YT YouTube2Text

YouTube2Text → Transcripts → DeepMind Veteran: How to Find the Next Big Opportunity in AI

DeepMind Veteran: How to Find the Next Big Opportunity in AI — Transcript

by Silicon Valley Girl · 8,420 words · 1,369 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00With AI, I believe you need that big
0:02ambition and that biggest mission
0:04possible.
0:04>> This is Max Jaderberg, an Oxford trained
0:07AI researcher whose startup was acquired
0:09by Google. He helped develop AlphaFold 3
0:12and is now president of Isomorphic Labs,
0:15founded by Demis Hassabis to design new
0:17medicines with AI.
0:18>> So AI can generate a lot of hypotheses.
0:21How do you know it's actually worth
0:24pursuing? [music]
0:24>> Instead of having to screen
0:26experimentally go through hundreds of
0:28thousands of different designs, [music]
0:30we can make 20 or 50 and then we go into
0:33the real world and actually make these
0:34molecules and test them out and see if
0:36they work.
0:37>> So when do you think we're going to see
0:38the first drugs designed by AI at a
0:40pharmacy?
0:41>> Well, it turns out
0:44>> Max, welcome to the podcast.
0:46>> Great. Thank you for having me.
0:47>> So happy to have you.
0:50Uh you're on an amazing mission uh to
0:52solve all disease with the help of AI.
0:56Can you talk to me what have you seen
0:59inside your work that made you believe
1:01that this is all possible?
1:02>> Maybe I'll rewind to to sort of say a
1:04bit where we began
1:06um and where we like started to see that
1:09actually this crazy mission of solve
1:11diseases like one we can actually go
1:12after. My background um has been in in
1:15like core core AI, core deep learning
1:17development. I started my research in
1:20computer vision and uh like
1:22uh things like ImageNet,
1:24how to train these big networks to get
1:27computers to see. Um I spent a long time
1:30at DeepMind working on all the general
1:33algorithms, new architecture design,
1:35early days of generative models,
1:37language models, and deep reinforcement
1:38learning.
1:40And through that was working on these
1:41like
1:42really hard
1:44challenge tasks like uh AlphaGo and
1:46AlphaStar. We were able to solve these
1:48like impossible challenges again and
1:50again.
1:52And maybe we hadn't got to AGI yet, but
1:54it was very clear that we we had this
1:56new technology with AI that was that was
1:58able to do like very different stuff.
2:00We're starting to think, okay, well,
2:01what should we really use this for?
2:03Like, what could be the impact on on
2:04humanity here?
2:06And at the same time, we start seeing
2:08the AlphaFold results internally in
2:10DeepMind starting to come to fruition.
2:12So this was
2:13an amazing project of trying to predict
2:15the structure of proteins, which is a
2:17very very hard and fundamental task of
2:18science. And we could start to see these
2:21results emerging of this AlphaFold
2:23network starting to get really really
2:24good.
2:26And this led to like a real conviction
2:28that you can actually
2:30train AI models that generalize in the
2:34space of biology and in the space of
2:35chemistry. And no one had really shown
2:37that before. And I think that
2:39proof point of AlphaFold was really
2:43allowed us to take the leap of faith to
2:44say, hey, we're at the right stage of
2:46the development of AI to really actually
2:48now start something in Isomorphic Labs
2:51to go full force towards using this to
2:55really solve human health and
2:57we really think about solving all
2:58diseases, that core mission for us to go
3:01hard behind.
3:02>> So from this AlphaFold moment to solving
3:04all diseases, what has to happen?
3:06>> AlphaFold isn't the solution to making
3:09medicines, to solving diseases. It's
3:11really that proof of concept that a
3:14solution could exist.
3:15You know, the to
3:17actually make a medicine to help solve a
3:19disease, it requires so many different
3:22things to come together. It's not just
3:23one problem. It's like a dozen really
3:25really tricky problems come together.
3:27That's why it's so hard to actually make
3:28medicines. It takes actually quite a
3:30long time to make a medicine. It takes
3:3110 years to to go from conception at the
3:34moment to [snorts] medicine for patients
3:36and the failure rate is really high.
3:38About 90% of medicines fail through
3:41development.
3:42>> Because we don't know how they're going
3:43to react, like how the proteins are
3:44going to react with the molecules, and
3:46this is where AI is helping as far as I
3:48understand this, right?
3:49>> there's there's so many different
3:51dimensions
3:52that come into play and which are hard.
3:54So yes, how do you actually go about
3:56solving that disease? What
3:58usually protein target do you need to go
4:00after? What's the what's the fundamental
4:02pathway
4:03and little molecular machine in the
4:05body, in the cell that you want to go in
4:07and modulate to make the disease state
4:09better? Can you actually make a molecule
4:11that's good and that works in in the way
4:13that you intended? If you take that
4:14molecule as a drug,
4:16will it actually get to the right part
4:17of the body? Will it be safe? Will it
4:19have toxic side effects? There's just
4:21such a long list of things that you need
4:23to design for and need to understand.
4:26Um and it's really really hard.
4:27And so the solution from a AI
4:30perspective and a model perspective is
4:32not is not actually just one model. It's
4:35like a whole collection of AlphaFold
4:36level breakthroughs. We think about
4:38maybe half a dozen plus AlphaFold-like
4:41breakthroughs that we have to
4:42>> to happen.
4:43>> that need to happen to then come
4:45together into one holistic drug design
4:47engine. We've been working really hard
4:49over the last five years on just
4:51systematically going through those half
4:52dozen plus AlphaFold-like breakthroughs.
4:55>> But have you identified them?
4:56>> Yeah, I mean we we know what those
4:58should be. We we we we think about
5:01okay, we want to make medicines in a new
5:03way. What are those like core
5:04superpowers, those superpower
5:06capabilities that if you could just give
5:09this superpower to a chemist, to a drug
5:11designer or a biologist, it would
5:13radically improve our ability to make
5:15medicines. So we can really identify
5:17what those are.
5:19That comes into things like
5:20understanding how cells work and
5:22different parts of disease biology work,
5:24how to design like different aspects of
5:26designing molecules, different aspects
5:28of understanding human toxicity. And
5:31then we can just go systematically after
5:33those
5:34to to create those superpowers. So we've
5:35identified those and we've been building
5:37the models towards that.
5:38>> How is AI helping you achieve what you
5:41want?
5:41>> It's really on
5:43two fronts. There's like the
5:44understanding fronts and then the
5:46ability to design new medicines fronts.
5:49So
5:50to make a medicine,
5:52you need to understand the disease
5:54biology first and foremost. And
5:57biology is a really like beautiful and
5:59complex beast.
6:01There's just so much complexity. We know
6:04such a tiny proportion of what there is
6:08to know, I think about about the science
6:09of biology and disease biology and how
6:11our cells work and
6:13crucially if you have a disease, how
6:15that changes the normal function of
6:17cells. And so we use AI and we create
6:20these models. These are like
6:21foundational models of biology
6:23that help us understand how how these
6:26systems work so it can help us
6:27understand
6:28how
6:30different proteins and the different
6:32building blocks of life, proteins, DNA,
6:34RNA, how they all fit together, how they
6:35form pathways which are the sort of
6:38signaling mechanisms for cells.
6:41If you have a disease, how that changes
6:43how these pathways operate and therefore
6:45that gives us a clue into how we can go
6:47in and maybe fix the situation. So
6:49that's one side of the equation is to
6:50understand how we work, how cell biology
6:54work, how disease biology works.
6:56And then the other side of the equation
6:57is okay, how how do you then design an
7:00intervention to that? So that's a
7:02medicine and a medicine is something
7:04like a molecule, like a small molecule
7:06you could take as a pill or something
7:08called an antibody or a peptide that you
7:10inject.
7:12Um how do you actually design a molecule
7:14that will go and modulate your
7:16biological system
7:17as you intended to solve that disease
7:19state? And so really AI is helping on
7:21those two fronts really really strongly
7:23for us.
7:23>> I'm trying to understand from the
7:25business perspective because what you're
7:26describing looks like very long-term
7:29research. How do you break it down into
7:31business goals and how do you know when
7:34to commercialize what you're doing?
7:36>> Great question. When we started the
7:38company, we we didn't want to go and sit
7:41in a dark cave and work away on the the
7:44core engine and models for
7:46a decade and then pop up with the
7:47solution. Now, hey, we're ready to
7:49commercialize this.
7:51The whole idea was to from day one
7:54be developing the core technology, the
7:57AI models, the platform and then also
7:59from day one to be able to like really
8:01try and use this for the the real
8:03hardest commercial problems. Very early
8:05in the company, we were doing real drug
8:07design,
8:08not just
8:09work that we instigated ourselves, which
8:11is really important, but also going out
8:13and just looking for the best people in
8:15the world, the best pharma companies,
8:17signed partnerships with them
8:19where
8:20they brought us literally like the
8:22industry's hardest problems. So you
8:23can't mark your own homework. You're
8:26really held against like the top
8:27standard of the industry.
8:29Um and that was tough, but it really
8:30forces forced us to develop
8:34our models and our AI drug design design
8:36engine in a way that's going to really
8:38speak to true commercial problems.
8:40Um and so that started
8:43our commercial
8:44um footing very early on in the company.
8:47And we think about sort of two aspects
8:50of of how we do drug design. We do
8:53internal programs, making medicines that
8:55we instigate ourselves, we fully fund
8:57ourselves, we own all of the IP around
8:59and we'll develop these these molecules
9:01ourselves.
9:03And then we work with partners to also
9:04do this um in in partnership. Uh so that
9:08builds
9:09two different commercial models, I
9:10think, that work really well together.
9:12>> There is one thing we still haven't
9:13solved perfectly on my remote team, and
9:16it is having one place where people and
9:18AI can actually work on the same
9:20project. Right now, research lives in
9:21one place, old scripts somewhere else,
9:24producer notes in another docs, and
9:26Claude is usually in his own chat. So
9:28every time we start a new episode,
9:29someone ends up copying old angles,
9:32questions, examples, and context back
9:34and forth. That's why we've been testing
9:35Miro, who's sponsoring this video. We'll
9:38put everything for the episode onto one
9:40Miro canvas, research, past scripts,
9:42audience feedback, questions, and
9:44decisions. Then we connect Claude to the
9:46same board through Miro's MCP server. So
9:48instead of the producer working in one
9:50place and Claude in another, they're
9:52working from the same project context.
9:54Claude can read what's already on the
9:56board, work with it, and write the
9:58output back there. The producer can
10:00review it, add her judgment, change the
10:02angle, and keep moving without
10:04constantly digging through old chats and
10:06documents. For us, it means less copying
10:08and re-explaining and more time actually
10:10working on the episode. If your team
10:12already uses AI, but the work is still
10:14scattered across chats and docs, check
10:16out Miro's MCP server. Link is in the
10:19description. Thanks to Miro for
10:20sponsoring this video.
10:21>> Do you have a like a business framework
10:23when you look at these new technologies?
10:25How do you know they're ready to be
10:28commercialized? Like when do they leave
10:30the lab and become a product?
10:32>> In general, I think this is different
10:34industry to industry. You can think
10:35about something like self-driving cars
10:38where you have to get to like 99.99999%
10:40accuracy before it's safe enough to
10:42deploy and it's actually like a
10:44commercial product.
10:45In the space of drug discovery,
10:48we're actually not very good before AI.
10:50We're actually not very good at
10:51predicting how good a molecule is going
10:54to work. The status quo of the industry
10:55is a very empirical process. It's sort
10:57of trial and error in the lab process
11:00which means that
11:01actually you don't need
11:03to have perfect predictive power
11:05before you can create real value,
11:08scientific value on your on your drug
11:10discovery programs. So we really think
11:12about what is that level of predictive
11:14accuracy that really changes the game
11:17for drug designers and helps us unlock
11:20problems that other people wouldn't be
11:21able to work on. And so when you hit
11:23that level of accuracy
11:25and that level of accuracy actually
11:26doesn't need to be very, very, very
11:28high, it can be just like a significant
11:31delta above what the rest of the
11:34industry might have access to. You can
11:35then start using that internally. We
11:37have this framework where we're doing
11:39this core
11:40frontier AI model development. We have
11:42lots of benchmarks and Evals that act as
11:45like proxies for how good these models
11:47are
11:48and then we just aggressively translate
11:50these onto live drug discovery programs
11:52to see
11:53that they and make sure that as we're
11:55making progress on our internal
11:56benchmarks, our static benchmarks, that
11:58translates into real scientific
12:00breakthroughs and therefore, you know,
12:01economic value on our live discovery
12:03programs.
12:04Um and oftentimes you can see that
12:08you're making progress on your machine
12:10learning benchmark
12:11and it 's not translating into reality.
12:13and that's because there's a disconnect
12:15between this proxy of your machine
12:17learning benchmark or your eval and the
12:19real problems. You have to go back and
12:21redesign your leaderboard and your evals
12:23to better reflect the real on the ground
12:25situation.
12:26>> It's interesting. How do you manage your
12:27ambition? How do you think about that?
12:29>> In this space and it comes back to
12:32making sure that we actually have the
12:34scope of ambition
12:35that's large enough,
12:37a business model that is expansive
12:39enough to mean that we can rationalize
12:42huge investment into our
12:44ability to create new models and to
12:46create this drug design engine. And and
12:48this is really back to the like the the
12:50beginnings of the company is really
12:52trying to assess like what is that big
12:55market potential? What's that
12:57potential impact for patients that could
12:59be?
13:00And how much of that can be realized
13:02differently now with AI? And that helps
13:05underwrite this amazing amount of
13:07investment into the core modeling and
13:08the core science which we then can have
13:10faith will translate at the end of the
13:12day.
13:12>> How do you stay
13:15true to your mission when the goal is so
13:17ambitious? Have you ever had periods
13:19when you know this is not working out,
13:21the technology is not there yet?
13:22>> Yeah, our mission is crazy. You know,
13:24solve all diseases. It's it's pretty
13:25audacious. Um we say that with a
13:27straight face. Like we really like feel
13:30and mean that mission. We know it's not
13:32going to happen overnight. So honest
13:34about that. We also like frankly it's
13:36not going to happen just within the four
13:38walls of ISO, but I believe you need
13:40that big ambition and that big like the
13:44the biggest mission possible
13:45because if you start reducing what you
13:48what you think your scope of impact can
13:50have, you start building out the company
13:51in different ways and you start
13:52designing different goals that don't
13:53quite quite match that that ambition.
13:56But then saying that this mission is too
13:57big for any one person to just hold in
14:00mind. How do you translate that into the
14:01day-to-day? So in my head I think about
14:03like a mission tree. You have the big
14:05solve all disease mission. Okay, what's
14:07what's one level down from that? Okay,
14:09well we need to be like solving the way
14:11we make medicines like and then on and
14:13on and on and so we think about missions
14:15throughout the company,
14:17even down to every individual.
14:19I want every individual to have their
14:20own personal missions that when you put
14:23these all together are speaking to our
14:25top level solve all disease mission now.
14:27And um and then thinking about, okay,
14:29what are the milestones, like the
14:30practical milestones
14:32that
14:33we can start to measure to help us
14:37litmus test, okay, are we still on this
14:38path? Are we pushing hard enough? Do we
14:40need to be doing something different?
14:41>> This is really fascinating. I was just
14:43talking to Greg Brockman, co-founder of
14:45OpenAI, and my key takeaway from that
14:47conversation was he told me we have to
14:49be working on our ambition because AI
14:51makes so many things possible. And
14:54something that you just said, if you
14:55have lower ambition, then the company
14:57works in a completely different way and
14:58you start solving for smaller problems.
15:01So you have to have that North Star. So
15:03if you have any tips for like the
15:05darkest days for founders when they're
15:07on a big mission but everyone around
15:09them is doubting and maybe even research
15:11is not proving that they're right, what
15:13can they tell themselves and what can
15:14they do?
15:15>> First of all, I think that's why it's
15:16important to actually have a mission,
15:18like know what that is,
15:19and to have spent the time to
15:21rationalize that and be like, yeah,
15:22okay, I can fully get behind this
15:25and just drive hard on this. For me,
15:28personally, the the ISO mission is one
15:30that I can just dedicate my life to and
15:33have no regrets behind um doesn't matter
15:35how hard
15:37it is. I think you have to translate
15:39this into
15:40like some signals of life. Does that
15:42make sense? So that you don't go too
15:45crazy ahead trying to pursue something
15:47really big that's going to take a long
15:48time. What are those little
15:51goal posts even if they're artificial?
15:52>> Yeah.
15:53>> I'll give you an example. Like we had
15:54this um particular modeling task
15:57and it was one that we like
16:00has been there since the beginning of
16:01the company. And we sent we set we want
16:03to solve all disease, sure, but how does
16:05this team on this particular modeling
16:06task like feel that day to day? Well, we
16:08set this like crazy ambitious
16:10quantitative target. You need to hit 80%
16:12accuracy on this particular benchmark.
16:15Oh, and by the way, the world is only
16:1610% good at it. So it feels like so out
16:18of reach,
16:19>> Yeah.
16:19>> but it's still like a very tangible
16:22number. And day-to-day you can see
16:24climbing 1%, 2%, 3%. So you're moving
16:27towards that. It still seems very far,
16:29but you can see progress against that.
16:30And actually as you compound this year
16:32over year,
16:33you can actually start to reach these
16:35things. So I think always trying to
16:36translate that we we try and translate
16:38these like really big qualitative ideas
16:41into
16:42just something quantitative. Even if
16:44it's slightly artificial, it can help
16:46focus teams, focus the mind in just like
16:49a very gamified competitive way. We
16:51literally call these leaderboards. And
16:52it doesn't matter what part of the
16:53company you're you're in, you you've got
16:55your quantitative goal that you're
16:57trying to grind towards.
16:58>> I love that. That's that's very very
17:00practical advice for founders. Yeah.
17:02Do you think uh there are industries
17:06that are ready for an Alpha Fold moment
17:09like biology when you're looking around?
17:11>> Yeah, so outside of biology
17:13um
17:14I think there's
17:16like a lot of opportunity to sort of
17:18find places that AI can now be applied
17:22and that still require an Alpha Fold
17:24like moment. So what what what what does
17:26this look like? We we have all of this
17:27amazing advancement on our core LLMs,
17:30our reasoning agents.
17:32Um these help us like understand data
17:36and reason about the world in, you know,
17:38the way we do, but you know, starting to
17:40far surpass us in terms of intelligence.
17:43Uh if you think about
17:44the biology space, that could be just
17:46like a human biologist will look at the
17:47data,
17:48look at some experimental data and try
17:50and draw conclusions for that and and
17:51hypothesize about what what experiment
17:53can be done next. Now these reasoning
17:55models are able to do that and starting
17:57to maybe even outperform people in doing
17:59that.
18:00But can they actually predict what the
18:02experimental outcome will be itself? Can
18:04they simulate the real world? Maybe not.
18:07Maybe not at this stage. Maybe one day
18:08they'll be able to train the models to
18:09do that as well.
18:11So I think there's going to be a lot of
18:13um applied AI domains where the
18:15reasoning models will come up with
18:16hypotheses or help you understand and
18:18analyze the world, but you still
18:20are bottlenecked by into the real world
18:22and poking the real world and trying to
18:24see what the response of that is. Doing
18:26a physical [clears throat] experiment.
18:27That's the case in biology. It's the
18:28case in, you know, material science and
18:30almost like any, you know, it's the case
18:32in in any form of like physical design
18:35work that you're doing. Uh, a lot of
18:37spaces of engineering. Um, and in those
18:40sort of cases, in in those sort of
18:42domains, there's I think there's an
18:43opportunity for AlphaFold-like moments,
18:45i.e., the ability to create these AI
18:48models that simulate part of that
18:51physical world so that instead of going
18:53and doing the physical experiment, you
18:55can poke your model instead and see what
18:57the response is. That's very
18:58complementary to the LLMs and reasoning
19:00models.
19:00>> And how do you know that the model is
19:02not ready to solve that yet? Have you
19:04ever had a situation where you have this
19:06idea,
19:07>> Mm.
19:07>> uh, but then you look at the model and
19:09you realize the technology is still a
19:11decade away from solving it.
19:13>> Yeah.
19:13I think a lot of that comes down to what
19:15data is available.
19:17Um, so machine learning AI only works
19:20obviously with
19:21by training on data. I think you have to
19:24start by looking, okay, what is this
19:26predictive capability that's needed
19:28and what data is there that might have
19:30some information content inside of it
19:33that could speak to helping predict that
19:35thing. And if you don't have much data
19:37or the data is very very noisy or it's
19:39very low dimensional, there's not
19:40actually much signal
19:41>> Mm-hmm.
19:42>> in this data,
19:43then
19:44there might not be actual scope with
19:47given the current data to make a lot of
19:49impacts on this modeling task.
19:51And that could also be an opportunity
19:53for people, by the way, because then you
19:55should actually start thinking, okay,
19:56well, what data could exist
19:59to unlock this problem? And I think
20:01there'll be many problems in this sort
20:02of space where if you look at the
20:04current like data that the world has
20:06access to,
20:08it's quite small, it doesn't have the
20:09right signal, it's too noisy,
20:11but maybe there's a new piece of data
20:13generating technology or way we can
20:14collect data or generate new data that
20:17could then suddenly unlock new domains
20:19that haven't been unlocked before.
20:21>> So you think the models are powerful
20:22enough so they can do way more than
20:23we're able to achieve just because we
20:25miss the data layer?
20:26>> I'm a deep learning maximalist. I think
20:28we have the algorithms
20:29that can solve
20:31almost any problem. We have the compute
20:34now. We have just an abundance of
20:35compute. Of course, everyone's compute
20:37constrained, but there's so much compute
20:39available.
20:40So it really comes down to
20:42if we have [snorts] the algorithms, we
20:43have access to compute, we've got
20:44amazing people that can put this
20:45together, then we need to to we need to
20:47make sure that we have the data. And a
20:49lot of these amazing applied AI
20:51potential domains, they're limited by
20:53the data that's available to to train
20:55on.
20:55>> That's a business idea.
20:57That's where we see a lot of businesses
20:58emerging, collecting data and
21:01making sure it's accessible to those
21:03models.
21:04When you hire people with no biology
21:06background who are good at AI, what
21:09exactly are you looking for?
21:11>> Like a mission like ours to solve
21:12disease
21:13with AI,
21:15it's a really complex problem. You need
21:17to hire people from lots of different
21:19disciplines. For better, for worse,
21:20there exists no world expert at solving
21:23disease or making drugs with AI because
21:25this field hasn't existed long enough.
21:28And if someone maybe claimed a world
21:30expert at, you know, drug design with
21:31AI, it's probably using a technology
21:33that we don't use internally. It's
21:35something outdated.
21:37So we
21:39believe that we have to hire
21:41best people of their respective
21:43disciplines.
21:45So the best AI scientists, the best AI
21:47engineers with the best software
21:49engineers, with the best chemists and
21:51the best biologists and the best
21:52clinicians,
21:54and then bring them all together
21:55and then ask them to start speaking each
21:57other's language and working together
21:58because those intersectional people
22:01haven't quite yet existed in these new
22:02spaces where AI is coming to these.
22:04And I think this will be the truth for
22:06quite a lot of industries, which means
22:07that we encourage hiring people who are
22:09just the best AI scientists and
22:10engineers. They've got no life sciences
22:11background. They don't know anything
22:12about making drugs, but they have to be
22:14willing and curious and open to learning
22:17this space. People who end up being
22:19fantastic AI scientists and engineers or
22:21fantastic chemists and biologists, they
22:22tend to be pretty smart people who love
22:24learning
22:25Um, and, you know, I think a lot of
22:28these fields are quite learnable by
22:30other people.
22:31Um, so as people come in, they learn
22:33about, you know, how to [clears throat]
22:34make drugs. They learn about how AI
22:36works.
22:37Um, and then internally we create this
22:39interdisciplinary type of scientist or
22:42engineer. And actually, there's there's
22:43another benefit that comes from that.
22:45Not just hiring the best people, but
22:48you get people looking at another
22:50discipline or another scientific field
22:51or technical field from a completely new
22:54perspective because they're kind of
22:55naive to it. And so you almost
22:58like by accident get this first
23:00principles type analysis and thinking
23:03for these domains. And that's exactly
23:04what you need to do. They need to do
23:05that first principles type thinking to
23:07build in these new spaces.
23:09>> Yeah, and I think this is how hiring for
23:11top companies is happening these days,
23:13right? Because you can't find someone
23:15for that job with like five years of
23:16experience just because it did not
23:17exist. And you mentioned you're looking
23:19for people who know how to learn and who
23:21are curious. Is there a question that
23:23you ask them to figure out their the
23:25exact type of person?
23:26>> I awesome, like honestly, like
23:29talk to me about learning a new field or
23:32like how did it feel to
23:34um, like when did you operate in a space
23:37that was like completely new to you? How
23:38did that feel? How did you approach
23:40that? To try and like see how people
23:42think in that way.
23:44Um,
23:45asking people like very honestly about
23:47where
23:48their technical work failed and how did
23:51you learn like figure out the solution
23:53to that and so start start to try and
23:56get people to like just recount and talk
23:58about what their their past experience
24:00has been like in these really
24:01uncomfortable learning situations where
24:03you feel like everything's going over
24:05your head. You don't understand the
24:07world around you. How did you feel in
24:08that moment? What did you do? And yeah,
24:11where did you get to?
24:12>> Max says his team brings together
24:14specialists from different fields and
24:16looks for people willing to learn from
24:18one another. And learning fast is one of
24:21the skills I keep hearing about from
24:23people building AI tools. I went back
24:25through my conversations with leaders at
24:27Replika, Anthropic, LinkedIn, and
24:29Microsoft AI, and found seven skills
24:31that came up again and again, from
24:33framing the right problem and checking
24:34AI's work to learning as the tools
24:37change. I put the first list in
24:39Futureproof, my newsletter. You can read
24:41the newsletter and subscribe through the
24:42link in the description. It is free.
24:44>> Okay, talk to me about using AI at work.
24:48So AI can generate a lot of hypotheses.
24:51How do you know it's actually worth
24:53pursuing?
24:54>> You're doing drug discovery. You're
24:55trying to make a medicine. Uh you're
24:57trying to design the next iteration of a
24:59molecule. AI can come up with choose
25:01your large number, like any
25:03thousands and thousands, millions of
25:05different hypotheses for next designs.
25:08Which ones of these are good and which
25:09ones are bad? You don't want to test
25:12necessarily all thousands or millions in
25:14the lab. So instead,
25:16we build a complementary set of of
25:18models, which are these like
25:19simulation models, these world models,
25:22virtual lab models
25:23that allow us to
25:25use instead of going into the lab
25:28and allow us to like actually start to
25:29predict what what what will be the
25:30response if we actually make these
25:32molecules and put them into the lab. So
25:33we
25:34we think about some models and
25:35generative models and design agents that
25:37come up with new hypotheses and designs
25:39and ideas, and then other models that
25:41are like virtual lab models that we ping
25:43off on a computer still uh to refine
25:46those and whittle those down to just the
25:48very best that we have the highest
25:49confidence will be good in the real
25:51world. And then we go into the real
25:53world and actually make these molecules
25:55and test them out and see if they work.
25:56That's the really time-consuming and
25:58expensive part of it.
25:59>> What's the percentage of those
26:00successful tests? Run me through those
26:02numbers.
26:03>> Yeah, and and you know, this is like one
26:04of those core internal metrics for the
26:06different parts of the pipeline that you
26:08want to see get better over time. It's
26:09at the point in time now where
26:12instead of
26:14having to screen
26:16experimentally go through hundreds of
26:18thousands of different designs, we can
26:20make 20 or 50 and there's such a high
26:24percentage of hits, you know, of of like
26:26experimental success there
26:28that we can just sort of like get rid of
26:30the old style of of just testing many
26:32many many molecules, thousands of
26:34designs. So that's the sort of thing we
26:36track internally. Again, as we make
26:38advancements on our core model
26:39benchmarks, we want to see that
26:40translate into that real experimental
26:42success rate.
26:44Um and the higher the experimental
26:45success rate, the less experimental time
26:47iterative experimental time you'll need.
26:49That's quite different from trying to do
26:50like a lab in the loop style where you
26:52don't you just want to interface with
26:54the real world quicker. We actually want
26:56to just bring the real world onto a
26:57computer as much as possible.
26:59>> Wow, this is interesting how the
27:00researcher's job is changing, right? Do
27:02researchers still generate hypotheses or
27:05do you think
27:06AI is so good at it now that the
27:09researcher's job has changed?
27:10>> I think it's changed and it's changed
27:12whether you're an AI researcher or a
27:14chemist. I can give examples on both.
27:16Like if you're an AI researcher,
27:18you used to um
27:21propose
27:22hypothesis of a change of a model
27:25um and then implement that and run the
27:27experiment and see if it works. Now
27:29we're seeing more and more like auto
27:31research happening on the grounds where
27:33an individual AI researcher is not
27:35proposing an individual idea,
27:38they're proposing the sort of design
27:40space of model changes and the goals of
27:44those of that idea. I want to increase
27:45the efficiency or throughput or accuracy
27:48with this sort of flavor. And the uh
27:51agents and models are coming up with the
27:53ideas and then running the experiments
27:55on that. As a chemist, it's a very
27:56similar thing.
27:57Before you want to make improvements on
27:59a molecule design, a chemist might be
28:01drawing out the molecule on a whiteboard
28:04and starting to change individual atoms,
28:06so coming up with different design
28:07hypotheses themselves.
28:09Now a chemist at ISO, a drug designer,
28:12is instead
28:13defining what's that design space? What
28:15are the goals of my next round of
28:17design? And that defines an objective
28:20for
28:21our design agents and then the agents
28:23are working autonomously to do all the
28:26individual designs and come up with the
28:28the next iteration.
28:29>> That's really fascinating. And apart
28:31from designing, what's your personal AI
28:33stack to run your productivity?
28:35>> I have um like a whole
28:39host of different agents that are like
28:41running across my like everything's
28:43plugged in. We use a lot of Gemini
28:45internally. Um we use a lot of all the
28:48different models. One example that's
28:50just been mind-blowing for me has been
28:52actually on like the
28:53analysis of data and the visualization
28:55of data
28:57um where
28:59you can imagine we we have lots of
29:01different operations um going on. And
29:05now with this sort of like agent stack,
29:07you can just plug in all of your company
29:08data. You can plug in all of your the
29:10different data across your different
29:12programs
29:13and start to actually like tease this
29:14out and visualize
29:16this in a way that
29:17normally this would take like you having
29:19an amazing designer next to you to do
29:22this. But instead this can just be
29:23driven purely by me as someone running
29:25the company. We use a lot of Gemini. We
29:26use a lot of like pretty much all the
29:27models. Um yeah including open source.
29:30>> I read that you have some drug design
29:32agents that run for two to four days
29:34consecutively. Is there something about
29:37prompting them to trust them to run for
29:40so long?
29:40>> This is the result of a lot of research.
29:43We we have an amazing research team in
29:45this space of like how do we create
29:46these design agents.
29:47Um so this comes together in terms of
29:50like the prompt, the harness, um how we
29:53define success
29:55as these agents run.
29:56>> How do you define success for an agent?
29:58>> This is less about accuracy. It's more
30:01um defining the goal space of the
30:02agents. So like what what is the agent
30:04trying to achieve? In something like
30:05drug design and I think this is common
30:07for a lot of design problems. Um there's
30:09not just like one quantity you're trying
30:11to achieve. There's probably like a
30:13whole host of different properties
30:15that you're trying to achieve. So
30:16getting the agent to understand okay
30:18what are all those different properties?
30:19What's the like the goal space for
30:21these? A lot of these properties are in
30:22conflict with each other. You might be
30:24pushing up
30:25property A that's getting really really
30:27good but as a result property B gets
30:29worse. So this isn't like a greedy
30:32optimization process. This is actually
30:34like this
30:35really open-ended search and
30:37optimization process where you need to
30:39be able to explore and exploit. It
30:41really goes back to these like core
30:43reinforcement learning type
30:45um ideas.
30:46>> And how do you decide which problem
30:48you're going to automate with an agent?
30:50Is there a framework where you're like,
30:51okay, I repeated this five times, now
30:53it's time to release it down and build
30:55something out?
30:56>> I think the um
30:58like you just have to be really eval
30:59driven and leaderboard driven in terms
31:02of um
31:03the development here. So I think it's
31:05very easy to fool yourself
31:08with agents at the moment that like, oh
31:10yeah, look, I can just automate this
31:12thing with agents and isn't this great?
31:14But like is the result good? Like is it
31:16saving you time? Is it creating more
31:18scientific value or economic value down
31:20down the line?
31:21And so it's really important to actually
31:23have
31:24your evals, your leaderboard of like
31:27what does good look like? How do you
31:28codify that? If you're lucky to work in
31:30a verifiable domain, this might be
31:32strongly verifiable on a computer. Like
31:34you can say whether the agent is right
31:35or wrong. In the space of biology and
31:37drug discovery, this isn't really
31:39verifiable on a computer.
31:41Um so it's a lot softer in terms of
31:43understanding if an agent is doing well
31:44or not. But trying to get down your
31:46evaluation benchmark
31:48>> Mhm.
31:49>> for every different part of your agent
31:50system and then really evaluate, are you
31:52making progress here? Is this getting
31:53better? Is this agent actually doing
31:55better than the human would in this
31:57scenario? Therefore, you know, or is it
31:59working faster than a human would?
32:02Just actually starting to empirically
32:03measure this and then then you can
32:05direct as well your development time
32:07into the right part of the stack.
32:08>> Before we keep going, can I ask you a
32:10small favor? If you've been enjoying
32:11these conversations, I'd really
32:13appreciate you subscribing. It's a free
32:15way to support me and the team behind
32:17Silicon Valley Girl, and it means a lot
32:20to us. Thank you.
32:22>> What part of your work you haven't
32:23automated yet and you think you will
32:25never automate?
32:26>> Yeah, so it's a really good question. I
32:28mean, I I spend a huge amount of my time
32:30talking to
32:31the different people around the company
32:33and managing the humans in, you know, in
32:37our company and and behind our mission.
32:39And I think that's something that's
32:41you know, I don't think is automatable
32:43at the moment and and probably shouldn't
32:45be solvable um
32:47at the moment.
32:48>> Yeah, I think sometimes you have a
32:4920-minute conversation and it changes
32:51your whole strategy and the way you
32:53think about the field,
32:55uh which is why I started this podcast.
32:57>> Yeah, yeah, yeah, and and I think I
32:59think again like these models can come
33:00up with so many amazing hypotheses and
33:02like you have to think, okay, what is
33:04the real
33:05like bottleneck resource? Like even as
33:07these models get to AGI, it's going to
33:09be like
33:10human attention,
33:12time, physical space. These are like the
33:14limiting resources that you know, never
33:16solvable, I think.
33:17>> Another bottleneck that actually Sam
33:19Altman mentioned, we had this creator
33:21Q&A and he was asked about using AI to
33:24cure cancer and he said we might see
33:26something already next year curing some
33:29type of cancer,
33:30but the bottleneck will be the FDA. Do
33:32you agree with him?
33:33>> So at the moment, takes a
33:36and rightly so, it takes a bit of time
33:38and it takes a lot of evidence
33:40to get through
33:41the regulatory process to make sure that
33:43okay, look, you think you've come up
33:45with an amazing medicine, can actually
33:47start testing this on people in clinical
33:49trials. You have to go through the
33:50regulatory process with the FDA
33:52and that's really important because you
33:55have to have patient safety sort of
33:56first and foremost in mind. So at the
33:59moment, you know, this takes a bit of
34:00time to collect data
34:02on all the different animal models and
34:05and all the different data points
34:07required to convince the experts that
34:09yeah, no, this this molecule looks like
34:12it should be safe in people
34:14and maybe it'll work for this particular
34:16disease um area. And only then can you
34:19start a clinical trial and actually
34:20start testing this out on people.
34:23And then when you actually do start a
34:24clinical trial and test this out on
34:25people,
34:26well, it turns out even with all of that
34:29data beforehand that goes into approving
34:31this clinical trial,
34:33still 90% of drugs fail in the clinic
34:35because there's some unintended side
34:38effect or some unintended toxicity. And
34:40so when you start these clinical trials,
34:42you have to be
34:43start quite slowly. Initially, you do
34:44what's called phase one trials before
34:46phase two. Phase one trials are all
34:47about figuring out
34:49if your molecule is toxic.
34:53Can patients tolerate
34:55a decent dose, therapeutic dose of this
34:58medicine? So you have to start really,
35:00really small and gradually work your way
35:02up. That takes a bit of time, but it's
35:03really necessary to protect patients um
35:07safety.
35:08But it's it's that way because we're not
35:09very good at predicting how well these
35:11molecules are going to work. We don't
35:12know the mechanisms of those. We can't
35:14predict very well the toxicity
35:16um of this space.
35:18We we do these animal models for
35:19toxicity, but these animal models,
35:21they're only like 30% predictive of
35:23human toxicity because a rat isn't a
35:26human. We see like a lot of opportunity
35:28to actually improve
35:30how good we are understanding what these
35:32medicines work. And over time,
35:34ultimately, I think everyone, us, the
35:37FDA, patients, everyone, everybody wants
35:40to get
35:41amazing new drugs to patients who need
35:43them as quickly as possible and as
35:44safely as possible. And so as you start
35:46to prove out the evidence that actually
35:48we can do this a lot safer and with a
35:50lot more predictive power,
35:52that should be able to like spur change
35:54of of of how long this takes and how
35:56quickly we can get these drugs to the
35:57patients who need them.
35:58>> So when do you think we're going to see
35:59the first drugs designed by AI at a
36:02pharmacy?
36:03>> Yeah, at at a pharmacy, I don't know. It
36:04takes it takes a bit of time to get
36:06through the full um process, but I think
36:08there's there's
36:09a lot of amazing things being designed
36:11at the moment. We're starting to see
36:12some things entering the clinic that
36:14have had that they've touched machine
36:16learning in some way. We're working
36:18really, really hard on our own, you
36:19know, pipeline of fully AI designed
36:21drugs
36:22um and getting those to the clinic as as
36:24quickly as possible.
36:25>> the most exciting for you if you can
36:26share?
36:27>> Yeah, we we have some just I think like
36:29really exciting programs in the space of
36:32oncology, so different types of cancer,
36:35um also immunology and inflammation, um
36:38like really trying to look at some of
36:39the core drivers of these diseases. And
36:42can we start to change how these core
36:44drivers work and and of course they get
36:45mutated or changed in disease states.
36:49Um so this is really exciting to see the
36:50progress.
36:51>> Which patients do you think are going to
36:52benefit first from the drugs that you're
36:55developing developing?
36:57>> I don't know. It's too early to say
36:58which ones will will will hit the clinic
37:00first. Yeah. What I can say is like
37:02we're pushing so hard on getting these
37:04first you know potential medicines to
37:07clinical trial to start seeing that
37:08impact on patients. We have some really
37:10really promising stuff. So you know
37:12touch wood like
37:14all of our data keeps progressing as it
37:16does and and yeah we're really excited
37:17about this.
37:18>> Okay keeping fingers crossed. Um I
37:20recently heard Ray Kurzweil said you
37:22have to live to 2030 because after that
37:25every year you live is going to add more
37:28than a year to your life because
37:29technology advances. Do you believe in
37:31this concept in general?
37:33>> I think we're getting we're going to get
37:35so much better at designing new
37:37medicines at understanding disease so
37:39much better that I don't know exactly
37:41what the rate of that progress is going
37:43to but I think there is very much
37:44something in this. We fast forward
37:4710 years, 20 years, 30 years. I think
37:49the way that we're going to be
37:51understanding disease, how we think
37:52about solving disease is going to be
37:54radically different. Even thinking about
37:56how to understand the onset of disease,
37:59how it works and therefore preventative
38:01measures before the disease manifests
38:03itself into like a really critical state
38:05should be very very different. That's
38:07going to require us to understand so
38:08much more about how you know biology
38:10works and and measure everything about
38:12us. You start thinking about well a lot
38:14of the pretty much all of the medicines
38:16that are out there and that we use
38:19it's like one molecule for the entire
38:21population of the world but like
38:24disease is so nuanced and it manifests
38:26itself in like different patients in
38:28very very different ways and you can
38:30imagine with AI there's going to be so
38:32much more opportunity to understand each
38:35individual case for patients much much
38:37better
38:38and then on the other hand design
38:41a much more bespoke solution for these
38:43patients. So I think we'll see like a
38:45track towards more personalized
38:46medicine,
38:47much more focused on the individual,
38:50um, so cohorts of patients, even down
38:52maybe to the individual itself over
38:54time.
38:54>> Can you make any like timeline
38:56predictions around that?
38:58>> We should be seeing this change now over
39:01the next couple of years as these, you
39:02know, new waves of medicines
39:04start coming into the clinics, like they
39:06should
39:07um, be able to tackle diseases in in new
39:10ways, like, you know, we want to
39:12ask another people right in the industry
39:14are like trying to um, go after these
39:16new targets.
39:18As we have this better way to engineer
39:19molecules and engineer new medicines,
39:21that's going to open up new frontiers of
39:22things to work on. So I think even over
39:25the next
39:2610 years we should be seeing like a big
39:28shift in in the types of
39:30uh, medicines that are available for
39:32patients and
39:33um, and yeah, so over this decade I
39:36think we're going to really start seeing
39:37that shift.
39:37>> That's fascinating and I wonder like
39:39what's going to happen to our life
39:41expectancy and what will be the next
39:43bottleneck?
39:44>> Yeah, you can think about there's
39:45obviously like
39:46very advanced stages of disease that we
39:48want to
39:49rectify, um, but then there are like
39:52there's a lot of disease that's just
39:53like associated with getting older and
39:56that's like afflicting
39:58our quality of life. That's before you
39:59even talk about trying to reverse and
40:01live longer. So I think there's a lot of
40:03like near-term opportunities um, to
40:06really improve our quality of life.
40:08And that's before you get into things
40:10like longevity and
40:12trying to reverse age of organs and
40:13specific things.
40:14>> Give advice to someone who really wants
40:16to work on AI for science in the next
40:18[clears throat] five years.
40:19>> I think um,
40:20don't neglect learning science itself.
40:22Even as AI gets better, as able to
40:24produce so so many hypotheses like where
40:26do we focus this compute? Like we still
40:28have finite amount of compute. Like we
40:30need to think about
40:32[sighs]
40:32like what are the core scientific
40:34questions for your respective field to
40:36ask? Where do we actually focus all of
40:38this amazing intelligence and explosion
40:40of intelligence with compute that's
40:42coming? Where do we actually focus this?
40:43What questions do we ask?
40:45um, and so I think you know, still like
40:49really truly understanding like the
40:52fundamentals of science and the
40:53mechanisms behind that is going to be
40:55imperative and then just be like really
40:57creative and aggressive with adopting,
40:59you know, agents and all of AI, I think
41:01around the space.
41:02>> Yeah, it comes down to asking the right
41:04questions and it really reflects how
41:06research works, right? It's it's knowing
41:09which question to work on.
41:10>> Yeah, the the process of answering those
41:12questions
41:13that's going to change. That's always
41:14been changing as new technology. But
41:17what questions to ask, where to
41:19like focus your finite resources,
41:22whether that's human time, agent time,
41:24wet lab work, it's going to have to be
41:26focused somewhere.
41:27>> Do you think it's it depends on the
41:30amount of knowledge that you have
41:31acquired? Does it depend on your taste,
41:33the quality of questions that you're
41:34asking?
41:35>> Yeah, I I think
41:37like
41:38taste is a real thing. I actually talk a
41:39lot about this internally like research
41:41taste and I think that is a function.
41:44Your research taste is your ability to
41:47sort of filter to what questions to ask
41:50and filter quickly to
41:53what could be some good ways to tackle
41:55answering those questions.
41:57And I think that research taste comes
41:59from
42:00you know, experience
42:02and that experience helps you build your
42:04intuition and your understanding of the
42:06domain. Yeah, choose your favorite
42:08scientific domain or area of like maths,
42:10engineering, science. Like these are
42:11really really complex.
42:13A lot of that complexity is almost like
42:14too high dimensional to hold
42:17in its true state in your head. So you
42:19come up with like mental frameworks and
42:21intuitions to help you navigate that
42:24as a scientist or an engineer. I think
42:26that's the thing that needs to be built
42:27up through hands-on experience and and
42:30yeah, just exposure to your to these
42:32fields. And also I think there's a lot
42:34to be said by
42:37like keeping generalists enough. Like I
42:39I I I'm a generalist like science
42:41scientist engineer. I love that and I
42:43think there's a lot that you can learn
42:45you can take from other areas of maths
42:47and science and apply
42:49elsewhere.
42:50And it's often those like crossover
42:51ideas that work really well.
42:53>> Building framework, I think it's just
42:55teaches your brain
42:57uh to understand what's going on and put
42:59it into simple words. Uh and I think
43:01it's a really great skill for the AI
43:04world.
43:04>> Yeah, yeah.
43:06>> Thank you so much, Max. That was so much
43:08more than just curing diseases. And I'm
43:09keeping fingers crossed for your mission
43:12and hopefully in the next few years
43:14we're going to see
43:15a great effect of AI on healthcare.
43:18>> Thank you so much. Been such a pleasure.
43:19>> If you've ever asked ChatGPT a health
43:22question, watch my conversation with a
43:24Stanford doctor Jonathan Chan. We talk
43:27about how the way you ask can shape the
43:29answer and where AI can be useful and
43:32where it can go wrong. Click here.

About this transcript

This page contains the full transcript of DeepMind Veteran: How to Find the Next Big Opportunity in AI by Silicon Valley Girl, generated from the public captions YouTube serves with the video. The transcript has 8,420 words across 1,369 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.