. And that's how AI is. And then if somebody is talking and bragging about the AI skills, and they couldn't spell AI three months ago, then maybe that's not the best source. Enterprise AI is entering a difficult phase, whereas even a few months ago, isolated experiments were celebrated. Now, organizations that are serious about AI success must turn those experiments into systems that can reimagine how work gets done. Now, Eli Lilly offers one of the clearest examples here.

0:38The company is combining its own supercomputer, decades of proprietary research, automated labs, and a growing biotech network to rethink how medicines are discovered, tested, and manufactured. Early results include new genomic insights and AI-assisted molecular designs. But the larger lesson reaches well beyond pharma. AI infrastructure matters only when it expands what experts can discover, build, and deliver. Joining us on today's show is the leader overseeing that effort. Lilly Chief AI Officer Thomas Fuchs, a machine learning scientist whose career spans NASA's Jet Propulsion Laboratory, computational pathology, and digital health.

1:17Thomas is now helping Lilly use AI across discovery, development, and manufacturing with one clear goal, getting medicines to patients faster. And the insights Thomas shared with me provide a potential roadmap others can use to replicate Lilly's success to date. This is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

1:46Thomas, welcome to the AI Proving Ground podcast. How are you today? Very well. Thank you, Brian. Thank you for having me. Certainly lots of excitement happening in the industry. And I can't think of another organization beyond Eli Lilly that's really made this big of an investment of a bet on AI infrastructure with LillyPod. just to start out, just to level set, tell us a little bit more about Lillipod and what its ambitions are and how you started to develop it.

2:16Dr. Thomas Fuchs: Yes, it truly is a big bet. And I'm very grateful to have a fabulous team here at Lillipod who was really supporting that from the beginning. First of all, it's all thanks, of course, to the vision of Diogo Rao and Dave Ricks that we actually embark on that journey. And then the fabulous team who actually build our supercomputer and all the scientists and machine learning and chemistry and biology and all the different areas were actually utilizing it. So the whole point of the exercise is really to build something that helps us, first of all, to control our own destiny, that we can really operate at a scale we need to, to be an AI-driven pharma company of the future so we can design, build, predict, establish, prove medicines for many, many more patients across the world.

3:12Dr. Thomas Fuchs: As you know, there's, of course, a dramatic need in all the areas Lily is working at and, of course, beyond. And LilyPod, our supercomputer, is one means beyond others, but a very important one to really drive that forward. And the reason to do that on-prem is really to have a resource our scientists and machine learning experts can use at will and at large scale. So it opens up the horizon. Instead of having just small runs here or there, you now can really think big. What could you do if you could actually train a model on hundreds or a thousand GPUs? What would that allow you to do? What kinds of medicines could you design?

3:54Dr. Thomas Fuchs: What kind of property prediction models are in manufacturing? What kind of digital twins could you build? And that change in thinking from resource-starved environment, like many of us who were in academia are used to, to an environment where you can really think big is a game changer for the organization. No, absolutely. And you talk about it becoming a control your own destiny type of situation. And I've seen you describe it before as compute being a means to an end. I'm curious from where you sit, what is that end, so to speak, that you're most excited to accomplish or maybe impatiently waiting to resolve itself?

4:36Dr. Thomas Fuchs: yes so my background is in machine learning of course i like the math behind it but we are not just doing that to write great papers the means to the the end here right is truly designing medicines faster producing them faster getting them to patients fast and everything that's necessary across that so we build ai throughout lily not only in the large language model space which is, of course, drastically important, and we can talk about inference later, but also, for example, diffusion models for molecular design or process optimization models in manufacturing, forecasting models in marketing and finance.

5:15Dr. Thomas Fuchs: So the space of AI is, of course, very, very large. And to get to that end, to design medicines faster and get them out faster, you have to build in all these areas. And that demands, of course, an enormous amount of compute, and that's what we built Lillipot for. Yeah, I mean, Thomas, you had mentioned kind of, you know, your background, perhaps a little bit in academia, but your career spans a wide gamut here. We're talking NASA's Jet Propulsion Lab, computational pathology, digital health, now certainly, you know, more enterprise AI scale inside Lilly. What did all of those kind of stops along the way?

5:53How did that shape how you think about AI, how it should be deployed on a massive scale such as Lilly? And maybe more of like, did it give you any unique insight in what it can do or how it should be treated?

6:05Dr. Thomas Fuchs: Yeah, first of all, it shows, of course, the beauty of machine learning and statistical learning. It can be applied from the microscopic scale, where we used it in pathology to differentiate cancer from non-cancer or identify molecular drivers on a microscopic level to a macroscopic level. Where at JPL, we used it on imagery from Mars or from the cameras from the Mars rover to differentiate sand from gravel, from rock, so it doesn't get stuck. And the methods behind that are quite similar, of course. It could be computer vision methods, and then, of course, these days it's all deep learning and transformer models.

6:46Dr. Thomas Fuchs: And what unifies all of these is they are certainly driven by you need a deep understanding of the underlying science. So be it a pathologist in oncology or be it a geologist on Mars. In all these areas, the ground truth is hard to come by and you actually have to incorporate all that knowledge you have from the domain experts. So it's truly an interdisciplinary effort. And it's even more so at Lilly because it goes beyond, again, pure chemistry or biology. You need everything downstream. Can you actually develop the drug afterwards? Could you scale it? And so you need truly an army of experts to get it done.

7:35Dr. Thomas Fuchs: That makes it interesting, but of course also difficult. And the really successful AI applications incorporate all of that. So that's one of many reasons why you could not just have an autoregressive large language model that will pump out the next drug that will not happen with the current technology. But if you have all these other building blocks in place, then you really can make a bet for the next drug. Yeah, you talk about getting AI into the end users, so to speak, the domain experts to kind of unleash it and see what they can do with it. I am curious, how has LilyPod AI, how has that changed some of the questions or the work that Eli Lilly scientists are working on?

8:21Has it made tangible impact thus far on just pushing the boundaries of what's possible?

8:27Dr. Thomas Fuchs: yes it already did so again first of all changing just the the the sheer scope opening the aperture is a cultural change that is drastic across all all functions and then second of all we have a lot of data of course we have a lot of proprietary data there's sometimes public data that can be used and we have also a lot of historic data so we have data not only from successful drugs but especially also from failed experiments and another reason why you couldn't have an ai scientist just trained on publications with successes of course the biggest part of science is all the failure right you fail you fail you fail until you succeed and if you can learn from all the failures which you we can internally right we have millions and millions of molecules where we know it doesn't work and you can learn from that, that is very powerful.

9:23Dr. Thomas Fuchs: And now with LilyPod, we can actually use all of the data. So for example, in protein design, we just could use 5 % of the structures in the past and now we can actually learn from all of our structures. In genetic medicines, we can learn from sequences at an awe-inspiring scale. And that's what compute allows you to do. It's really a scientific instrument. It's like a microscope that allows you to go deep or a telescope that allows you to go further. And compute just allows you to take so much more data into account. And then you can also train larger models or more difficult models. You can iterate.

10:01Dr. Thomas Fuchs: You can use reinforcement learning on top of it together with our domain experts. And that's really a flywheel that gets us better and better models. so in molecular design we have the first molecules that go to through internal testing before they go in vitro in in genomics we have insights in some of our medicines we didn't have before so they are already now after just a few months tangible things that that are very useful but of course the the future is a wide open land here very excited what we can do in the months and the years to come because it is difficult right you will not we cannot in the years to come have one model that pumps out a functioning drug that space is too complicated the mechanisms of a single cell go far beyond our that even our human language can just describe or it's much more difficult than building a language model on text and that's challenging but exciting and necessary If you think of all the need and suffering in the world triggered by disease, that's an area where we should bring AI to bear.

11:16You mentioned scientists learning as much from failure as they do success. And certainly that's in the data that Eli Lilly is using in Lillipod. I am curious, data is such an obstacle and hurdle for many organizations that we talk to. What type of decision-making had to go into data architecture or strategy? Or were you already in a good spot when you started to undertake this initiative?

11:42Dr. Thomas Fuchs: So with data, you are never in a good spot. It's always a moving target, right? It's important to understand if you wait until a time when you think all your data is cleaned and sorted, and then you put a bow around it, and only then you're going to do machine learning and AI, then you're going to wait until eternity. eternity. You have to work with what you have. And we have fabulous teams beyond mine that deal with that. We have a software product engineering team that's revamping all our data infrastructure and data access mechanisms in an authentic way, in a very unique way that really accelerate what we can do, not only in discovery, but also in enterprise data.

12:25Dr. Thomas Fuchs: and then a big advantage we truly have beyond let's say a tech company that tries to go into that space we have thousands of chemists and biologists you have to understand that data without that understanding it's kind of a shale it doesn't it doesn't yield the results you want to and then we also are embarking on producing a lot of new data because although we have a lot that's the biology space and chemistry space, the data is still scarce, right? Very, very scarce. For example, if you think about small molecules, there are more drug-like, potential drug-like molecules than particles in the universe.

13:06Dr. Thomas Fuchs: And so we only have a teeny tiny space of that. How do we get more? We are building agentic labs, automated labs, that really automated loop, how we can produce more of the data. So we have this design, test, make, and analyze loop really becomes a design test make and learning loop. So a continual learning loop where we build very large models on lily pots, propose designs or all kinds of other things. So what we're doing with these agentic labs and data factories is we really want to get to a world where we can do continual learning. So you have your very large machine learning models on lily pots that propose designs, for example, of molecules.

13:46Dr. Thomas Fuchs: Then you send them automatically to the lab where it's completely automated with robotics and the molecules are produced, the experiments are run. Then you can get the data back, you update your models, and that gives you that flywheel where you really do design, test, make, and learn from it. And if we are able, or when we are able to do that at scale, that yields then breakthroughs. We are very, very confident will change how drugs are developed in the future. Well, Thomas, obviously what you're talking about here, that's a very sophisticated process there. I'm curious, can you think of any type of obstacle or hurdle that you experienced along the way that surprised you and maybe offer a little bit of insight into how you overcame any of that?

14:36Dr. Thomas Fuchs: So there are, of course, many hurdles. It's...

14:44Dr. Thomas Fuchs: so we often are led astray especially if you think of chat applications in llms to think these things are easy because the language models are amazing especially in code so you enter a prompt and you get something out but you think about designing a molecule you are in on a totally different scale. Chemistry never behaves as you want to. Biology for sure does not, because we just do not understand enough. You need enormous expertise across all of that. And even before you start in actually building a supercomputer, that does need expertise. It's, of course, financially beneficial to do it on-prem.

15:28Dr. Thomas Fuchs: And in addition, you can use all your date. Without constraints, you control your security. It's a geopolitical hedge, etc., etc., etc. But of course, it needs expertise. It needs someone who truly understands, of course, the GPUs you're using, the storage layers you're using. You have to build the infrastructure, water, cooling, etc. And then you need a great team to run that. And that alone, just to get to the instrument, right? But then when you have it, then you can scale up and scale out and we are increasing our capabilities constantly. So when you think about going beyond the design and you go into clinical development and you need enormous machineries, of course, to run trials, to have all your partners across the world actually do that.

16:17Dr. Thomas Fuchs: And the patients who are willing or eager, of course, to participate and then hopefully it works. And then you have to produce it and you have to roll it out, et cetera, et cetera. So it's a humongous endeavor. And that's why, of course, companies like Lilly exist. And we are doing that quite well at that scale. And AI plays a role in all these areas and LillyPod does too. A number of times thus far, you've mentioned the benefits of having all of this on-prem as owned infrastructure. That's also a conversation that a lot of organizations that we deal with here have within their executive ranks, the buy versus build.

16:54So walk me through some of that decision-making on why the on-prem model made more sense for Lilly as opposed to taking advantage of any external capacity.

17:07Dr. Thomas Fuchs: First of all, of course, you have to be very precise about your use cases. We are great users of all kinds of cloud infrastructure and hyperscalers. And for many, many use cases, especially classical SaaS software, etc., etc., the cloud is the right way to go. And my teams implement a lot of their work in the cloud. And it's especially useful also to, we use it to run inference on the finally trained models often in the cloud. So there are many, many good reasons to do it cloud-based. But when it comes to training machine learning models, there are many reasons you would want to control it yourself and you want to do it on-prem.

17:54Dr. Thomas Fuchs: Finance is one. There is a huge difference just in scale. What we could do price-wise if you do it on-prem, when you already have a data center as we do instead of going to the cloud. But other reasons is also you can build really for your use case. We don't have to serve the world. We can really build the supercomputer, the infrastructure and applications for what we need at Lilly. Then you can combine the GPUs and the storage layers. Something people often forget is that GPU-adjacent storage in the cloud is more expensive often than the GPUs. And on-prem, you can do that very elegantly. and you also can combine it again with access to all your proprietary data.

18:37Dr. Thomas Fuchs: You can control security around it. Second of all, we live in volatile times and geopolitically hedge against supply chain disruptions. I think it's a wise thing to do. That also means you can run applications independent of the cloud or the internet, which is important when in the future we use it, for example, for manufacturing applications or other areas which are really sensitive. You don't want to be disrupted in any way. And when you train these models, that's what the compute intensive part is. You need hundreds of thousands of GPUs to do that at scale. And that's very hard to do on a consumer cloud.

19:23Dr. Thomas Fuchs: So, you know, Brian, these days it's kind of a no-brainer. You talk to your AI startup XYZ, they're all doing that. The VC funds are buying classes for the startups. And in my view and many of my colleagues, it was, of course, the case for a long time. We built the first DGX in healthcare in 2016. We used high-performance computing 20 years ago for large random forests in oncology. compute is always what drives modeling at scale. It's a little bit of a different scheme, and one reason why it's foreign to many is that often also science is very hypothesis-driven and focused in a specific area. So many biologists want to investigate, let's say, only one single kinase until end of days, which is very, very valuable.

20:18Dr. Thomas Fuchs: But if you have all the data, you need larger models, And then you can drive also larger models. And then you can use more data again. So it's really a very beneficial loop. And I think many saw that very early. And it's really the ones who didn't, who are surprised now that you need that. That's Cain. And then you need, of course, a team who believes in it. And that starts with the CEO and goes down to everybody at the organization to help to build that. And then also to use it and feed it with ideas. If the market dynamics are such right now that that on-prem decision at a certain scale is, to your point, a no-brainer, do you see those dynamics holding for the foreseeable future?

21:04Or what would have to change, knowing that the AI landscape is always in a state of change, for any of that thinking to shift?

21:10Dr. Thomas Fuchs: A very good question. Hard to predict the future there. I mean, you could, of course, imagine. So in the past, for example, we usually only trained on-prem and did all inference in the cloud. We did that with the startup I founded in Pathology page, and we did that before as well because it was just easier to deploy. You need 1 ,000 GPUs to train, but you could do inference on just two or three in the cloud. Large language models and prediction time inference compute changes that picture. Case in point, one application we are using Lillipot now, in addition to everything we discussed, building our own foundation models, is actually use it to run large language models locally.

22:01Dr. Thomas Fuchs: Token costs are exploding drastically. And you can imagine an organization like ours with 50 ,000 employees and a lot of demand on agents and agentic work. There's an enormous amount of tokens that need to be processed, an enormous amount of applications for large language models. And there are good arguments to do that also on-prem. So we are implementing the largest open source models on-prem with all kinds of harnesses so people across Lillic can use that. And then imagine your world when you do not have just a set number of tokens you can use per month or a set number of dollars you can use, but you suddenly have near infinite tokens.

22:43Dr. Thomas Fuchs: And very fast, because the B300s we are using are optimized for floating point four, which most of the models are implemented in. So our local inference is much, much faster than all the third party providers in the cloud. So when you can do that, again, you can think differently. that goes back to also what we discussed in discovery. The whole point is really changing how you view the world, right? Go away from that restrictive thinking of constrained resources to a world of abundance. Abundance in GPUs, we build larger molecular design models, and now abundance in tokens. So how would actually your authentic loop in clinic or development look like if you don't have these constraints?

23:26Dr. Thomas Fuchs: So that's a very, very exciting space. And so we're using part of Lillipod to run these local models and open that up for everyone. We also implemented something called Lillicode, where you can really choose the models you use. You can choose to run it through the cloud and use frontier models. But most of the cases, the open source models are maybe six or eight months behind. And for most of the applications, that's completely fine. So there again, scale would be depth, although the difference in depth is not that big anymore. But again, it opens the aperture. That's the exciting part here.

24:23secure and streamline your core network services with Infoblox's intelligent DDI platform. Yeah, you talk about abundance and, you know, the potential of near infinite tokens, I can just kind of sense the CFOs of the world having big eyes here of being like, Oh, hold on a second. I mean, it is an interesting conversation as we start to talk about tokenomics and a lot of organizations around the world, kind of shifting from, hey, let's experiment, experiment, experiment to now saying, let's figure out this cost control perspective. How does Eli Lilly go about thinking about that tokenomics? Because to your point, like you said, you have tens of thousands of scientists who are potentially using a lot of tokens.

25:07Dr. Thomas Fuchs: Yeah, so our token costs, like everybody else's, are exploding, which is on one hand a good sign. People are really using it. So everybody is trained and is, of course, should use large language models and AI beyond that. wherever they can, wherever meaningful, wherever appropriate, of course. And it shows enormous impact throughout the organization. So there's a plethora of examples where it adds value. But the models are getting bigger. The chain of thought loops at the end, if you call it reasoning or thinking, are getting deeper. And that, of course, drives token costs into the sky. and it will just increase in the years to come.

25:51Dr. Thomas Fuchs: And again, one of the solutions, and we think a big one, is just running your own models. And if you already have a compute cluster that's optimized for deep learning, it's easy to do. So we actually only need a small fraction of lily pods to cover trillions of tokens we need for all our agent decoding, for example, for all the electronic notebook analysis, for a lot of the work we do with large language models throughout. To the extent you can, tell me more about LilyCode. I think you said it was a resource that you have that you can help choose your model. Was that born out of the tokenomics discussion or what other value does that have?

26:37Dr. Thomas Fuchs: It even started before. It's a means to help not only the software engineers but also the statisticians, for example, who are big adopters of it. And now many of the scientists to give them choice, which models make most sense for them. We're expanding that also now with a choice of harness, which harnesses make most sense for them. And it also allows us, of course, to make sure cybersecurity can control everything to make sure that no patient data ever is compromised. And now we can just also route these requests to our local models that are then, again, you could just think at a totally different scale if you employ these.

27:24Yeah, you mentioned cyber here now. So, I mean, certainly no conversation about AI is taking place these days without any type of, you know, cyber implication. What is, you know, what type of cyber initiatives did Eli Lilly bake into Lillipod from the get-go? How did you think about cyber going forward? And could you have imagined, even when you started this several years ago, that we'd be in a situation we are today where cyber is so much up and center?

27:52Dr. Thomas Fuchs: Yeah, it's absolutely dramatic. Again, not only because we are protecting Lilly's data, sometimes we are protecting the patient's data. And that is, of course, paramount. We are in an enormously lucky situation at Eli Lilly that we have a CISO, a cybersecurity officer who is enormously experienced, worked also in military applications for a long time and really defends us against state actors that are literally coming after Lilly, of course, every single day. And AI, of course, again, like all dual use technologies, helps the attackers, but it also helps the defenders. And our cybersecurity team under Andrea Abel is using AI in many, many different ways.

28:37Dr. Thomas Fuchs: to protect Lilly. And it's a great example. They even helped some of the frontier labs to better protect what they are doing because we are under such stress from, again, criminal organizations, state actors and so forth. And that team does a fabulous job. Again, an argument to also do more on-prem because you can control it completely. There are no dangers of leaks to third party. is or control by third parties what you can or cannot do. You've given a number of benefits and rightly so for the on-prem model. I'm curious, what are any of the risks or downsides that you're monitoring to make sure that you're not just over-indexing on one side?

29:24Dr. Thomas Fuchs: Yeah, so first of all, you really have to understand what you're doing and you cannot naively approach that. I had the fortune that I built clusters in the past in machine learning and healthcare in academic medical centers and in startups. And we have a great team of people who did that in the past. So you have to really understand how difficult it is, what's easy, what's not easy. If you have never done that before, it's probably too daunting a task. But the benefit is if you already did it, then you also understand that it can be done. And you see the benefits. You're not scared. Most people are very scared about it because it's new thinking.

30:06Dr. Thomas Fuchs: strangely for a lot of execs, right? For 20 years, everybody was trained, just go to the cloud. And now AI, of course, completely changes that equation. Then besides cybersecurity, of course, which is important, it needs a lot of software and different approaches. So our machine learning scientists and everybody is trained not just to use, for example, interactive notebooks, which would be a waste of that resource. You have to properly submit your workloads and jobs So you can scale across GPUs and nodes, and so everybody has to be trained to do so. And then I think these are the main downsides.

30:51Dr. Thomas Fuchs: It's really education. Of course, you also make sure that we have enough load, which is easy to produce, that the resource is really used to its maximum constantly. but we have a long queue of jobs that are waiting to get executed. So that upside certainly outweigh the downsides. For our scenario, that's something I want to emphasize. There are a million, million use cases where it's much easier for a startup or a company and more meaningful to use a cloud provider to do their work. But if you want to do science at the edge, at the total top of what scientific exploration with AI looks like, you need to do it and control it yourself at that scale on-prem.

31:39I mean, you talk about the forefront of scientific exploration here. I'll just ask, what is kind of at the forefront? What is the bleeding edge of what is possible with AI as it relates to scientific exploration, medicinal exploration, and what Eli Lilly can accomplish on behalf of humankind?

31:59Dr. Thomas Fuchs: So it actually leads back to your first question. It's a means to an end, and the end is very, very clear. So one area, for example, genetic medicines, rare diseases. That's an area where Lilly is investing a lot. I think we all think a big part of the future is in genetic medicines, and that just needs a total new kind of model. So that's not something where you just take an off-the-shelf machine learning model and throw in your data. So you have to come up with completely new models to handle that. You can start or add nucleotide language models or other structural models, but the teams are constantly developing new models to do that.

32:43Dr. Thomas Fuchs: That's why it's so exciting. And then you can use language models and agents to orchestrate all that work and then use these tools or these specialized models for the heavy lifting. And that then gives you an ecosystem of agents and tools to do that at scale. So that is very exciting when you pair it with automated data production. So data factories, automated labs. and both together, so really that authentic science, if you want, or really an AI-driven loop for science, that is enormously exciting because that pushes you to insights you couldn't do before. We have examples where molecules were co-designed with AI and the model came up with new motifs that were very interesting for the chemists and now they're exploring the same motive in other circumstances of other molecules and so forth.

33:38Dr. Thomas Fuchs: And that's exactly what you want. It's the same way Magnus Carlsen is now using moves. He learned from AI. Our chemists are using moves. They learned from AI. So it's really a partnership there. So that's an exciting part. Another one is reinforcement learning at scale. Again, that comes back to what makes Lilly special, and that's our people. and there is, I think, I couldn't think of another company in the world with that kind of team at scale and certainly not in the tech world where you then really compare domain experts with these models to make the models better so the models can serve the chemists better.

34:20Dr. Thomas Fuchs: At the end, AI is a fabulously amazing tool, but it has no will of volition. It has to be yielded by humans. And I think these are the front lines where we really push the envelope. And then if you think of ROI, that's, of course, manufacturing, right? The more devices and drugs we can get out to patients, the better for the patients and, of course, the better for the company. And optimization in manufacturing yields by far more ROI than our whole operation costs. You took over as Chief AI Officer in 2024. So you've had a couple years of doing this now, but it might still be a new title or role for a lot of listeners out there, a lot of organizations out there.

35:09Walk us through some of the thinking about why organizations might want to consider creating an office or a Chief AI Officer, what the role means and how the role might evolve moving forward.

35:24Dr. Thomas Fuchs: like many roles or titles it always depends how you fill it and what you do with it here at lily it's truly horizontal across everything we do i had the enormous fortune to inherit three fabulous teams and we did a restructuring it's run more like little startups so we have dedicated squads between five and seven people where you have machine learning phds ml ops engineers, software engineers, but then also domain experts, chemists, biologists, line engineers in the same teams, and they go very, very deep in these areas. And we are taking on the projects that are difficult, that need these kind of experts, and are very prone to fail.

36:07Dr. Thomas Fuchs: At the end, if you really do machine learning AI at the front line, it's the scientific loop. You fail, you fail, you fail, you fail, and often you have nothing to show but a better understanding of the underlying problems. And that's how you make progress there. And when we find something that is really useful for the company or for the patients, then we have excellent software development teams that can take that over and business and product teams to roll that out and scale out these applications and use these models. We also offer these models to the broader world through TuneLab where biotechs can use our models and also help to make them better in a federated way.

36:46Dr. Thomas Fuchs: and so that whole ecosystem is necessary to really roll out this the applications throughout let's say there's a listener or an organization out there right now that's considering adding the chief ai officer role from your perspective having done it like i said for a couple years now what type of advice or guidance would you give them on hey here's you might think it's this way right now but you get into this position and here's what you're here's what you're in store for?

37:17Dr. Thomas Fuchs: Yeah, so first of all, the unfiltered answer is, and we talked a lot about chemists and biologists, right? We couldn't do that. You cannot do small molecule design in AI without the chemists who know what they are doing. Because every scientific endeavor looks simple until you actually dive deep. You need this chemist. And the same way we respect these disciplines in chemistry and biology or in medicine, that also has to be done on the chief AI side. So your CMO should be, and it always is, a medical doctor because you have to understand your field. Sometimes you see the mistake that companies hire chief AI officers who are not machine learning trained experts for decades.

38:16Dr. Thomas Fuchs: And the main reason that is a problem is if you are not grounded deeply in your science, then it's easy to be swept away by all the hype and fluff that goes on, especially in AI these days. And then you become the play ball for all kinds of third-party company PR and vendors and whatnot. So to really change how you do AI in an organization, you have to understand what is real, what is not real, what is just PR talk to race the next round, and what LLM actually can do and cannot do, how it will change the world or not change the world. And the same way we respect that expertise in all these areas, it also has to be respected in machine learning, of course.

39:11Well, what's your own personal framework for cutting through that hype? How do you identify red flags versus how do you identify where there might be some substance and you might need to pull a little further?

39:25Dr. Thomas Fuchs: so i have the benefit that i have excellent team so i learn most of course from my people be it and in in in computational biology or pathology or or machine learning and all these other areas and if you're deep into that then you you see again what works and what doesn't work i know that's very very difficult for the general population what's very important if you look for voices out there is again look who really knows what they are talking about for a long period of time right today ai it's like soccer everybody's talking about it everybody has opinion and nobody is actually playing it in the world-class league right very few people and that's how ai is and then if somebody is is talking and bragging about the ai skills and they couldn't spell ai three months ago, then maybe that's not the best source.

40:19Dr. Thomas Fuchs: Or if a company tries to raise the next round, which is their job to oversell what they are doing, they also might not be the right source or they want to go public, then their statements might not be the right source to really ground your knowledge, especially if you are responsible in our case to build AI that at the end of the day impacts patients. So that's a big responsibility I think a chief AI officer has also to ground the company, to not get swept away by that hype and help everybody internally to differentiate the weed from the chaff. and if you do that right then you can still keep and even get more excitement about ai throughout because the people now know what it's real what what is real where they really should put their energy behind instead of being constantly anxious because they hear the next hyperbole that all white collar work will be replaced i think last year was the prediction already didn't happen will not happen this year or next year, or this or that type of ball that has nothing to do with what you can actually do with the eye.

41:35Dr. Thomas Fuchs: And that is, of course, change how we learn, how we teach, how we work completely, but in a meaningful way, in a grounded way. In our case, also at a very large-scale way. But like all science, it just gets more interesting if you strip away the fluff and the hype and the BS and focus on what's real. Yeah, without knowing exactly, you know, what level of internal resistance to AI Eli Lilly may have faced, I think it's fair to say every organization has some pocket of resistance to AI. I would suspect Eli Lilly may be less than others given its scientific background. But how did you and your teams go about trying to drive that adoption, make sure people understood, felt comfortable, had trust in the systems?

42:23Was that a playbook that you already had in place or did you have to build that on the fly?

42:29Dr. Thomas Fuchs: That's something, of course, I'm dealing with my whole career. So that was, as you can imagine, think about AI 10 years ago. That was really an uphill battle or 15 years ago. To convince a physician to use AI 15 years ago, good luck, right? So of course, I prefer to live in these times than in the times back. So we really bruised and scarred of doing machine learning for 20 years. So it all seems easy these days. But the way to make it work in organization is, of course, it has to be let from the top down. So when I came here, Diogo Rao, CDI, already had AI employed at a scale I was flabbergasted to see.

43:11Dr. Thomas Fuchs: I did not expect that. And I could expand on top of that. And then Dave Riggs is so amazingly tech forward he runs his own open source model at home tell me tell me another trillion dollar ceo who is capable and willing to do that so working in that kind of environment right is with all these open doors makes it so much easier that does not mean we need we do not need very exact and detailed business plans and have to hash out until the last detail when we want to make an investment, why we are making that, what it will yield for the company and the patients at the end of the day. To some degree, it's even more difficult because all the leadership of Lilly understands what you're talking about.

44:01Dr. Thomas Fuchs: But on the other hand, they also see the promise. And now we have, over the last two years, enormous impact. So we could easily show that what we build by far brings more return of investment than it all costs, orders of magnitudes actually, that you get more trust and you can go further. And that's necessary. Again, at the end of the day, the only right Lilly has to exist if we make life of patients better. The only right my role exists if I can contribute to that. The only reason we are using AI and building compute clusters is if we can show we do that. If we do not do that, we shouldn't exist.

44:45Well, fantastic. Thomas, thank you so much for your time here today. This was a great conversation. Thank you for the partnership. And hopefully we'll have you on again soon. We'll be talking about the next wave and iteration of AI. There will probably be something new here in the next couple of weeks, I suspect.

44:59Dr. Thomas Fuchs: Thank you, Brian, for the thought-provoking conversation. It was a lot of fun. Okay, thanks to Thomas for joining us today. The lesson here is simple. AI creates value when it's treated as a scientific and operational capability, not a technology showcase. The infrastructure matters, but only when it expands what experts can discover, build, and deliver. This episode of the AI Proving Ground podcast was co-produced by Nas Baker and Kara Kuhn. Our audio and video engineer is John Knobloch. My name is Brian Felt. Thanks for listening. See you next time.

45:37Thank you.