Talip Uçar
Inside Boltz's open-source bet on the future of AI-driven drug discovery

Nicola: Hi Talip, great that you made it to this episode of Models & Molecules.

Talip Uçar: Thanks so much for the invite, Nicola. Really good to see you.

Nicola: I want to talk to you about what you guys are building at Boltz. But before getting there, I would like to look a bit at what you have been doing before that and your journey. For instance, we met while you were working at AstraZeneca, but I know that before that you even worked on NVIDIA. So how did that all come about, what brought you from one company to the other and now, and now to Boltz?

Talip Uçar: Yeah, that’s a very good question really. I was thinking about it today actually. I always make decisions based on the problems that interest me and the problems that are challenging. And I choose the path based on that alone and nothing else, it’s not about the titles, it’s not about money. When I was at school, I basically got into chip design and that was really interesting at the time and very, very challenging. So that kicked off my career, and I ended up basically working in Silicon Valley for so many different companies, including NVIDIA. So, I worked on the designs of GPUs and CPUs. In fact, I even worked on the quantum computer or the parts of the quantum computer.

So that was really exciting at the time, but then I felt that at some point, technically speaking, I didn’t have much to learn. I was sort of bored. And at the time, AI sort of exploded with AlexNet. And I thought that AI is the future really. I decided to basically sell everything that I have and go back to school, get a formal training in machine learning and move to the UK, and start all over again.

Nicola: Wow, that’s quite a jump, right? And also, because indeed you were a hardware person, right? So, designing chips and…

Talip Uçar: Yeah, it was a completely different domain, completely different… I mean, I changed both locations and the career and everything. So, it was a big, big change in my life. And it’s just the basically prospect of working in this new field, which was very exciting at the time. And it is still exciting. So basically, I gave up everything and then just started all over again.

Nicola: Maybe one question. Because you’re saying that, you thought AI was very interesting at that point in time. How did you come to the conclusion “This is very interesting or this is the future” or what did you feel in that, in that respect?

Talip Uçar: At the time I was at NVIDIA, so when AlexNet happened, I was at NVIDIA, so I was exposed to all these different ideas and where the future is going basically. It was so clear to me that in every layer of life, there will be AI basically. It was just a matter of time. The sooner I jump in, the better it is, I thought basically. And some of my friends and colleagues, they also started moving to AI, but they chose to stay in the hardware field. So, at the time, basically one popular thing to jump on was these AI chips or AI-based hardware. So, a lot of my friends still stay in the same domain, sort of started to transfer to AI in that same sort of framework. But to me, the most interesting part was the algorithmic side. So, I decided to basically just change the entire field and just go back to school and start all over.

Nicola: Brave choice, I think.

Talip Uçar: I don’t know. Looking back, yeah, it was brave. Would I do the same thing now? I’m not sure.

Nicola: But that’s very funny what you’re saying because sometimes also people ask me, “Why did you do a certain choice in the past?”. And sometimes I reply, you know, it was also a bit of naivete. You know, I did not know really what would have entailed a certain choice, but I just felt that that was the right choice at the time. And with a bit of naivete, I did it. And it seems that you are also, you know, saying something like that. I don’t know if I will do it again, but at that time I felt that was the right choice.

Talip Uçar: Yeah, I think you really need that. You have to be naive about things otherwise you sort of refrain yourself from doing anything.

Nicola: Yeah, absolutely. But I think you did also another jump, right? Because not only did you go from hardware to software to algorithms, but you also eventually jumped to very different domain, biology and such. So, which is in a way another, how would you say, magnitude of a jump. How did that come about?

Talip Uçar: Yeah, so when I was coming out of school I decided to stay in the UK and not to go back to Silicon Valley. Because I really enjoy staying in the UK. And when I looked around, at the time, AI in healthcare was still sort of in its initial winning. Like people were questioning whether AI will be useful in the context of healthcare. And most people basically were skeptical of it, and they were sort of right at the time, actually, because deep learning-based algorithms and models, they didn’t really work well with the healthcare either. Most classical methods, they tend to work better. So basically, deep learning-based approaches, they were still at the beginning, so to speak.

I thought that basically eventually things sort of come around and this, I mean, if I have a good shot, basically since I am starting over from scratch, healthcare would be a nice place to start because now you can actually pioneer things because no one is doing it yet. So that was the thinking. So, I got my first job at a startup based in Oxford, because I moved to Oxford, and there basically we worked on this end-to-end clinical AI pipeline. So basically, we were getting data from different NHS trusts and then anonymized that data. That was basically patient data, electronic health records. We anonymized that data, then trained models on it and gave insights about the patient population and do it in an end-to-end fashion. And that sort of exposed me to the problem, this messy data problem that we have in the healthcare space, especially in the clinic. After a short while, I was pulled into AstraZeneca by my former boss, Lindsay Edwards. And basically, he was building his AI team. So, I became his first hire. And then basically we started building this foundation of AI within AstraZeneca. In my mind, that was like one of the first teams who really started doing modern AI between pharma.

Nicola: When was that actually, roughly?

Talip Uçar: That was 2020.

Nicola: OK, so we can say that it’s five or six years now back there.

Talip Uçar: Six years ago, exactly. That’s sort of like some of the pharma started to invest in it.

Nicola: Quite early. So even before, let’s say the ChatGPT moment, there was already a vision.

Talip Uçar: Yes, not in all of the pharma, only like few of them. I can think of only like a few of them who had this initial start and conviction. And I think they made the right choice in my mind.

Nicola: Yeah, I think history is giving them certainly some credit. But I think even in the context of bio and pharma, you did a few jumps, right? Because you said that you started working on clinical health records first. But now when we met the first time we you were already working on molecules, right?

Talip Uçar: Yeah, so basically when I joined AstraZeneca I stayed on the clinical side of things. But then AlphaFold 2 happened, and it was clear that basically if AI will make an impact in life sciences, it will start from the drug discovery part. It was the most practical application. Data is cleaner and is sort of easier to collect, easier to control, and easier to build models. I decided to sort of move in that direction and lucky for me, AstraZeneca kickstarted a project called MLAB at the time, basically it’s called Machine Learning Driven Antibodies and Biologics.

The goal was basically to build this end-to-end pipeline where you start from data, and you have AI pipeline in the middle all the way to the lab. And I think that was the biggest project within AstraZeneca. So, I sort of forced myself into that project because I thought that was super exciting. Over the course of the project, we have done some amazing work, I would say. Of course, we couldn’t disclose it, but I think we were far ahead of many people. In fact, I should say actually at the time, I even proposed this to some of the leaders within AZ.

There are two ways to go about it really. One is you spin out the project and do this work under a startup and make AstraZeneca as your first customer or do it internally. To me, it sort of makes sense to do it outside because then you don’t have to deal with the legacy workflows, internal politics and all of that.

Nicola: Right.

Talip Uçar: Because we had all the ingredients: great scientists, access to lab, everything that you need, you already had it. Our main challenge was the politics and internal dynamics and so on.

Nicola: Of course, larger corporations are difficult to move and steer, right? So there are many components, many responsibilities, risk management, all kinds of different things that are different in a startup, in which you can be much more agile in making moves and taking some risks, which I guess bring us to nowadays, and your latest, you know, jump that’s in your career, but in your interest on joining Boltz. I guess this has part to do with that, I imagine, but I’m curious to know how you got to work with the guys at Boltz.

Talip Uçar: Part of the reason was again my frustration with like how things are being done in a big organization. It is slow and if you want to do something you need to get like 10 different approvals from 10 different people. So that was frustrating. And also, what happened was when Boltz was introduced it came as an open source that was the first open source model. And it was quickly adopted by everyone. So it was obvious that basically there was a market for it. Basically, market for a foundation model that can be used for structure prediction and so on. So that sort of opened my eyes.

But then what happened was then Gabriele reached out to me for getting feedback on Boltz. So I was like, instead of me giving you feedback, why don’t you come to AstraZeneca, present your work and then get the feedback from the scientists directly.

So they came in, they presented their work. It was a very popular session. I think a lot of people showed up. A couple of weeks later, they reached out to me again, basically saying that they are thinking about starting a company and so on, and asked for my help in some regards. And also, I was convinced that basically, like as I was talking to Gabriele and Jeremy and Saro, like I felt that we have the chemistry. Like you have to have the chemistry with people if you want to do a good job.

Nicola: Super important. I totally agree with that. Chemistry is as important as science when it comes to running a startup.

Talip Uçar: Exactly. And that sort of pulled me into it as well, actually. Like, if it didn’t have the chemistry, I wouldn’t go for it.

Nicola: You mentioned something that I found interesting. You said that you, when you moved internally to adopt Boltz, a lot of, or at least when Boltz was introduced, a lot of the scientists jumped on it. So, and this is something that I find interesting because I do wonder, for an open source model, how much easier and not necessarily open source, but any model, how much easier for a scientist to start adopting it? I can imagine that if people never used such an instrument, they might have a difficulty to approach it first. So, somehow, of course, being open source makes it in a way more accessible, but you still need to bridge a certain, I guess, usability gap.

Talip Uçar: That’s true, it depends on your level of comfort using these tools basically. Within AstraZeneca for example, the people who adopted Boltz first were the computational biologists because they have some skills to install these tools and use them. Not in the best ways sometimes because they don’t know how to, they don’t know some of the details, but they could use it good enough to get the job done so to speak.

But for anyone outside of computational biology, I think they had, let’s say, more difficulty, I would say. Hence, it makes sense to sort of build a platform on top of Boltz and make it available to scientists. So that was like a next natural conclusion. So, it’s not easy to install these tools and then use them, so why not build a platform and then make it available to everyone?

And another challenge is actually, even if you know how to install it and use it, you don’t have access to easy compute. By building a platform on top of, let’s say, a dynamic compute, you make it very accessible to everyone, really.

Nicola: And this is something that indeed Boltz is doing at the moment, right? So, because you know that there is a platform built on top of, let’s say the Boltz AI engine. So that’s currently in progress, right?

Talip Uçar: Exactly. So basically, we have been building the platform in the last, let’s say eight months, although guys started earlier, and it’s on a very scalable compute. So, we can scale up and down the compute depending upon the demand. And we make it available to scientists, research groups, enterprises, and so on.

Nicola: Which brings me a bit to the concept or let’s say the vision of Boltz, which I’m curious also to discuss a bit with you, right? In a way, okay, based on open source, it’s also for public benefit, I think type of company, which is very interesting. And at the same time also reminds me a bit about certain things about maybe how OpenAI started. So some sort of for public benefit kind of opportunity, but of course that changed through time. I’m just wondering how do we foresee Boltz maybe going forward and you know, what’s the vision there?

Talip Uçar: Yeah, I mean I can speak of the intention now. I don’t know how it’s going to pan out in the future. At least at the moment, our focus is really to build the technology and make it accessible to everyone. And that’s the important point actually. So basically, there are two ways to go about it. You build the technology, keep it behind closed doors, use it in-house, develop your therapeutics. That’s what companies like Isomorphic Labs are doing.

But you can choose a second path which might be more impactful if you want to basically cure all diseases in the shortest amount of time. That second path is really to build the technology, make it available to as many people as possible so that the people can work on multiple problems at the same time.

You can have this like a compounding effect. So basically, the scale is not linear anymore, but it is compounding. So, every time a new team uses our platform, it means that one more output, one more learning. If 200 teams use it, it’s even better.

Talip Uçar: But really the motivation is to basically serve the people, have that public benefit in mind. I can speak for myself and hopefully for some other folks as well within Boltz that we are not really incentivized by money. We really think that this is an important technology, and it shouldn’t be staked behind closed doors. The goal is to push the frontier and make it available to everyone.

Nicola: Which makes a lot of sense. And I was reading, of course, about Boltz and it’s nice because they have a section in actually your own website, which says, you know, how we make money, which I think to me is good because it’s good to create something for public benefit, but you need to be realistic on how that can be achieved. And in a way it needs to be sponsored. Otherwise, you know, it’s very difficult to create public benefit. So I think from that perspective, I can totally see it. And when it comes to the technology, where do you think that what’s the idea behind Boltz? So where, do you think, you know, you have an edge?

Talip Uçar: I don’t want to think of it as an edge, but what we want to do is basically enable people, design molecules, any modality basically: be it small molecules or biologics. It should be like a one-stop shop, is that the term?

Yeah, so basically, I think we are one of the maybe few companies who focus on all of the modalities, both in the small molecules and biologics. Most companies focus on just one modality or one problem, basically. What we wanted to do was basically to build technology across different modalities, basically unify it, unify the platform such that it can be used for predicting structures of any modality or designing any modality for any other modality basically. So unification is, I don’t know if it’s a competitive advantage, but we think that basically future is going be the one where you unify things.

In fact it goes back to AlphaFold 2 times. In AlphaFold 2, what problem did they solve? They tried to solve the problem of predicting the structure of monomeric proteins. Then they moved on to multi-chain proteins. Then with AlphaFold 3, did unified modalities. Now you could predict the structures of proteins in complex with DNAs, RNAs, small molecules in complex with proteins. So, there was this sort of unification going on for a while. With BoltzGen that we introduced last year, we unified a structure prediction with molecule design. So that story will continue basically.

Nicola: Yeah. But do you think that in the short term, this makes your roadmap more difficult or easier? Because I can imagine that if you try indeed to, I can think two things, right? I can think that on the one hand, you can borrow information from multiple aspects, right? And that can, you can learn better. On the other hand, your optimizing function or what you’re trying to achieve, it’s also more difficult to explore your landscape, you know, that you’re exploring with your model. So, I can see both ways. What do you think there?

Talip Uçar: There is definitely a tension there. So basically, you can learn across different problems. That’s what we are doing. And at the same time, it means that you have to divide your energy into multiple sub-problems, basically. So, you have to pick and choose your battles. And it’s a hard thing to balance, to be honest. So far, we have been able to do well. And we are scaling the company. So, we are hiring people to be able to address those problems at the same time, in parallel in a much better fashion. I think it’s doable. we have been able to do well so far, and I feel that we will continue to do well. Although, again, it’s a hard thing to trade off.

Nicola: Yeah, I can imagine also maybe touching upon the data aspect of it, right? I think if you are trying to, let’s say, learn from a single modality, then you, of course, need all the data from that one modality. Therefore, you need to generate a lot of data for that modality. Do you think therefore that you can also leverage the fact that you can get data from all these different modalities and learn from it rather than focusing and generating a lot of data from a single modality? Do you think that’s an advantage? So, what’s the data strategy for you guys?

Talip Uçar: When I mentioned that unification of modalities, so basically Boltz is sort of a unified model that can predict structures of different modalities. What you see in these machine learning models is that as you basically incorporate data from different domains and modalities, they tend to learn better. Models basically take advantage of commonalities across different domains and modalities. So that’s already a benefit. And when it comes to our strategy for data collection or generation is basically what we care is not the amount of data, but it’s the diversity and the nature of the data. You can have, let’s say, one billion data points on a single thing. That’s not really that important. What I would prefer is like much smaller scale data, but diverse. So that’s sort of what we are putting our attention and energy into. And we basically collaborate. We have some collaborations with third parties to acquire their data or have some partnerships, let’s say. We have some strategies which I cannot talk about, at least for some of them.

Nicola: But basically you are gathering some data from either partnership or generating it that allow you also to further improve the model, designing your own data sets, I guess, or something like that. Because I always wonder whether you can really advance a model just with, let’s say, second-hand data, for lack of a better term, or whether you do need to generate specific data to guide the training of the model.

Talip Uçar: So you have to do many different things. So even like the publicly available data, I don’t think we are able to maximize our gains from it. I think there is still some juice there to, is that a term? There are still things to extract from it basically. Be it like pattern data or be it like some other data in the public domain. I don’t think people are taking good advantage of it.

Nicola: From an algorithmic point of view, you mean? Or meaning that we can still improve or create better models to make more use of the data that we have? That’s what you’re saying?

Talip Uçar: There are things to gain from the algorithmic side, but there are also like… If someone basically puts some effort to record publicly available data and just spend the time to clean it up. Even that gives you a good enough mileage to push it a little bit further basically. We just have to do a little bit of dirty work.

Nicola: The curation part.

Talip Uçar: The curation part. I think that’s like a very much underappreciated side of machine learning. People just assume that the data is clean, or data is what it is, and then we should just train the model on it. But if people spend a little bit of time, just a little bit more time on the cleaning side of data, I think models can do a much better job of it, even the same algorithms basically.

That’s the public data side of things. But of course, you still have to acquire, let’s say, private data or generate private data for some of the problems that you are trying to solve.

Nicola: Maybe touching a bit upon those elements, because we talk about algorithms, models, data. What do you think is going to be the first one that will get commoditized, if any? So, we talked about infrastructure, data, models. Do you think that some of those aspects will be commoditized in the years to come?

Talip Uçar: I mean infrastructure is already sort of commoditized, is it not?

Nicola: I think in a way, I’m curious to hear your opinion, then I can share what…

Talip Uçar: Okay, so when it comes to infrastructure, most of the companies including us, rely on third-party providers for compute, for data search and so on. So, in my mind, that problem is already sort of solved.

Nicola: In a way, yeah, let’s say that there are many parties that can offer that, that’s a type of service because in a way, let’s say, as you are saying, it’s kind of solved. So, there are multiple parties that offer that, that there is not that much variance between the different parties maybe.

Talip Uçar: And, when it comes to models, I don’t want to say commoditized, but to be honest with you, to solve our current problems, you can solve most of our current problems with the current algorithms and models, it comes down to data. Like we don’t have the good data to solve those problems to begin with. So, I think data will be the one that will give you competitive advantage over everything else. I don’t think it will be the models. OK, I take it back.

You can, let’s say, DeepMind, when they came up with AlphaFold 3, if they didn’t disclose it, probably it will be a competitive advantage for them. But I think eventually open source community would come up with something that’s sort of equivalent. So, it will take a little bit more time. But I think the ultimate in this space, in the life sciences space, the ultimate, actually this is true for mission learning, it’s not just life sciences. Data is the ultimate competitive advantage.

Nicola: Yeah.

Talip Uçar: Whether it will be commoditized, I’m not sure. Because that’s, like, if it is my competitive advantage, I wouldn’t, I wouldn’t, yeah.

Nicola: No, yeah, no, I, in a way, I totally agree with you, right? So, I think data’s a series of competitive advantages that provides, including the fact that if you have the right data, you can build better models. Or you also know where you can go with your training in a certain way, because data guides you in knowing better how to create better models. So that’s at least what I think about data. But also I’ve heard concept, for instance, of dark labs or fully automated labs, at which point, I guess, also the data generation, in a way, I’m not saying maybe get commoditized, but maybe we will have an Instagram moment for biological data in which we will be able to produce enough data to train the models. I don’t know, but I just, I find the interesting concept to think about.

Talip Uçar: Yeah, I mean that’s a really good question to think about. I don’t know. I like people always, like I see a lot of different companies working on data generation, some of them are high throughput, some of them are low throughput. And people try to optimize lab work. I think I feel at this very moment, actually, no one sort of cracked the problem. And I think it will take a while before we see it. I mean, there needs to be, let’s say, AlphaFold 2 moment for data generation. If someone comes up with some nice data generation platform, for example, that is high throughput and high quality. I don’t think it’s going to happen, but even if it’s the same medium throughput and medium quality, that will serve you well. That would already sort of enable you to solve a lot of different problems. But the question of commoditization, I don’t know again.

Nicola: I think what I was trying to think about is, let’s suppose what’s going to be the hard part, right? What is, would be the part that, you know, all of this we can give to a, maybe a vendor thinking about from a pharma perspective, right? All of this we can outsource, but this thing, you know, we want to have certain, you know, own it or something like that. That’s more or less what I was thinking, right? So, because it’s like, I don’t know, email clients are commoditized, right? So, nobody really cares which clients you’re using for reading your email in a way, because it’s a solved problem, as you put it. And I think at different times, there will be maybe different type of commoditization or different parts will be commoditized, but.

Talip Uçar: Okay, so I okay I think I’m sort of understanding what you’re trying to say. Let’s say from the pharma perspective, actually… overall, like how I think about this is like following: you have these different layers of the stack. You have data generation, you have the infrastructure in between, and then you have the compute and you have the science on top of it. There is already what do call a fragmentation basically or basically there are companies focusing on data generation alone, and there are companies focusing on infrastructure alone, and model development like us basically building the platform on top of it. And I think the pharma, like some pharma have been trying to build everything themselves. I think they might come to the realization that building everything yourself is not the most efficient way to do it. And therefore, maybe they should just focus on science, basically, and license everything else. I don’t know how things will sort of shape in the future, but as far as I can tell, there will be less ownership of different layers and across them basically there will be some cross talk.

I always try to give an analogy to hardware. So, let’s say in the hardware space you have, let’s say Google or Dell. And when they build the hardware, they buy different components from different providers and then put them together to build the final product. But then you have the Apples of the world who tries to build everything themselves, the software, the hardware, the chip, and so on. And basically, do this vertical integration and both models seem to work well.

So maybe we will see something similar where, let’s say, some companies will basically try to do everything themselves because they can control everything. And if they can do a good job of it, actually, it might give them a competitive advantage like Apple. And it might work well. And in some cases, you will see companies, that will sort of outsource everything except like one small domain where they will sort of put things together and then specialize on that one basically. And that might work well as well.

Nicola: And I think it’s interesting because that brings me to another point that I was curious to know your opinion about because I have seen a couple of different attitudes from, let’s say, tech companies. You mentioned Google, but you we can mention AWS, big tech companies, and pharma companies.

For instance, the other week I was in New York at AWS Symposia and there some people came up with a very interesting definition at the end, defining their company and they said: “We want to be the technology company of today to be the medicine company of tomorrow”, which I thought it’s a very interesting approach. On the other end of the spectrum, you see AWS or, you know, Anthropic, for instance, they just acquired this, this new company for, for a crazy valuation. They want to enter into the biotech world and the medicine making company. So, these two worlds seem to collapse. What’s your take there, having been in both?

Talip Uçar: It is very interesting. These companies, they have to be tech companies, really. They have to adopt AI and engineering culture and so on to be able to do a good job of it. But whether they can do it or not, I have my doubts. Pharma doesn’t have the culture and the mindset to do a good job in this space, in this new space. You need a different mentality. So I think, I might be wrong, but I think that people who are coming from the tech side moving into biology, they might have a better chance, just because they can go faster. And also, they are so naive and that naivety might work in their favor.

Nicola: They don’t know how difficult biology is. So, in a way they will try.

Talip Uçar: Yeah, I mean, I ran into that problem when I moved into life sciences. I was so naive. And probably I am sometimes. It sometimes gives you this competitive advantage basically.

Nicola: You just try.

Talip Uçar: You just try. Exactly.

Nicola: You apply it. Yeah. And you see what you can achieve. So, I think that that’s a great attitude to have.

Talip Uçar: Exactly. And also, the people with this tech culture, tech mindset, they are quite flexible. They can adapt much faster. Versus in the pharma, again, it’s a completely different thinking really. Like things are slow, people shy away from things, they don’t try things, they play it safe.

Nicola: Yeah, much more risk-averse.

Talip Uçar: Exactly. They, just play it safe and that’s the culture. Hence, it doesn’t play well with this tech culture or AI. Let’s say, within big organizations, you will see this clash between the two cultures.

Nicola: Which brings back to you joining Boltz and trying to accelerate that in that respect.

Talip Uçar: Yeah, I mean, even like when I started working with the guys, I sort of knew it already, like even before I joined, even when I was at AstraZeneca, because I came from tech background, so I already knew it. But it has become even more clear to me that if you want to work at the intersection of AI and bio, you have to have that tech culture, tech mindset first. Like the way that we work at Boltz, for example, and the way that we approach problems is completely different than how one would do in pharma.

Nicola: Can you give an example? Because I think that would be quite interesting.

Talip Uçar: Since I joined guys eight, nine, ten months ago, I have run basically maybe, I don’t know, 10x more experiments than I have done at AZ in the maybe in the last three-four years in terms of in terms of model development, basically running experiments for model development alone. If I had to do it in pharma, for example, I had to ask for funding first and then wait for, I don’t know, like for the annual budget to come out, blah, blah, someone needs to give approval, and you don’t even know if you’re going to get it. It’s a different world. Here, if I want it, I just say, hey, I want to do this. And then we just do it. So, it’s a completely different mindset.

And same thing goes with model development. So, on any given day, we try 10 different things. And we see what might work. And then we just go in that direction. Basically, it gives you opportunity to iterate faster to see what works, what doesn’t work, and then you just move, and move, and move. There is no way you can do that in the big organizations or in the big pharma.

Nicola: Yeah, I guess it is double, right? So, in a way it’s being in a big organization, but also being in a sort of a risk-averse organization, traditionally like pharma. And the combination of the two, of course, plus, I guess, the tech mindsets, when it comes to adopting these technologies that are definitely changing very rapidly, right? So yeah.

Talip Uçar: And also one thing I realized, again, I knew this one before. It just made me sort of convinced that the compute is the ultimate enabler. At this day and age, if you want to have the competitive advantage over everyone else, it’s really the compute. Because it allows you to run so many different experiments at the same time. And then you see, let’s say… Again, it gives you an insight in terms of what might work. Then you just, by the time you are, let’s say, 10th experiment, you’re already like far ahead of everyone else, basically. Versus in pharma, for example, I had like much, I had, as a group, we had like much less access to compute, which is sort of a disadvantage if you want to compete in this space. Again, as a small company, as small as we are, we have enormous resources to do well, basically.

Nicola: Yeah, which I think is great. Maybe because we are running a bit out of time, but I’m truly honestly interested to hear your opinion about a few more things. So, I hope that we can we can still continue for a minute. And one of them is the let’s say, let’s call it I don’t know, agentic vibe or agentic fever that that we are seeing. So, if I go to conferences, there is a lot of talking about using different types of agentic workflows or things like that. How do you stand about that and how Boltz maybe is thinking around those elements?

Talip Uçar: Yeah, I think it’s great. To be honest with you, I don’t like hype, but I see a lot of benefit to these agentic workflows, especially for orchestration and to basically increase efficiency in the in the workflows. I think they are quite useful, and they will be quite useful. I mean, we will be making some announcements in the upcoming weeks in this space as well.

Nicola: Interesting.

Talip Uçar: So, we definitely think that it is useful as an orchestration tool. But it’s not going to solve the problem. Basically, these things are just tool users, orchestrate things basically. At the end, what differentiates you is the technology that you are building underneath.

Nicola: A productivity augmenter, but you need to have the basis to leverage those.

Talip Uçar: Exactly, yeah. But I think that’s great. Even between Boltz for example, we do use agent workflows for software development, even like in some cases, partial to model development.

Nicola: As I said, I was just generally curious to know your opinion, because it’s something that also for us, it’s of interest. For us, it’s more about how people interact and run these workflows, right? Because for us as a company, of course, we are very much interested in giving the people the tool to run workflow in the same automated way or whatnot. So, we’re really interested in what we have seen is that there is already a shift in which people are moving maybe a little away from certain interfaces and having a more orchestrated-by-agents type of approach.

Talip Uçar: Yeah, I think like the technology itself is like interesting. What makes it useful is actually how you integrate it really, like the UI part of it. You might make a mess out of it if you don’t do a good job, if it’s not intuitive. But if you can basically put the pieces together, actually, it’s a very useful technology, I would say.

Nicola: Another random question, en passant. If you could choose more context or more inference speed, which one would you go for?

Talip Uçar: More speed… It depends on the part of the pipeline. Let’s say if I’m doing target ID, target identification, maybe I don’t care that much about the inference time, but I care about the accuracy and therefore maybe the context. But if I’m doing virtual screening, I care about the inference time and therefore, yeah, maybe less context. So, it is situational, I would say.

Nicola: We usually end with the contrarian question. So, which in this case is, what’s a belief that you have about AI and the current industry that maybe most of your peers will not agree with.

Talip Uçar: For a long time, I thought that basically, I told you this one before, I believe that in life sciences, in the context of big pharma, what you really need is a unified data fabric, which means that basically connecting every stage of the pipeline where every team reads and writes from the same data layer. You want to do it for two reasons, let’s say. The first reason is that if you are designing a molecule that needs to be safe, that needs to have certain properties, that needs to basically work for certain subpopulation, you have to incorporate all that downstream data in the design process. For you to be able to do it, you need access to that downstream data.

The second reason is that the moment you design a molecule to be a drug, you want to be able to track the entire lineage, whatever attribute people measure on it or collect on it, we should be able to track it in the future.

Why is it contrarian? Because people think that it’s not going to be possible. The main problem is that basically there is this obstacle, the virtual obstacle between the drug discovery part and the clinic.

Usually, we cannot connect the data across because of privacy issues, patient data and so on. But I believe it should be possible. As I said, in my very first job, we built this clinical AI pipeline using anonymized data. Therefore, access is not an issue. So, I don’t see why it will be a problem to basically build this layer and then have all this data from different stages. And that’s sort of what you need if you want to really solve the problem.

Today, like in my mind, internet is the biggest data generator. It’s a unified data layer, so to speak. We need something like that in life sciences. I don’t know basically if we can solve any of the problems without a unified data layer, really. I don’t know if I’m making sense.

Nicola: No, I think it’s a great answer, to be honest, and a very good one. And I think indeed the problem maybe is more organizational than technological, but I agree with you, right? So, I think we need to connect the patient data, the final very end outcome with the beginning if we want to really improve the whole process.

Talip Uçar: One thing that’s really underappreciated is this nature of the feedback loop in different domains. For example, if you are working in vision or audio, I don’t need to be an expert to decide the output of a generative model is crap or not. If it generates an image, I know if this is good enough or not. So that feedback loop is super tight. So that’s why you see very fast progress in those domains versus that feedback loop is quite fragmented in life sciences because you have data generation, you have model development, you have experimentation, you have biological interpretation, and all of these different stages are owned by different domain experts. And they work in these isolated layers. And as you go across the pipeline, the information is transformed. Then basically what becomes important is that basically you try to keep the signal preserved, so to speak. So, if you don’t have like a, let’s say, very well-connected pipeline, your model might perform well locally. It looks coherent locally, but it breaks apart the moment you try to connect these different stages. And that’s the problem actually. It’s a system-level problem. So, if the system fails, it may not be the case that your model failed, but maybe something else failed. You don’t know.

Nicola: It’s super insightful. I really never thought about this thing in which in the for instance for imaging, everybody can tell immediately, “Okay, this is not good enough” or you know “this is better than that”. But yeah, of course, in such a process, the feedback is so much further away that it might never come back. So yeah, super insightful.

Talip Uçar: Yeah, I mean most of the time you don’t even get it back. It goes into a black hole.

Nicola: Exactly, exactly. All right, I think we need to go. But it has been really a pleasure to have this conversation with you, Talip. And I hope that at some point in the future, we can have another one. And perhaps you can keep us up to date with what you guys are building at Boltz. But so far, well, just thank you for joining us today.

Talip Uçar: Thank you so much again for the invite, Nicola. This was a pleasure.

Nicola: Thank you.