Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Silicon Valley GirlSilicon Valley Girl

AI’s Next Trillion-Dollar Market Isn’t Software | Max Jaderberg, President of Isomorphic Labs

AI can generate thousands of ideas. I asked Max Jaderberg how to tell which ones are worth pursuing.

Max JaderbergguestMarina Mogilkohost
Nov 10, 202643mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

0:00 – 3:02

Intro

Max Jaderberg 0:00 With AI, I believe you need that big ambition and that biggest mission possible.

Marina Mogilko 0:04 This is Max Jaderberg, an Oxford-trained AI researcher whose startup was acquired by Google. He helped develop AlphaFold 3 and is now president of Isomorphic Labs, founded by Demis Hassabis to design new medicines with AI. So AI can generate a lot of hypotheses. How do you know it's actually worth pursuing?

Max Jaderberg 0:24 Instead of having to screen experimentally, go through hundreds of thousands of different designs, we can make 20 or 50, and then we go into the real world and actually make these molecules and test them out and see if they work.

Marina Mogilko 0:36 So when do you think we're gonna see the first drugs designed by AI at a pharmacy?

Max Jaderberg 0:41 Well, it turns out-

Marina Mogilko 0:44 Max, welcome to the podcast.

Max Jaderberg 0:46 Great. Thank you for having me.

Marina Mogilko 0:47 So happy to have you. Uh, you're on an amazing mission, uh, to solve all disease with the help of AI. Can you talk to me, what have you seen inside your work that made you believe that this is all possible?

Max Jaderberg 1:02 Maybe I'll rewind to, to sort of s- say a bit where we began, um, and where we, like, started to see that actually this crazy mission of solving disease is, like, one we can actually go after. My background, um, has been in, in, like, core, core AI, core deep learning development. I started my research in computer vision and, um, like, uh, things like ImageNet, how to train these big networks to get computers to see. Um, I spent a long time at DeepMind working on all the general algorithms, new architecture design, early days of generative models, language models, and deep reinforcement learning. And through that was working on these, like, really hard challenge tasks like, uh, AlphaGo and AlphaStar. We were able to solve these, like, impossible challenges again and again, and maybe we hadn't got to AGI yet, but it was very clear that we, we had this new technology with AI that w- that was able to do, like, very different stuff. We're starting to think, "Okay, well, what should we really use this for?" Like, what could be the impact on, on humanity here? And at the same time, we start seeing the AlphaFold results internally in DeepMind starting to come to fruition. So this was amazing project of trying to predict the structure of proteins, which is-

Marina Mogilko 2:17 Mm

Max Jaderberg 2:17 ... a very, very hard and fundamental task of science. And we could start to see these results emerging of this AlphaFold network starting to get really, really good. And this led to, like, a real conviction that you can actually train AI models that generalize in the space of biology and in the space of chemistry, and no one had really shown that before. And I think that proof point of AlphaFold was really-- a- allowed us to take the leap of faith to say, "Hey, we're at the right stage of the development of AI to really actually now start something in Isomorphic Labs to go full force towards using this to really solve human health." And we really think about solving all disease as that core mission for us to go hard behind.

3:02 – 5:38

Why Developing a Drug Can Take 10 Years

Marina Mogilko 3:02 So from this AlphaFold moment to solving all diseases, what has to happen?

Max Jaderberg 3:06 AlphaFold isn't the solution to making medicines, to solving diseases. It's really that proof of concept that a solution could exist. You know, the-- to actually make a medicine to help solve a disease, it requires so many different things to come together. It's not just one problem. It's like a dozen really, really tricky problems come together. That's why it's so hard to actually make medicines. It takes actually quite a long time to make a medicine. It takes 10 years to, to go from conception at the moment to, uh, a medicine for patients, and the failure rate is really high. About 90% of medicines fail through development.

Marina Mogilko 3:42 Because we don't know how they're gonna react, like how the proteins are gonna react with the molecules, and this is where AI is helping, as far as I understand this, right?

Max Jaderberg 3:49 Yeah. There, there's, there's so many different dimensions that come into play and which are hard, so yes. How do you actually go about solving that disease? What usually protein target do you need to go after? What's the, what, what's the fundamental pathway and little molecular machine in the body, in the cell that you want to go in and modulate to make the disease state better? Can you actually make a molecule that's good and that works in, in the way that you intended? If you take that molecule as a drug, will it actually get to the right part of the body? Will it be safe? Will it have toxic side effects? There's just such a long list of things that you need to design for and need to understand, um, and it's really, really hard. And so the solution from a AI perspective and a model perspective is not, is not actually just one model. It's like a whole collection of AlphaFold-level breakthroughs. We think about maybe half a dozen plus AlphaFold-like breakthroughs-

Marina Mogilko 4:42 Mm. That need to happen

Max Jaderberg 4:43 ... that we have to-- They need to happen-

Marina Mogilko 4:44 Mm

Max Jaderberg 4:44 ... to then come together into one holistic drug design engine. We've been working really hard over the last five years on just systematically going through those half dozen plus AlphaFold-like breakthroughs.

Marina Mogilko 4:55 But have you identified them?

Max Jaderberg 4:56 Yeah, I mean, we, we know what those should be. We, we, we, we think about, okay, we wanna make medicines in a new way. What are those, like, core superpowers, those superpower capabilities, that if you could just give this superpower to a chemist, to a drug designer, a biologist, it would radically improve our ability to make medicines? So we can really identify what those are. That comes into things like understanding how cells work and different parts of disease biology work, how to desi-- like, different aspects of designing molecules, different aspects of understanding human toxicity. And then we can just go systematically after those to cr- to create those superpowers. So we, we've identified those, and we've been building the models

5:38 – 10:22

How AI Helps Scientists Design New Medicines

Max Jaderberg 5:38 towards that.

Marina Mogilko 5:39 How is AI helping you achieve what you want?

Max Jaderberg 5:41 It's really on two fronts. There's, like, the understanding front and then the ability to design new medicines front. So to make a medicine, you need to understand the disease biology first and foremost, and, uh, biology is a really, like, beautiful and complex beast. Um, there's just so much complexity. We know such a tiny- proportion of what there is to know, I think, about, about the science of biology and disease biology and how our cells work, and crucially, if you have a disease, how that changes the normal function of cells. And so we use AI and we create these models. These are like foundational models of biology that help us understand how, how these systems work, so it can help us understand how different proteins and the different building blocks of life, proteins, DNA, RNA, how they all fit together, how they form pathways, which are the sort of signaling mechanisms for cells. Uh, if you have a disease, how that changes how these pathways operate, and therefore, that gives us a clue into how we can go in and maybe fix the situation. So that's one side of the equation is to understand how we work, how cell biology work, how disease biology works. And then the other side of the equation is, okay, how do, how do you then design an intervention to that? So that, that's a medicine, and a medicine is something like a molecule, like a small molecule you could take as a pill or something called an antibody or a peptide that you inject. Um, how do you actually design a molecule that will go and modulate your biological system as you intended to solve that disease state? And so really, AI is helping on those two fronts really, really strongly for us.

Marina Mogilko 7:23 I'm trying to understand from the business perspective, 'cause what you're describing looks like very long-term research. How do you break it down into business goals, and how do you know when to commercialize what you're doing?

Max Jaderberg 7:36 Great question. When we started the company, we, we didn't want to go and sit in a dark cave and work away on the, the core engine and models for a decade and then pop up with the solution, "Now, hey, we're ready to commercialize this." The whole idea was to, from day one, be developing the core technology, the AI models, the platform, and then also from day one to be able to, like, really try and use this for the, the real hardest commercial problems. Very early in the company, we were doing real drug design, not just work that we instigated ourselves, which is really important, but also going out and just looking for the best people in the world, the best pharma companies, sign partnerships with them, where they brought us literally, like, the industry's hardest problems. So you can't mark your own homework. You're really held against, like, the top standard of the industry. Um, and that was tough, but it really forces-- forced us to develop our models and our AI drug desi- design engine in a way that's gonna really speak to true commercial problems. Um, and so that started our commercial, um, footing very early on in the company, and we think about sort of two aspects of, of how we do drug design. We do internal programs, making medicines that we instigate ourselves, we fully fund ourselves, we own all of the IP around, and we'll develop these, these m- molecules ourselves. And then we work with partners to also do this, um, in, in partnership. Uh, so that builds two different commercial models I think that work really well together.

Marina Mogilko 9:12 There is one thing we still haven't solved perfectly on my remote team, and it is having one place where people and AI can actually work on the same project. Right now, research lives in one place, old scripts somewhere else, producer notes in another docs, and Claude is usually in his own chat. So every time we start a new episode, someone ends up copying old angles, questions, examples, and context back and forth. That's why we've been testing Miro, who is sponsoring this video. We put everything for the episode onto one Miro canvas, research, past scripts, audience feedback, questions, and decisions. Then we connect Claude to the same board through Miro's MCP server. So instead of the producer working in one place and Claude in another, they're working from the same project context. Claude can read what's already on the board, work with it, and write the output back there. The producer can review it, add her judgment, change the angle, and keep moving without constantly digging through old chats and documents. For us, it means less copying and re-explaining and more time actually working on the episode. If your team already uses AI, but the work is still scattered across chats and docs, check out Miro's MCP server. Link is in the description. Thanks to Miro for sponsoring this video.

10:22 – 13:12

How to Tell When AI Is Ready for Real Work

Marina Mogilko 10:22 Do you have a, like, a business framework when you look at these new technologies? How do you know they're ready to be commercialized? Like, when do they leave the lab-

Max Jaderberg 10:31 Yeah

Marina Mogilko 10:31 ... and become a product?

Max Jaderberg 10:32 In general, I think this is different industry to industry. You can think about something like self-driving cars, where you have to get to, like, ninety-nine point nine nine nine nine nine percent accuracy before it's safe enough to deploy and it's actually, like, a commercial product. In the space of drug discovery, we're actually not very good-- before AI, we're actually not very good at predicting how good a molecule is gonna work. The status quo of the industry, it's a very empirical process. It's sort of trial and error in the lab process, which means that actually you don't need to have perfect predictive power before you can create real value, scientific value on your, on your drug discovery programs. So we really think about what is that level of predictive accuracy that really changes the game for drug designers and helps us unlock problems that other people wouldn't be able to work on. And so when you hit that level of accuracy, and that ac- level of accuracy actually doesn't need to be very, very, very high, it can be just, like, a significant delta above what the rest of the industry might have access to. You can then start using that internally. We have this framework where we're doing this core, uh, frontier AI model development. We have lots of benchmarks and evals that act as, like, proxies for how good these models are, and then we just aggressively translate these onto live drug discovery programs to see that they-- and make sure that as we're making progress on our internal benchmarks, our static benchmarks, that translates into real scientific breakthroughs and therefore, you know, economic value on our live discovery programs. Um, and oftentimes you can see that You're making progress on your machine learning benchmark, and it's not translating into reality.

Marina Mogilko 12:13 Yeah.

Max Jaderberg 12:13 And that's because there's a disconnect-

Marina Mogilko 12:15 Mm

Max Jaderberg 12:15 ... between this proxy of your machine learning benchmark or your eval and the real problems, and you have to go back and redesign your leaderboard and your evals to better reflect the real on-the-ground situation.

Marina Mogilko 12:26 It's interesting. How do you manage your ambition? How do you think about that?

Max Jaderberg 12:29 In this space, and it, it comes back to making sure that we actually have the scope of ambition that's large enough, a business model that is expansive enough to mean that we can rationalize huge investment into our-- the ability to create new models and to create this drug design engine. And, and this is really back to the, like, the be- the beginnings of the company is really trying to assess, like, what is that big market potential? What's that potential impact for patients that could be, and how much of that can be realized differently now with AI? And that helps underwrite this amazing amount of investment into the core modeling and the core science, which we then can have faith will translate

13:12 – 16:58

How to Turn a Big Mission Into Measurable Goals

Max Jaderberg 13:12 at the end of the day.

Marina Mogilko 13:13 How do you stay, uh, true to your mission when the goal is so ambitious? Have you ever had periods when, you know, this is not working out, the technology is not there yet?

Max Jaderberg 13:22 Yeah, our mission is crazy. You know, solve all disease. It's, it's, it's pretty audacious. Um, we say that with a straight face. Like, we really, like, feel and mean that mission. We know it's not gonna happen overnight, so honest about that. We also-- Like, frankly, it's not gonna happen just within the four walls of Iso. But I believe you need that big ambition and that big, like, the, the biggest mission possible, because if you start reducing what you, what you think your scope of impact can have, you start building out the company in different ways, and you start designing different goals that don't quite, quite match that, that ambition. But then saying that this mission is too big for any one person to just hold in mind, how do you translate that into the day-to-day? So in my head, I think about, like, a mission tree. You have the big solve all disease mission. Okay, what's, what's one level down from that? Okay, well, we need to be, like, solving the way we make medicines. Like-- And then on and on and on. And so we think about missions throughout the company, even down to every individual. I want every individual to have their own personal missions that when you put these all together are speaking to our top-level solve all disease mission now. And, um, and then thinking about, okay, what are the milestones, like the practical milestones that we can start to measure to help us litmus test, okay, are we still on this path?

Marina Mogilko 14:39 Yeah.

Max Jaderberg 14:39 Are we pushing hard enough? Do we need to be doing something different?

Marina Mogilko 14:41 This is really fascinating. I was just talking to Greg Brockman, co-founder of OpenAI, and my key takeaway from that conversation was he told me we have to be working on our ambition 'cause AI makes so many things possible. And something that you just said, if you have lower ambition, then the company works in a completely different way, and you start solving for smaller problems, so you have to have that North Star. So do you have any tips for, like, the darkest days for founders when they're on a big mission, but everyone around them is doubting, and maybe even research is not proving that they're right? What can they tell themselves, and what can they do?

Max Jaderberg 15:15 First of all, I think that's why it's important to actually have a mission. Like, know what that is, and to have spent the time to rationalize that and be like, "Yeah, okay, I can fully get behind this and just drive hard on this." For me personally, the I- the Iso mission is one that I can just dedicate my life to and have no regrets-

Marina Mogilko 15:34 Mm-hmm

Max Jaderberg 15:34 ... behind. Um, d- doesn't matter how hard it is. I think you have to translate this into, like, some signals of life, if that makes sense, so that you don't go too crazy-

Marina Mogilko 15:45 Mm-hmm

Max Jaderberg 15:46 ... you know, trying to pursue something really big that's gonna take a long time. What are those little goalposts, even if they're artificial?

Marina Mogilko 15:53 Yeah.

Max Jaderberg 15:53 I'll give you an example. Like, we had this, um, particular modeling task, and it was one that we like-- has, has been there since the beginning of the company. And we sent-- we set-- We want to solve all disease, sure, but how does this team on this particular modeling task, like, feel that day-to-day? Well, we set this, like, crazy ambitious quantitative target. You need to hit eighty percent accuracy on this particular benchmark. Oh, and by the way, the world is only ten percent good at it.

Marina Mogilko 16:17 Mm.

Max Jaderberg 16:17 So it feels, like, so out of reach.

Marina Mogilko 16:19 Yeah.

Max Jaderberg 16:19 But it's still, like, a very tangible number. And day-to-day, you can see climbing one percent, two percent, three percent, so you're moving towards that. It still seems very far, but you can see progress against that. And actually, as you compound this year over year, you can actually start to reach these things. So I think always trying to translate that. We, we try and translate these, like, really big qualitative ideas into just something quantitative. Even if it's slightly artificial, it can help focus teams, focus the mind in just, like, a very gamified, competitive way. We literally call these leaderboards. And it doesn't matter what part of the company you're, you're in. You've, you've got your quantitative goal that you're trying to grind towards.

16:58 – 19:00

Where the Next AlphaFold Moment Could Happen

Marina Mogilko 16:58 I love that.

Max Jaderberg 16:59 Yeah.

Marina Mogilko 16:59 That's, that's very, very practical advice for founders.

Max Jaderberg 17:01 It's very practical. [chuckles]

Marina Mogilko 17:02 Yeah. Do you think, uh, there are industries that are ready for an AlphaFold moment like biology when you're looking around?

Max Jaderberg 17:11 Yeah. So outside of biology-

Marina Mogilko 17:13 Mm

Max Jaderberg 17:13 ... um, I think there's s- like, a lot of opportunity to sort of find places that AI can now be applied and that still require an AlphaFold-like moment. So what, what, what does this look like? We, we have all of this amazing advancement on our core LLMs, our reasoning agents. Um, these help us, like, understand data and reason about the world in, you know, the way we do but, you know, starting to far surpass us in terms of intelligence. Uh, if you think about the biology space, that could be just like a human biologist will look at the data, look at some experimental data, and try and draw conclusions for that and, and hypothesize about what, what experiment can be done next. Now these reasoning models are able to do that and starting to maybe even outperform people in doing that. But can they actually predict what the experimental outcome will be itself? Can they simulate the real world? Maybe not. Maybe not at this stage. Maybe one day they'll be able to train the models to do that as well. So I think there's gonna be a lot of, um, applied AI domains where the reasoning models will come up with hypotheses or help you understand and analyze the world, but you still A bottlenecked by going into the real world and poking the real world and trying to see what the response of that is, doing a physical experiment. That's the case in biology. It's the case in, you know, material science and almost like any-- and it's the case in, in any form of like physical design work that you're doing, uh, a lot of spaces of engineering. Um, and in those sort of cases, in, in those sort of domains, there's an-- I think there's an opportunity for AlphaFold-like moments, i.e., the ability to create these AI models that simulate part of that physical world so that instead of going and doing the physical experiment, you can poke your model instead-

Marina Mogilko 18:56 Yeah, instead

Max Jaderberg 18:56 ... and see what the response is. That's very complementary to the LLMs

19:00 – 20:55

How to Tell if You Have the Right Data for AI

Max Jaderberg 19:00 and reasoning models.

Marina Mogilko 19:00 And how do you know that the model is not ready to solve that yet? Have you ever had a situation where you have this idea-

Max Jaderberg 19:07 Mm.

Marina Mogilko 19:07 ... uh, but then you look at the model, and you realize the technology is still a decade away from solving it?

Max Jaderberg 19:13 Yeah. I think a lot of that comes down to what data is available. Um, so machine learning AI only works obviously with-- by training on data. I think you have to start by looking, okay, what is this predictive capability that's needed, and what data is there that might have some information content inside of it that could speak to helping predict that thing? And if you don't have much data, or the data is very, very noisy, or it's very low dimensional, there's not actually much signal-

Marina Mogilko 19:41 Mm-hmm

Max Jaderberg 19:42 ... in this data, then there might not be actual scope w- given the current data to make a lot of impact on this modeling task. And that could also be an opportunity for people, by the way, because then you should actually start thinking, okay, well, what data could exist to unlock this problem? And I think there'll be many problems in this sort of space where if you look at the current d-- like data that the world has access to, it's quite small or it doesn't have the right signal. It's too noisy. But maybe there's a new piece of data-generating technology or way we can collect data or generate new data that could then suddenly unlock new domains that haven't been unlocked before.

Marina Mogilko 20:21 So you think the models are powerful enough, so they can do way more than we're able to achieve just because we missed the data layer.

Max Jaderberg 20:26 I'm a deep learning maximalist. I think we have the algorithms that can solve almost any problem. We have the compute. Now we have just an abundance of compute. Of course, w- everyone's compute constrained, but there's so much compute available. So it really comes down to if we have the algorithms, we have access to compute, we've got amazing people that can put this together, then we need to m-- to-- we need to make sure that we have the data.

Marina Mogilko 20:48 Mm-hmm.

Max Jaderberg 20:49 And a lot of these amazing applied AI potential domains, they're limited by the data that's available to, to

20:55 – 23:09

Why Max Hires AI Engineers Without a Biology Background

Max Jaderberg 20:55 train on.

Marina Mogilko 20:55 That's a business idea.

Max Jaderberg 20:56 Yeah. [chuckles]

Marina Mogilko 20:57 [chuckles] That's where we see a lot of businesses emerging-

Max Jaderberg 20:59 Exactly

Marina Mogilko 20:59 ... collecting data and-

Max Jaderberg 21:00 Exactly

Marina Mogilko 21:00 ... um, making sure it's accessible to-

Max Jaderberg 21:03 Yeah

Marina Mogilko 21:03 ... those models. When you hire people with no biology background who are good at AI, what exactly are you looking for?

Max Jaderberg 21:11 Like a mission like ours, solve a disease with AI, it's a really complex problem. You need to hire people from lots of different disciplines. For better or for worse, there exists no world expert at solving disease or d-- making drugs with AI because this field hasn't existed long enough. And if someone maybe claimed they're a world expert at, you know, drug design with AI, it's probably using a technology that we don't use internally. It's something outdated. So we believe that we have to hire best people of their respective disciplines, so the best AI scientists, the best AI engineers with the best software engineers, with the best chemists and the best biologists and the best clinicians, and then bring them all together and then ask them to start speaking each other's language and working together because those intersectional people haven't quite yet existed in these new spaces where AI is coming to these. And I think this will be the truth for quite a lot of industries, which means that we encourage hiring people who are just the best AI scientists and engineers. They've got no life sciences background. They don't know anything about making drugs. But they have to be willing and curious and open to learning this space. People who end up being fantastic AI scientists and engineers or fantastic chemists and biologists, they tend to be pretty smart people who love learning. Um, and you know, I think a lot of these fields are quite learnable by other people. Um, so as people come in, they learn about, you know, how to make drugs. They learn about how AI works. Um, and then internally, we create this interdisciplinary type of scientist or engineer. And actually, there's, there's another benefit that comes from that, not just hiring the best people, but you get people looking at another discipline or another scientific field or technical field from a completely new perspective because they're kind of naive to it. And so you almost, like, by accident, get this first principles type analysis and thinking for these domains, and that's exactly what you need to do. You need to do that first principles type thinking to build in these new spaces.

23:09 – 24:45

The Interview Question Max Uses to See How You Learn

Marina Mogilko 23:09 Yeah, and I think this is how hiring for top companies is happening these days, right? Because you can't find someone for that job with, like, five years of experience-

Max Jaderberg 23:17 Yeah

Marina Mogilko 23:17 ... just because it did not exist. And you mentioned you're looking for people who know how to learn and who are curious.

Max Jaderberg 23:22 Mm.

Marina Mogilko 23:22 Is there a question that you ask them to figure out they're the exact type of person?

Max Jaderberg 23:26 I ask them, like, honestly, like, "Talk to me about learning a new field," or like, "How did it feel to..." um, like, "When did you operate in a space that was, like, completely new to you? How did that feel? How did you approach that?" To try and, like, see how people think in that way. Um, asking people, like, very honestly about where their technical work failed and how did you learn, like, figure out the solution to that. And so st-start to try and get people to, like, just recount and talk about what their, their past experience has been like in these really uncomfortable learning situations where you feel like everything's going over your head. You don't understand the world around you. How did you feel in that moment? What did you do? And yeah, where did you get to?

Marina Mogilko 24:12 Max says his team brings together specialists from different fields and looks for people willing to learn from one another. And learning fast is one of the skills I keep hearing about from people building AI tools. I went back through my conversations with leaders at Replit, Anthropic, LinkedIn, and Microsoft AI and found seven skills that came up again and again, from framing the right problem and checking AI's work to learning as the tools change. I put the first list in Future Proof, my newsletter. You can read the newsletter and subscribe through the link in the description. It is free. Okay,

24:45 – 26:58

How AI Narrows the Search to 20 or 50 Molecules

Marina Mogilko 24:45 talk to me about using AI at work. So AI can generate a lot of hypotheses. How do you know it's actually worth pursuing?

Max Jaderberg 24:54 You're doing drug discovery, you're trying to make a medicine, uh, you're trying to design the next iteration of a molecule. AI can come up with choose your large number, like any thousands and thousands, millions of different hypotheses for next designs. Which ones of these are good and which ones are bad? You don't wanna test necessarily all thousands or millions in the lab, so instead we build a complementary set of, of models, which are these like simulation models, these world models, virtual lab models, that allow us to use instead of going into the lab, and allow us to like actually start to predict what the-- what, what will be the response if we actually make these molecules and put them into a lab. So we, we think about some models and generative models and design agents that come up with new hypotheses and designs and ideas, and then other models that are our, like, virtual lab models that we ping off on a computer still, uh, to refine those and whittle those down to just the very best that we have the highest confidence will be good in the real world. And then we go into the real world and actually make these molecules and test them out and see if they work. That's the really time-consuming-

Marina Mogilko 25:58 Yeah

Max Jaderberg 25:58 ... and expensive part of it.

Marina Mogilko 25:59 What's the percentage of those successful tests? Run me through those numbers.

Max Jaderberg 26:03 Yeah. And, and, you know, this is like one of those core internal metrics for the different parts of the pipeline that you wanna see get better over time. It's at the point in time now where instead of having to screen, experimentally go through hundreds of thousands of different designs, we can make 20 or 50.

Marina Mogilko 26:23 Mm-hmm.

Max Jaderberg 26:23 And there's such a high percentage of hits, you know, of, of like experimental success there that we can just sort of like get rid of the old style of, of just testing many, many, many molecules, thousands of designs. So that's the sort of thing we track internally. Again, as we make advancements in our core model benchmarks, we wanna see that translate into that real experimental success rate. Um, and the higher the experimental success rate, the less experimental time and iterative experimental time you'll need. That's quite different from trying to do like a lab in the loop style where you don't-- you just wanna interface with the real world quicker. We actually wanna just bring the real world onto a computer

26:58 – 28:29

What Researchers Do When AI Comes Up With the Ideas

Max Jaderberg 26:58 as much as possible.

Marina Mogilko 26:59 Wow, this is interesting how the researcher's job is changing, right? Do researchers still generate hypotheses, or do you think AI is so good at it now-

Max Jaderberg 27:08 Yeah

Marina Mogilko 27:08 ... that the researcher's job has changed?

Max Jaderberg 27:10 I think it's changed, and it's changed whether you're an AI researcher or a chemist. I can give examples on both. Like if you're an AI researcher, you used to, um, propose hypothesis of a change of a model, um, and then implement that and run the experiments and see if it works. Now we're seeing more and more like auto research happening on the ground, where an individual AI researcher is not proposing an individual idea, they're proposing the sort of design space of model changes-

Marina Mogilko 27:42 Mm

Max Jaderberg 27:43 ... and the goals of those, of that idea. I wanna increase the efficiency or throughput or accuracy with this sort of flavor. And the, uh, agents and models are coming up with the ideas and then running the experiments on that. As a chemist, it's a very similar thing. Before, you want to make improvements on a molecule design, a chemist might be drawing out the molecule on a whiteboard and starting to change individual at- atoms, so coming up with different design hypotheses themselves. Now, a chemist at ISO, a drug designer, is instead defining what's that design space? What are the goals of my next round of design? And that defines an objective for our design agents, and then the agents are working autonomously to do all the individual designs and come up with the, the next iteration.

28:29 – 29:30

Max’s Personal AI Stack

Marina Mogilko 28:29 That's really fascinating.

Max Jaderberg 28:30 Yeah.

Marina Mogilko 28:30 And apart from designing, what's your personal AI stack to run your productivity?

Max Jaderberg 28:35 I have, um, like a whole host of different agents that are like running across my-- like everything's plugged in. We use a lot of Gemini internally. Um, we use a lot of all the different models. One example that's just been mind-blowing for me has been actually on like the analysis of data and the visualization of data, um, where you can imagine we, we have lots of different operations, um, going on, and now with this sort of like agent stack, you can just plug in all of your company data. You can plug in all of your-- the different data across your different programs and start to actually like tease this out and visualize this in a way that normally this would take like you having a, an amazing designer next to you to do this. But instead, this can just be driven purely by me as someone running the company. We use a lot of Gemini. We use a lot of like pretty much all the models. Um, yeah, including open

29:30 – 30:46

Inside a Drug Design Agent That Runs for Days

Max Jaderberg 29:30 source.

Marina Mogilko 29:30 I read that you have some drug design agents that run for two to four days-

Max Jaderberg 29:34 Yeah

Marina Mogilko 29:34 ... consecutively. Is there something about prompting them to trust them to run for so long?

Max Jaderberg 29:40 This is the result of a lot of research. We, we have a, an amazing research team in this space of like, how do we create these design agents? Um, so this comes together in terms of like the prompt, the harness, um, how we define success a- a- as these agents run.

Marina Mogilko 29:56 How do you define success for an agent?

Max Jaderberg 29:58 This is le- less about accuracy. It's more, um, defining the goal space of the agent. So like what, what is the agent trying to achieve? In something like drug design, and I think this is common for a lot of design problems, um, there's not just like one quantity you're trying to achieve. There's probably like a whole host of different properties that you're trying to achieve. So getting the agent to understand, okay, what are all those different properties? What's the, like, the goal space for these? A lot of these properties are in conflict with each other. You might be pushing up- Property A is getting really, really good, but as a result, property B-

Marina Mogilko 30:29 Yeah

Max Jaderberg 30:29 ... gets worse. So this isn't like a greedy optimization process. This is actually like this really open-ended search and optimization process where you need to be able to explore and exploit. It really goes back to these like core reinforcement learning type-

Marina Mogilko 30:44 Mm-hmm

Max Jaderberg 30:45 ... um, ideas.

30:46 – 32:22

How to Tell if an AI Agent Is Worth Using

Marina Mogilko 30:46 And how do you decide which problem you're gonna automate with an agent? Is there a framework where you're like, "Okay, I repeated this five times. Now it's time to really sit down and build something out"?

Max Jaderberg 30:56 I think the, um... Like you just have to be really eval driven and leaderboard driven in terms of, um, the development here. So I think it's very easy to fool yourself with agents at the moment that like, "Oh, yeah, look, I can just automate this thing with agents-"

Marina Mogilko 31:12 Yeah

Max Jaderberg 31:12 ... "and isn't this great?" But like is the result good? Like is it saving you time? Is it creating more scientific value or economic value down, down the line? And so it's really important to actually have your evals, your leaderboard of like what does good look like?

Marina Mogilko 31:28 Mm-hmm.

Max Jaderberg 31:28 How do you codify that? If you're lucky to work in a verifiable domain, this might be strongly verifiable on a computer. Like you can say whether the agent is right or wrong. In the space of biology and drug discovery, this isn't really verifiable on a computer. Um, so it's a lot softer in terms of understanding if an agent is doing well or not, but trying to get down your evaluation benchmark-

Marina Mogilko 31:48 Mm-hmm

Max Jaderberg 31:49 ... for every different part of your agent system, and then really evaluate, are you making progress here? Is this getting better? Is this agent actually doing better than the human would in this scenario, therefore, you know, or is it working faster than a human would? Just actually starting to empirically measure this, and then, then you can direct as well your development time into the right part of the stack.

Marina Mogilko 32:08 Before we keep going, can I ask you a small favor? If you've been enjoying these conversations, I'd really appreciate you subscribing. It's a free way to support me and the team behind Silicon Valley Girl, and it means a lot to us. Thank you.

32:22 – 33:17

The Part of His Job Max Keeps Human

Marina Mogilko 32:22 What part of your work you haven't automated yet and you think you will never automate?

Max Jaderberg 32:26 Yeah. It's a, it's a really good question. I mean, I, I spend a huge amount of my time talking to [chuckles] the different people around the company and managing the humans in, you know, in our company and, and behind our amazing mission, and I think that's something that's, you know, I don't think is automatable at the moment and, and probably shouldn't be automated, um, at the moment.

Marina Mogilko 32:48 Yeah. I think sometimes you have a 20-minute conversation, and it changes your whole strategy and the way you think-

Max Jaderberg 32:53 Yeah

Marina Mogilko 32:53 ... about the field, uh, which is why I started this podcast. [chuckles]

Max Jaderberg 32:57 Yeah, yeah, yeah. And, and I think, I think again, like these models can come up with so many amazing hypotheses, and like you have to think, "Okay, what is the real like bottleneck resource?" Like even as these models get to AGI, it's gonna be like human attention, time, physical space. These are like the limiting resources that-

Marina Mogilko 33:15 Yeah

Max Jaderberg 33:15 ... you know, are never solvable, I

33:17 – 35:58

Why Promising Drugs Still Fail in Clinical Trials

Max Jaderberg 33:17 think.

Marina Mogilko 33:17 Another bottleneck that actually Sam Altman mentioned, we had this creator Q&A, and he was asked about using AI to cure cancer, and he said we might see something already next year curing some type of cancer, but the bottleneck will be the FDA. Do you agree with him?

Max Jaderberg 33:33 So at the moment, takes a-- and rightly so, it takes a bit of time, and it takes a lot of evidence to get through the regulatory process to make sure that, okay, look, you think you've come up with an amazing medicine. Can I actually start testing this on people in clinical trials? You have to go through the regulatory process with the FDA, and that's really important because you have to have patient safety-

Marina Mogilko 33:56 Yeah

Max Jaderberg 33:56 ... sort of first and foremost in mind. So at the moment, you know, this takes a bit of time to collect data on all the different animal models and, and all the different data points required to s- to convince the experts that, yeah, no, this, this molecule looks like it should be safe in people, and maybe it'll work for this particular disease, um, area. And only then can you start a clinical trial and actually start testing this out on people. And then when you actually do start a clinical trial and test this out on people, well, it turns out, even with all of that data beforehand that goes into approving this clinical trial, still 90% of drugs fail in the clinic because there's some unintended side effect or some in- unintended toxicity. And so when you start these clinical trials, you have to be-- start quite slowly. Initially, you do what's called Phase 1 trials before Phase 2. Phase 1 trials are all about figuring out if y- your molecule is toxic. Can patients tolerate a, a decent dose, therapeutic dose of this medicine? So you have to start really, really small and gradually work your way up. That takes a bit of time, but it's really necessary to protect patients', um, safety. But it's, it's that way because we're not very good at predicting how well these molecules are gonna work. We don't know the mechanisms of those. We can't predict very well the toxicity, um, of this space. We, we do these animal models for toxicity, but these animal models, they're only like 30% predictive of human toxicity-

Marina Mogilko 35:25 Mm-hmm

Max Jaderberg 35:25 ... because a rat isn't a human. We see, like, a lot of opportunity to, to actually improve how good we are at understanding will these medicines work. And over time, ultimately, I think everyone, us, the FDA, patients, everyone, ev- everybody wants to get amazing new drugs to patients who need them as quickly as possible and as safely as possible. And so as you start to prove out the evidence that actually we can do this a lot safer and with a lot more predictive power, that should be able to, like, spur change of, of, of how long this takes and how quickly we can get these drugs to the patients

35:58 – 40:14

When AI-Designed Drugs Could Reach Pharmacies

Max Jaderberg 35:58 who need them.

Marina Mogilko 35:58 So when do you think we're gonna see the first drugs designed by AI at a pharmacy?

Max Jaderberg 36:03 Yeah. At, at a pharmacy, I don't know. It takes, it takes a bit of time to get through the full-

Marina Mogilko 36:06 Yeah

Max Jaderberg 36:06 ... um, process, but I think there's, there's a lot of amazing things being designed at the moment. We're starting to see s- some things entering the clinic that have had, you know, that they've touched machine learning in some way. We're working really, really hard on our own, you know, pipeline of fully AI-designed drugs, um, and getting those to the clinic as, as quickly as possible.

Marina Mogilko 36:24 What's the most exciting for you, if you can share?

Max Jaderberg 36:27 Yeah. We, we have some just, I think, like, really exciting programs in the space of oncology, so different types of cancer, um, also immunology and inflammation, um, like really trying to look at some of the core drivers of these diseases and can we start to change how these core drivers work, and, and of course, they get mutated or changed- In disease states. Um, so this is really exciting to see the progress.

Marina Mogilko 36:51 Which patients do you think are gonna benefit first from the drugs that you're develop- developing?

Max Jaderberg 36:57 I don't know. It's too early to say which ones will, will, will hit the clinic first, yeah. What I can say is, like, we're pushing so hard on getting these first, you know, potential medicines to clinical trial to start seeing that impact on patients. We have some really, really promising stuff, so, you know, touch wood, like, all of our data keeps progressing as it does, and, and yeah, we're really excited about this.

Marina Mogilko 37:18 Okay, keeping fingers crossed.

Max Jaderberg 37:19 Yeah.

Marina Mogilko 37:20 Um, I recently heard Ray Kurzweil said you have to live to 2030 because after that, every year you live is gonna add more than a year to your life-

Max Jaderberg 37:29 Mm

Marina Mogilko 37:29 ... because technology advances. Do you believe in this concept in general?

Max Jaderberg 37:33 I think we're getting-- we're gonna get so much better at designing new medicines, at understanding disease so much better that I don't know exactly what the rate of that progress is gonna ... But I think there's very much something in this. We fast-forward 10 years, 20 years, 30 years, I think the way that we're gonna be understanding disease, how we think about solving disease is gonna be radically different. Even thinking about how to understand the onset of disease, how it works, and therefore preventative measures before the disease manifests itself into, like, a really critical state should be very, very different.

Marina Mogilko 38:06 Yeah.

Max Jaderberg 38:07 That's gonna require us to understand so much more about how, you know, biology works and, and measure everything about us. You start thinking about, well, a lot of the-- pretty much all of the medicines that are out there and that we use, it's like one molecule for the entire population of the world.

Marina Mogilko 38:22 Mm-hmm.

Max Jaderberg 38:23 But, like, disease is so nuanced, and it manifests itself in, like, different patients in very, very different ways. And you can imagine with AI, there's gonna be so much more opportunity to understand each individual case for patients much, much better, and then on the other hand, design a much more bespoke solution for these patients. So I think we'll see like a track towards much more personalized medicine, much more focused on individual, um, sort of cohorts of patients, even down maybe to the individual itself over time.

Marina Mogilko 38:54 Can you make any, like, timeline predictions around that?

Max Jaderberg 38:58 We should be seeing this change now, over the next couple of years as these, you know, new waves of medicines start coming into the clinics. Like, they should, um, be able to tackle diseases in, in new ways. Like, you know, we want to-- us and other people, right, in the industry are, like, trying to, um, go after these new targets. As we have this better way to engineer molecules and engineer new medicines, that's gonna open up new frontiers of things to work on. So I think even over the next 10 years, we should be seeing like a big shift in, in the types of, uh, medicines that are available for patients, and, um, and yeah, so o- over this decade, I think we're gonna really start seeing that shift.

Marina Mogilko 39:37 That's fascinating, and I wonder, like, what's gonna happen to our life expectancy and what will be the next bottleneck.

Max Jaderberg 39:44 Yeah, you can think about there's obviously, like, very advanced stages of disease that we want to rectify, um, but then there are like-- there's a lot of disease that's just, like, associated with getting older, and that's, like, afflicting our quality of life, and that's before you even talk about trying to reverse and live longer. So I think there's a lot of, like, near-term opportunities, um, to really improve our quality of life, and that's before you get into things like longevity and trying to reverse age of organs and specific

40:14 – 43:34

What to Learn if You Want to Work in AI for Science

Max Jaderberg 40:14 things.

Marina Mogilko 40:14 Give advice to someone who really wants to work on AI for science-

Max Jaderberg 40:18 Mm-hmm

Marina Mogilko 40:18 ... in the next five years.

Max Jaderberg 40:19 I think, um, don't neglect learning science itself. Even as AI gets better, is able to produce so, so many hypotheses, like, where do we focus this compute? Like, we still have finite amount of compute. Like, we need to think about, like, what are the core scientific questions for your respective field to ask? Where do we actually focus all of this amazing intelligence and explosion of intelligence with compute that's coming? Where do we actually focus this? What questions do we ask? Um, and so I think, you know, still, like, really truly under- un- understanding, like, the fundamentals of science and the mechanisms behind that is gonna be imperative, and then just be, like, really creative and aggressive with adopting, [laughs] you know, agents and all of AI, I think, around the space.

Marina Mogilko 41:02 Yeah, it comes down to asking the right questions, and it really reflects how research works, right? It's, it's knowing which question to work on.

Max Jaderberg 41:10 Yeah. The, the process of answering those questions, that's gonna change. That's always been changing as new technology. But what questions to ask, where to, like, um, focus your finite resources, whether that's human time, agent time, wet lab work, it's gonna have to be focused somewhere.

Marina Mogilko 41:27 Do you think it's-- it depends on the amount of knowledge that you have acquired? Does it depend on your taste, the quality of questions that you're asking?

Max Jaderberg 41:35 Yeah, I, I think, um, like, taste is a real thing. I actually talk a lot about this internally, like research taste, and I think that is a function. Your research taste is your ability to, um, sort of filter to what questions to ask and filter quickly to what could be some good ways to tackle answering those questions. And I think that research taste comes from, you know, experience, and that experience helps you build your intuition and your understanding of the domain. Yeah, choose your favorite scientific domain or area of, like, maths, engineering, science. Like, these are really, really complex. A lot of that complexity is almost, like, too high-dimensional to hold in its true state in your head, so you come up with, like, mental frameworks and intuitions to help you navigate that, uh, as a scientist or an engineer. I think that's the thing that needs to be built up through hands-on experience and, and yeah, just exposure to your f- to these fields. And also, I think there's a lot to be said by, um, like, keeping generalist enough. Like, I, I, I'm a generalist, like, science, scientist engineer. I love that, and I think there's a lot that you can learn, you can take from other areas of maths and science and apply elsewhere. Um, it's often those, like, crossover ideas that work really well.

Marina Mogilko 42:53 Building framework, yeah, I think it just teaches your brain, um, to understand what's going on and put it into simple words.

Max Jaderberg 43:00 Mm-hmm.

Marina Mogilko 43:00 Uh, and I think it's a really great skill for the AI world. [laughs]

Max Jaderberg 43:04 Yeah, yeah.

Marina Mogilko 43:06 Thank you so much, Max. That was so much more than just curing diseases, and I'm keeping-

Max Jaderberg 43:10 Such a pleasure

Marina Mogilko 43:10 ... fingers crossed for your mission, and hopefully in the next few years we're gonna see a great effect of AI on healthcare.

Max Jaderberg 43:18 Thank you so much. Been such a pleasure.

Marina Mogilko 43:19 If you've ever asked ChatGPT a health question, watch my conversation with a Stanford doctor, Jonathan Chan. We talk about how the way you ask can shape the answer and where AI can be useful and where it can go wrong. Click here.

Episode duration: 43:34

In the uListen app, the transcript lights up word by word as you listen. Get uListen

Transcript of episode yIFhTVMOH98

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.