0:25Welcome everybody. So I'm Alistair Berisford the current head of the department of computer science and technology also known as the computer laboratory. It's my great pleasure this afternoon to welcome Demis back to Cambridge. So Demis studied computer science here in Cambridge in the 1990s at a time when the lab was based just next to this lecture hall and where Robin Walker who I'm pleased to say is here today was Demis' director of studies at Queens College. Uh I was discussing earlier with Demis and we think that this is where he had his
0:54first Cambridge lecture maths at 9:00 am on the first Thursday of Mikmouth's term. So this seems a fitting place for him to return to.
1:03So Demis had already made several incredible achievements uh by the time he arrived in Cambridge. He was the chess master and second highest rated under 14 player in the world. And after completing his schooling a year early, instead of backpacking around Europe, he instead took a job in the computer games industry where he co-designed and was the lead programmer for the computer game theme park.
1:27After graduating from Cambridge with a first- class degree, Demis returned to the games industry, at first working at Lionhead Studios and subsequently forming his own company. However, there was clearly a passion in him for fundamental scientific research. And so, Demis returned to academia, this time to UCL, where he studied for a PhD in cognitive neuroscience, graduating in 2009.
1:49He stayed on at UCL until 2011 when he left to co-found DeepMind, an AI research lab which was acquired by Google in 2014.
1:57Demis and colleagues at Google Deepmind have gone on to make several seinal contributions to science. Highlights include Alph Go, which was the first computer program to beat professional human players at the board came and alphafold a computer program which is able to predict protein structure. is stemis's contributions to AlphaFold for which he was awarded a share of the 2024 Nobel Prize in chemistry.
2:21Now alongside his incredible intellectual contributions over this period, he's also been a fantastic supporter of the university uh including funding for academic positions and significant support for students from underrepresented groups both in the computer lab and at Queens College. And Demis' uh passion and support for the next generation of computer scientists is the motivation for our lecture today.
2:43and I'm sure he will not only help us understand how to accelerate science uh scientific discovery with AI but also inspire the next generation of students in the room to change the world too. And with that, I would like to welcome Dennis to the stage.
3:00[applause] [applause] Thanks Alistister for that lovely introduction and um it's so great to be back at Cambridge. I always have a warm feeling when I um sort of my homecoming back to Cambridge and specifically this lecture hall as Alistister reminded me I think it is the first lecture hall I was in. It's always been my favorite lecture hall. Uh I remember telling and I see a lot of my old friends here from my Cambridge days
3:27um Aaron I think about uh that one day maybe I'd come back to give a lecture in here and talk about announcing AGI and maybe a robot would walk on and uh astound everyone. I'm not going to do that today um to to disappoint you, but maybe in a few years time I'll come back again and I'll give that lecture. Um but you know this is amazing place. It's such an inspiring place and I'm going to talk a little bit about how Cambridge has inspired my whole career actually and hopefully is going to do the same
3:55for many of you the students in in the room.
3:59So for me it started uh uh my journey on AI started with games and specifically chess as Alistister mentioned. So I was playing chess from the age of four years old and uh very seriously for the England junior teams and things like that and it got me thinking about thinking itself you know how does our mind come up with these plans with these ideas um how do we problem solve and how can we improve obviously when you're playing chess at a young age and you're trying to play competitively you're trying to improve that process uh and it
4:28was fascinating to me perhaps more fascinating than even the games I was playing was the actual mental processes behind it uh and in AI and and computers. I came across computers and AI for the first time in the context of chess. Uh and trying to use very early chess computers like the one on the right here. I think this was my first ever chess computer. Um there were physical boards where you had to actually had to press the squares down to make move the pieces. Uh and of course we were supposed to be using these chess computers to train opening
4:58theory and and learn more about chess.
5:01But I remember being fascinated by the fact that someone had programmed this lump of inanimate plastic to actually play chess really well against you. And uh I was sort of really fascinated by how that was done and how um how someone could program something like that. And I ended up experimenting myself uh in my early teenage years with an Omega 500 uh computer, amazing home computer back in the in the late 80s and early 90s and building those kinds of AI programs myself to play games like Oll. And
5:30really that was the first my first taste of AI and and I was hooked from then on and that's you know I decided from very early on that I would spend my my entire career trying to push the frontiers of of this technology. So then that led me to uh to Cambridge um and which was really my three years here were incredibly formative for me.
5:52Um and I went to a school in in north comprehensive school in North London. No one had ever got to sort of no one had gone to Oxbridge for in sort of living memory but and the reason I wanted to come to Cambridge was all these inspiring stories that I'd heard about um what happened at Cambridge. all these amazing people that I used to read about their biographies and the work they done um especially people like Crick and Watson in the in the top left there and I remember particularly a film uh the race for the double helix was amazing
6:22film from the 80s if you haven't seen it with Jeff Goldblum one of his early parts you know with all the enthusiasm that he plays all of his parts as he was Watson and they were having just such an amazing time discovering roaming around Cambridge you know working on things like DNA and I thought like that's what I want a piece of that and I want to feel what it's like to be at the frontier of discovery and what could be more exhilarating and and that film actually really brought it to life uh
6:50what that might be like um and then of course all of my my my heroes my scientific heroes a lot of them had gone through Cambridge people like Alan Turing and Charles Babage of course in in the lecture hall that we we now sit in um and uh and and even places like the Eagle Pub where if you start at Queens College. One of the tours they give you on the first day is to go and show you the the the put sort of uh uh uh table that they were discussing and
7:17and and the DNA structure uh around and uh you can't help but be inspired by that and walking down King's Parade. Um and I almost felt like the the the intellectual giants of the past were were almost speaking to you from the stones. Um and and that's how I felt, you know, going for a late night burger at Gardinas. um that that was what was inspiring me around like all of these amazing people that had walked those same steps uh over hundreds of years and and that's the history that is sort of
7:46unrivaled here at Cambridge that I think we can still draw on and take inspiration from today. Uh and then there's a picture of me and Aaron there, one of my best friends from Queens, you know, obviously on on the mathematical bridge there.
8:00And then finally, you know what when we Alistister mentioned obviously the Nobel Prize and it was an honor of a lifetime to go and collect that in in Stockholm um in December. Uh amazing one week of of activities but my favorite activity was when you get to sign the Nobel book in the Nobel Foundation and uh and and that's the book there and one of my pictures of and I started sort of leafing through the book. You sign your name and then you leaf back. you wonder if you know is cricket in there and of course he is and then you go back
8:28further and then Einstein signatures there and it's just mind-blowing really and I started to spend an hour just photoing every page of the uh of the book um so you know it's full circle for me of of of that picture and and then really seeing that film in the in the in the in the late 80s.
8:48So then uh you know in 2010 we started DeepMind in London as uh really at the time it was a kind of an Apollo program effort is the way we thought of it uh for trying to build artificial general intelligence you know AI that was truly general and could perform all the cognitive capabilities that that humans are capable of. So it would be a truly uh general AI system. In fact you know the idea for that really comes from Turing and Turing machines. So something
9:16that's able to uh compute anything that is computable uh as Turing showed with his Turing machines. And really that's been the foundation for me uh of uh and one of the main things that I carried with me from the lectures here at Cambridge was all these theoretical underpinnings of computer science and computation theory that that people like Turing and Shannon famously did in the 40s and the 50s.
9:41So we started in 2010 and it's amazing like it's 15 years ago which in some ways isn't that long ago but when we started out deep mind almost nobody was working on AI which is hard to believe today given that almost everyone seems to be working on AI today. Uh in just a matter of just over a decade uh things have gone you know accelerated incredibly and obviously we've been part of that very exciting journey.
10:04So our mission at deep mind from the beginning was um uh uh we talk about building AI responsibly to benefit humanity but the way we used to articulate it when we started out was uh in a two-step process. Step one solve intelligence step two use it to solve everything else and um and it seemed very outlandish at the time in 2010 and you could imagine trying to pitch venture capitalists on that on the basis of that mission. It seemed pretty crazy but I I still fundamentally believe uh in that today and I think more and more
10:33people are realizing that AI um built in this general way could have these kind of profound and transformative impact on almost any field uh uh which is obviously the second part of that mission statement and I think uh uh for me that that involves accelerating scientific discovery itself and medicine and advancing our understanding of the universe around us.
10:58So back when we started out there were basically two ways and in fact when I was studying here in the 90s there are two ways to build AI broadly speaking there's the expert system way which is you kind of pre-program an expert system directly with the solution. Um things like deep blue that beat Carrie Kasparov at chess very famously in the 90s actually while while I was studying here. Um that would be the pinnacle example probably of an expert system.
11:23But the problem with these expert systems was and why they never really scaled to full general intelligence is they can't deal with the unexpected. Um if you you know if something unexpected happens that you didn't already cater for um there's nothing in the system that will allow it to to deal with that.
11:40Um and they sort of were inspired by logic systems. Um and they were quite rigid and and fragile and brittle because of that. Whereas the modern day approaches are built on learning systems. So these are systems that are able to learn for themselves and learn directly from experience or data from first principles um and really inspired by more neuroscience ideas. Um and the obviously the the promise of these systems that we have today is that they can go beyond the knowledge potentially that we as the programmers or the system
12:08designers already know how to solve. And of course that's extremely valuable in areas like um scientific discovery.
12:17So we started in the in the early 2010s with um with games of course and I've used games many times in my life first of all to train my own mind then I used to build games and AI4 computer games and then finally in a third way to train up our AI systems and games are the perfect proving ground for AI systems.
12:36Um you can start with very simple games like Atari games from the 70s. uh and really this this this system DQN was the first time anyone had built an end-to-end learning system that could learn directly from raw data. So in this case the raw pixels on the screen and it's it's not told anything about the games or anything about what it's controlling. It's just told to maximize the score based on this uh uh video stream input pixel stream input.
13:03So we were able to master all different Atari games uh uh uh in sort of around 2013. Then we took these uh uh systems and we scaled them up to really the I would say the grand challenge of games AI which is um can you you you create systems that can play the game of go um at world champion level or beyond. And go of course is probably the most complex game that humans have ever invented. It's thousands of years old.
13:31So it's also the oldest game and one of the most elegant games. Um but one of the ways you can just see the complexity of go is that there are 10 to the^ 170 possible positions in go. Um so that's more way more than there are atoms in the observable universe. And um the important reason point about that is that you cannot come up with a strategy in go using brute force techniques. It would be impossible. It would be totally intractable. So you have to do something much smarter. Um, and we famously in
13:59basically 2016 won a million-dollar challenge match against 10 times world champion Lisa Dole and one of the legends of the game, the South Korean grandmaster, and was watched by 200 million people around the world. Um, and not only did Alph Go, our system win that match, importantly, it actually came up with new original Go strategies.
14:21Even though um we've played go for thousands of years and professionally for hundreds of years, um it was still able to find never seen before strategies, most famously uh this move 37 move uh here in red uh during game two, which um if you watch the documentary on this, you'll see which is on YouTube, you'll see uh that how surprised the the best players in the world were that were commentating on the game uh about this move. it was sort of an unthinkable move and yet it decided
14:50um the G this game too in favor of Alph Go um a 100 moves later. So again that told me about the potential for these types of systems to invent and discover new knowledge. So here of course we're just talking about game knowledge but um obviously my dream was to generalize this to all areas of um scientific discovery.
15:14So how do these systems work? um we we basically train up these neural networks through a system of selfplay. So this is actually Alph Go and also um subsequent systems uh that Alph Go Zero and Alpha Zero that generalized what we've done for Go to play any two-player uh game from scratch. And you start off with a version one of the system that doesn't really know anything about the game, just the rules. And it plays randomly and you play say a 100,000 games against
15:42itself uh of this system. And that creates a new database of um game positions from those 100,000 uh games.
15:51And from that, you train a second version, a slightly better version of the model, version two, uh that's trained to predict um what the likely moves are to be played in any one position and also who is more likely to win, which side, black or white, is likely to win from that position. Uh and what percentage chance do they have of winning? Uh and then you can use that version two to um you play against version one uh and in a 100 game matchoff and then if it wins by
16:20significant uh margin so in this case 55% win rate you replace the version one with version two and you create a new uh database of games that are slightly higher quality and then you learn a version three system. Uh, and if you do this and you repeat it around 17 18 times, you go from playing randomly in the morning to 24 hours or less later, you're you're by the version 17 or 18, you're stronger than um the world champion level. So, it's quite an
16:48incredible thing uh process to see this this self improvement process playing out uh in a very very short amount of time.
16:58So if we think about what these neural networks are doing is you're kind of uh reducing down this intractable search space of you know 10 to the^ 170 possibilities down to something that's much more tractable in a few minutes of uh compute time and uh it's doing this by narrowing down uh using the neural network to efficiently guide the search mechanism. So if you think about this tree of possibilities as uh and each of the nodes in this tree is a go position
17:27um then instead of having to to to look at sort of every possibility you can actually use the neural network to guide you just down the most um the most uh uh uh most interesting and most useful lines to examine. So in this case the ones in blue and then um after you've run out of thinking time you pick the best line the most promising line that you've seen thus far. So I I in this case uh this this particular line in purple.
17:55So this then leads to you know we we then play did not just go but any two-player perfect information game and it was even able to discover new strategies and new styles of playing chess um which which is kind of extraordinary but given that chess computers were so strong already. So, uh, programs like Stockfish and Alphaz was able to beat Stockfish at the time, uh, uh, in at chess, which is almost sort of impossible to do. And not only did it beat Stockfish, uh, so Alpha Zero
18:25here is in white, uh, playing against Stockfish, who's black, but in this particular position, one of the most famous games that Alpha Zero played, it's called the immortal Zugwang game.
18:34Um, white is winning here um, because it favors mobility over material. So, most chess computers favor material, and you'll see that black, those of you who play chess, you'll see that black has more material, um, but actually can't move any of its pieces. They're all stuck in the corner. And this is Alpha Zero sacrificed material for this mobility. And actually, for human grand masters uh, and top chess players, this is not only a very effective style, it's very beautiful aesthetic style uh, to
19:02and to play chess in. So um it's kind of amazing that Alpha Zero was able to sort of discover this new way, this new dynamic way of playing. And in fact, some of the top chess players in the world commentated about this. So Gary Kasparov, my all-time sort of favorite chess player, he said the programs usually reflect priorities and prejudices of the programmer. But because Alphazero learns for itself, I would say that its style reflects the truth. And then the current world champion at the time, Magnus Carlson, said uh, you know, read uh, and looked
19:32at these games and and read about the books that were written about Alpha Zero and said, "I've been influenced by one of my heroes recently, one of which is Alpha Zero." So, he actually incorporated a lot of these ideas uh, into his own game to dominate the chess scene for, you know, almost a decade now.
19:50So we did all these landmark breakthroughs in games AI and over the first I guess sort of decade of of of deep mind's existence but of course these were just the training ground for what we wanted to do and uh was just a a means to an end. It wasn't the end in itself to play these games much as I love games um it was to create these algorithms that could be generally useful for tackling real world problems.
20:15So um what we look for in real world problems not only scientific problems uh but actually uh industrial problems as well and we look for sort of three different criteria that make it suitable the problem suitable to be tackled by these types of AI systems and and ideas and algorithms that we developed for playing games. Um number one is we look for problems that can be described as massive combinatorial search spaces. So um so usually far too complex, far too
20:45many combinations to brute force the solution. So um but maybe there's some kind of structure that we can learn about with annual networks that can guide that search very efficiently.
20:56Secondly, uh we look for problems that can be described with a clear objective function or some sort of metric uh that you can optimize against. So in games that's very easy. It's things like maximizing the score or winning the game. But actually there's a lot of real world problems that you can boil down to a few metrics or a few objective functions that you're trying to maximize. And then finally, of course, you need quite a lot of data uh or experience to learn from and or ideally it's and a kind of accurate and
21:26efficient simulator so you can generate more synthetic data uh to augment uh the real data that you have. And it turns out uh that there are a lot of problems uh that can be couched in these terms if you're looking at the problem from this angle uh and including many many problems uh important problems in science. And the one that I always had in mind actually from my days of first seeing coming across the problem here at Cambridge as an undergrad is the protein folding problem uh which I'll just
21:55quickly describe to you for those that don't know about biology and proteins. So proteins are incredibly important. They're they're the building blocks of life. Pretty much every every uh uh uh function in the living body depends on proteins. Um from your neurons firing to your muscle fibers twitching. So really proteins are what makes life possible.
22:17And the protein folding problem then is is really easy to describe. It's basically a protein is is defined by its uh desri is is is defined by its gene sequence, its genetic sequence uh which then specifies an amino acid sequence uh which uh in nature then folds up spontaneously into usually a very beautiful uh protein structure. So you go from this genetic sequence to a protein structure and the reason the protein structure the 3D structure is
22:45very important is it goes a long way to defining what function it has what it does uh in the body. So um it's not doesn't totally describe the function but it it has a big part to play in what what it actually does uh in nature. So the protein folding problem then is is this problem of can you predict the protein structure directly from this one-dimensional amino acid sequence. Can you predict that computationally that incredible 3D structure uh uh from that
23:16So why is this such a hard problem? Well um uh uh Levental who's a famous researcher protein researcher in the in the 60s described this uh uh uh conjecture that that that became called it became known as the leal's paradox which is that he calculated there roughly you know 10^ the 300 possible shapes that an average protein can take um and yet somehow in nature and in the body these proteins fold up spontaneously in a matter of
23:44milliseconds. So that's the paradox like if there's so many possibilities how does nature do this right and basically how does physics uh achieve this um and that gives you hope that this must be tractable computationally in some reasonable amount of time because physics does solve this problem um you know billions of times a second in the body and um furthermore what attracted me to this um problem was that there was a bianual competition called the CASP competition which is like you can think
24:14of it as like the Olympics for protein folding and it happens every two years and it's run by some amazing people led by professor John Malt uh University of Maryland and um it's been running for since 1994 and it's a great competition because they work with experimentalists and uh who who painstakingly find these structures using very exotic and expensive equipment like electron microscopes and um and they newly discovered structures that haven't been
24:43published yet. they put it into the competition. So you actually the the the the competition organizers know what the ground truth is but the computational teams you know hundreds of teams enter every year every every competition every couple of years and they try with their computational methods to predict that those structures and it's usually around a 100 proteins that are in the competition and then at the end of the summer um they reveal what the true structures are and you can compare the predicted ones and their distances the
25:11error in the predictions uh to to the real structures.
25:17So we entered uh AlphaFold one actually for the first time in 2018 and we started this Alpha Fold project in 2016 actually pretty much the day after we c we got back from the AlphaGo match in Soul in Korea. We felt that and I felt that we were ready we had the techniques that were mature enough and ready to now be applied outside of games and to try and tackle uh really meaningful uh uh problems. We call them kind of root node problems because they open up if they
25:46could be solved they they open up whole new branches and avenues of discovery that can be built on top and protein folding was was a prime example of that.
25:54So we started working in 2016. We we AlphaFold one was ready after a couple of years and we entered it into the CAS 13 competition and you can see for the decade prior uh these bar charts are showing the winning score of the winning team uh uh in the hardest category actually the hardest proteins that were being predicted. And you can sort of think think of it as a percentage accuracy of how many of the atomic how many of these amino acids have you got in the right position within a certain tolerance uh within the sort of width of
26:24an atom you need to predict it within.
26:26And um you can see there was sort of not much progress for a decade and we were stuck at this um 60 points level and you we for the which is effectively if you got to 90 you would be within the width of an atom. So you'd be atomic accuracy and that's what we were told by experimentalists was the accuracy you'd had to reach so that it was competitive with experimental methods. so that experimentalists could actually rely on
26:56these predictions rather than having to necessarily do the laborious painstaking uh work to um to find that structure.
27:04And uh just as a rule of thumb, my biologist friends would always tell me that you know it take a PhD student their entire PhD so four five years to fold uh to find the structure of just one protein. um and um and and there 200 million proteins uh known to science and 20,000 uh uh u proteins in the human uh proteome.
27:27So uh so we with AlphaFold one we were able to uh win this competition and and sort of um be better by almost 50% than the next best system and Alphafold one for the first time introduced machine learning techniques as the main component of the system uh for the first time. Uh but it was not enough to to reach this atomic accuracy. We actually had to go back to the drawing board with what we'd learned and rearchitect it uh for Alpha Fold 2 from scratch. uh using all the learnings we had from alpha fold
27:56one to finally reach this atomic accuracy and that led the organizers to declare that the problem had been solved in uh the end of 2020.
28:06So this is an example of how alpha fold works visually. So you can see on the left hand side here is a very complex protein. Uh the ground truth is in green, the predicted uh structure is in blue and you can see how closely the blue overlaps the green. Uh and then on the right hand side you can see how alpha fold 2 works. It sort of builds up that structure in an iterative process.
28:27It kind of recycles itself actually over 192 steps. Um and then builds out starts as a as a scrunch ball of of protein matter amino acids and then it builds out more and more plausible structure to and at the end it sort of refineses the last parts of it until it has the finished prediction.
28:46Um we we immediately uh then tried to because alpha fold is so accurate but it's not just accurate it's also extremely fast. It's able to fold proteins in a matter of seconds an average protein. We realized quite quickly that we could actually fold all 200 million proteins uh known to science. And we over the course of a year we used a lot of computers on the Google cloud to to fold all of them and then put them out freely uh on a database with our colleagues at EMBLE
29:14EBI just up the road uh at the Sanger Center um uh just outside of Cambridge.
29:20And we provided that for free unrestricted access to anyone in the world to use it. Um, and that 200 million proteins, if you think about how long it takes to do that experimentally, four or five years, it's kind of like a billion years of PhD time done in one year. So, um, you know, it's it's kind of amazing to think about how much science could be accelerated. And it opened up whole new avenues of exploration because um many of these uh structures especially for the less
29:48wellstudied organisms like some certain types of plants they're very important for science and for agriculture um and agriculture research but that there there's there's that you wouldn't have any almost none of those structures will have been found and be available. So now those are all available and also 200 million you can look at them at an aggregate level and sort of look at structural structures um across species uh and uh kind of meta structures and see what the commonalities are uh
30:18through evolution. So there's really interesting new avenues of of of branches of of structural biology now that are being explored. And of course we thought about safety from the beginning. Uh and we take our responsibility very seriously at the forefront of AI. And in this case we you know consulted with over 30 biocurity and bioeththics experts to make sure that um what we were putting out into the world um the benefits of that far outweighed any any uh risks associated with it. And now I'm very proud to say
30:48that you know over two million researchers are using it um from pretty much every country in the world. um and uh and it's been sort of cited over 30,000 times now and it's become a standard tool uh in in in biology research now and many of you in the audience who are PhD students hopefully you're using it and making use of it and it's just um part of the standard cannon now that is used for biology research.
31:12Uh it's been amazing to see what other researchers have done uh with all of this uh technology and all these structures. Um, I've just called out uh six of my favorite uh uh examples. Uh people at University of Portsmouth are using it, a research group using it to tackle plastic pollution in the environment, trying to design new enzymes which are types of proteins uh that can digest plastic. Um we're working with the Fleming Center on antibiotic resistance. Um neglected
31:42diseases uh like tropical diseases that affect the poorer parts of the world. We work with drugs for neglected diseases institute. Um and here's a good example of where we can accelerate research in those areas where you know whether it's malaria, leechmanasis, um Zika virus, a lot of those structures are not known.
32:00Uh but now they can go straight to drug discovery because uh they have a lot of the information about the structures for those viruses and and bacteria.
32:10uh and then there's been a lot of fundamental research being done on things like uh finding the the structure of the nano pore complex which is a very important protein that that that lets in nutrients in and out of the nuclear pore of the cell. Um there's amazing work at the Broad Institute done with drug delivery sort of designing molecular syringes redesigning uh proteins that can deliver drugs targeted to a particular part of the body. uh and it's even being used in in things like looking at mechanisms of of fertility.
32:40So the the amount of things that it's being used for is sort of almost every area of biology and medical research now is is uh making use of alpha fold. We've continued in the last few years to develop uh more struct more uh uh uh uh developments and improving the systems.
32:58We released Alpha Fold 3 uh earlier this year uh for academics to use and we've extended now alpha fold 3 to to deal with um interactions. So um you can think of alphafold 2 as a picture of the the static uh uh protein structure but really biology is a dynamic process. So you need to understand how how um different biological elements interact with each other. So of course proteins with other proteins but also proteins with others other molecules important to
33:27life things like DNA and RNA and also ligans so uh small molecules which uh you know things like drug compounds how does the protein bind with um with that compound and then we have a separate set of work alpha proteio which is doing the reverse of alphafold but making use of still of of the alphafold techniques where if you want to design a novel protein maybe that doesn't exist in nature for a particular job, a particular function,
33:56what is the amino acid sequence and the genetic sequence that will give you that structure. So um so it's kind of like running it in reverse and trying to design new structures that will do novel things um and again could be extremely useful uh uh in in for for designing drugs and things like anti antibiotics and antibodies.
34:19So taking a step back then what is uh you know if I look at all the work we've done in the last 15 years um what's the implications uh for science and and also machine learning um and if you think about what we've done with first of all our games work uh and then now with the scientific work that we've been working on of which alpha fold is our best example um it's all about making this search tractable you have this incredibly complex problem there's many many possible solutions to the problem and you've got to find uh the optimal
34:48solution kind of needle in the haystack of that uh enormous combinatorial search space and you can't do it by brute force. So you have to learn this neural network model. Um so it sort of learns about the topology of the problem uh so that you can efficiently guide the search to reach your uh to maximize or find the optimal solution to the objective that you have in mind. And I think this is an incredibly general way uh uh in general solution an incredibly
35:17general way to approach um a whole myriad of problems. And so we think about back to the go example. So we're trying to use these systems to find the best go move but you could also change those nodes to be uh chemical compounds and now you're trying to find the best molecule in chemistry space in chemical space and the best molecule you know and this this is the beginning of drug design and and that will bind specifically to the target you're interested in but nothing else. So it
35:47reduces the side effects and the toxicity of that compound. Um and it's a very very similar um techniques that we're using in order to design these molecules now as the next steps as we as we move more and more into drug discovery.
36:01So I think in biology at least um I feel like we're entering a new era now of what I like to call digital biology. So you know I think of biology in its most fundamental level as an information processing system you know that's trying to resist entropy around it. And I think that's basically what life is. Um, of course, it's a phenomenally complex and emergent information processing system.
36:24Uh, and I think that's where AI comes in. Just like maths, uh, and the maths that I learned in this room was the perfect description language for physics, um, and phys physical phenomena. I think that AI is potentially the perfect description language for biology. Um, it's perfect for dealing with the complexities of the emergent behaviors and interactions that you get in a dynamic system like biology.
36:48And I think AlphaFold is a proof point of that. Um and I hope when we look back in 10 years time, it won't be an isolated breakthrough, but it actually um has sort of heralded in this new era um golden era of digital biology. And we're trying to progress that ourselves.
37:04We started a new spinout company, Isomorphic Labs to build on our Alphafold technology, and move more into the chemistry space that I was just talking about, and actually try and reimagine drug discovery from first principles with AI. And right now, it takes an average of 10 years um for a drug to be developed. Um and it's extraordinarily expensive. It costs billions and billions of dollars. And so I'm thinking why can't we use these techniques to reduce that down from years to months maybe even one day weeks
37:33just like we reduce down the discovery of protein structures from potentially years down to now minutes and seconds. And so we think of this as a kind of doing science at digital speed. So, um trying to bring the best of what we do in the technology area to the um the natural sciences.
37:55And my dream one day is to be able to create maybe a kind of virtual cell uh a computational cell perhaps of something very simple like a yeast cell. Um and that you can actually run experiments in silicon on it and the predictions that you get out of the virtual cell will actually tell you uh and form your your real world experiments in the lab. Um and you can reduce down a lot of the search that's done in the wet lab uh and actually um more use the wet lab for validation steps rather than the very
38:24expensive um and slow search process.
38:29Of course AI and we've been using AI not just in biology but it can be used for science, mathematics, medicine more generally. Um, and we've had a whole range of breakthroughs, not just in the biological sciences, but from things in health like identifying eye disease from retinal scans, discovering new materials, um, helping with plasma container infusion reactors, faster algorithms, um, so AI discovering better algorithms for itself like faster matrix multiplication, doing weather
38:58prediction, and even helping with quantum computers and error correction in quantum computing. And that's just a small example of some of the work we've been doing in the last um two three years. And I think AI will be almost be applicable to pretty much every field.
39:12And uh and I always encourage actually uh universities to start thinking very seriously about multi-disiplinary work where you apply AI to the right questions in a particular specialist field. And I think there's many many uh advances to be made over the next 5 to 10 years by doing that.
39:31So I'll just end then with a little bit of uh a more general view about not just AI for science but the path to AGI and how close we are to that and our more general work on the original mission of of AGI. And um we've been making a lot of advances in in all areas of general understanding of the world. We sometimes call them world models. Um so we're particularly proud of our new video model called V2 which was just released at the end of last year. It's it's um
39:59state-of-the-art video generation and um it's able to generate these videos just from a text description, right? Or or a single static image. So, you know, and uh actually although some of these videos may not seem that impressive, if you think about um this chopping the tomato one, this is like the cheuring test for video models because usually you get, you know, the tomato comes magically back together or you're chopping through the fingers or the knife moves off somewhere. And it's actually, if you think about what the
40:27systems had to do to really understand the physics of the world, uh, or this or the bubbles around this blueberry here, right? It's just generating that from text, you know, blueberries dropping into a glass of water. And it's it's doing all the physics of this correctly or the motion of um, you know, this these little cartoon characters or the bee here. Um, it's kind of mind-blowing really. And I think even if you told me 5 years ago um that this would be possible without sort of building in some special understanding of physics or
40:56something I would have told you that that's seems unlikely that that would be possible but yet somehow these learning systems are able to learn about real world physics um just from watching you know many many YouTube videos and um it's it's pretty crazy that that's possible. So we've done that. We've gone a step further with Genie2, which of course bringing my games uh hat back in here. And this is taking those VO models a step further. So now with a text instruction, you can generate a whole
41:25game. So you know, here we at the bottom here, we said generate a playable world as a robot in a futuristic city. And it just comes up with this and you can control it with the, you know, QWE keys and the arrow keys. Um, at the moment it's only consistent for a few seconds. Um, but we we're we're working to extend that to so that the the consistency of the game world lasts for many minutes.
41:46Um, and so then you've really got what I would call a world model, a really un an understanding of the real world. Um, and and how interactions in that real world work and the physics of the real world work.
41:58Um, of course we've been working very hard on the safety aspects of this. uh and from very from the very beginning in 2010, we were working on um uh planning for success even though almost nobody was working on AI back then. Um we we imagined that it would be a 20-y year mission and actually amazingly we're sort of on track 15 years in and um we we were sort of planning for success if we were to build these kinds of transformative systems and technologies.
42:26um it would come with a lot of uh uh responsibility as well to make sure they get deployed in a safe uh and responsible way. And one of the uh systems we technologies we built is called synth ID which um invisibly watermarks uh actually using an AI system adversarial AI system uh slightly adjust the pixels or the text or the or the audio um imperceptibly to the human ear or eye. um but it can be detected by a detection system that these were
42:55synthetically generated images whether that's audio um image or video. And of course it's going to become increasingly important um as uh as as these technologies become widely deployed that we're able to easily distinguish between synthetically generated images and real images.
43:13So AI has been, you know, it's this incredible potential uh to to help with our greatest challenges from climate to to health, but obviously this is going to affect everyone. Um so I think it's really important that we engage not the just it's not just the technologists deciding this, but that we engage with a wide range of stakeholders from society.
43:33So I've been really pleased in the last couple of years that one of the consequences of AI becoming mainstream is that many governments have got interested in it and many parts all parts of society and I think it's been great to see these international summits. Actually the UK hosted the first one in Bletchley Park a couple of years ago bringing together heads of government with academia and civil society to discuss uh these technologies um how to put gu the right guard rails on it how to make sure we embrace the opportunities um but we mitigate the
44:02risks um that are coming down the line and uh I think that's going to become increasingly important um given the exponential improvement uh that we're seeing with these technologies. So my my my shorthand for this is to say, you know, with with a lot of Silicon Valley tries to the kind of mantra in Silicon Valley is is like move fast and break things. And of course that's that's created a lot of advances, a lot of of the technologies we all use every day today. But I think it's not appropriate in my opinion for this type of transformative technology. I think
44:32instead you know we should be trying to use uh the scientific method and approach it with the kind of humility and respect that this kind of technology deserves and you know not we we don't know a lot of things there are a lot of unknowns around how this technology is going to develop it's so new um um and I think with exceptional sort of care and foresight we can get all the benefits uh and minimize the downsides of this but I think only if we um start the research
45:00and the debate about that out.
45:04So, just to end then, we're now building our own big multimodal models that try and take the best of all of these different models I've shown you and put it into one system. Uh, we call it the Gemini series. Our latest one's Gemini 2.0 that some of you may have tried, which is state-of-the-art across many leading benchmarks. Um, we're using it to further the I'm very excited about the next generation of assistance. I call it universal assistance. We call it project Astra where actually you have it on your phone or some other devices maybe glasses and it starts it's being
45:33an assistant that you can take around with you in the in the real world and it helps you in everyday life to um uh to to get to enrich your life or to make you more productive.
45:44And the next step then in AI is combining what I've shown you with Alph Go, these kinds of um agent-based models that able to um efficiently search through and find a good solution uh to a problem in a in a limited domain. in this case in games, but we want to actually build those types of search systems and planning systems on top of much more general models like Gemini, these world models that understand uh how the real world works and then can
46:12then plan and and achieve things in the real world. And of course, that's key to things like robotics uh working, which I think in the next two three years is going to be a huge area um that's going to going to have massive advances in it.
46:28So I'll just finish then by just having a a slight conjecture about what does this all mean if we think back to Turing and all of the work that he did to lay down the foundations uh of computer science and um and I think that if you see the work that we've done I see myself as a kind of Turing's champion in a way like he how far can the cheuring machines and this idea of classical computing go and I feel like um probably one of the lectures I took in this room was one of my
46:57favorite things to think about is the P equals MP problem which is uh a famous problem in computer science of you know what sorts of problems are tractable on classical systems and um and there's obviously a lot of great work going on in quantum computing systems uh many of that here in Cambridge and also at Google we have one of the top comput uh quantum computing groups in the world and there's a lot of things that are thought to require quantum computing to solve a lot of real world uh systems that we would like to understand a model
47:26and my conjecture is that actually classical Turing machines basically classical machines that these types of AI systems are built on can do a lot more than we previously gave them credit for. And if you think about alphafold and protein folding, you know, proteins are quantum systems. You know, they operate at the atomic scale and one might think you need quantum simulations to actually be able to find the structures of proteins. And yet we were able to approximate those solutions uh
47:54uh with our neural networks. Um and so I think you know one one potential idea here is that any pattern that can be generated or found in nature so i.e. it has some real structure physical structure can be efficiently discovered and modeled by one of these classical learning algorithms like alpha fold. Um and if that turns out to be true, I think that has all sorts of implications uh for quantum mechanics and and actually fundamental physics. Um which you know is something that I hope to
48:23explore and many of my colleagues hope to explore maybe with the help of these classical systems as well to help us uncover what the true nature of reality might be. And that leads me back to the whole reason why I started my path on AI many many years ago is that I always believe that AGI uh built in this way could be the ultimate general purpose tool to understand the universe around us and our place in it. Thank you.
48:57[applause] Great. We have a time for some questions if people have questions. A first hand shot up just here.
49:08Hi, thank you. Uh, thank you for the great talk. Um, because you have a background in neuroscience. Um, and you really like to think in terms of root node problems. Was there ever a root node problem you came across in neuroscience that you thought was worth tackling and still worth tackling to understand biological and artificial intelligence? better.
49:31Yeah, there's many. I mean, I that's what I studied for my PhD actually was um memory but also imagination. So, future thinking kind of planning. So, I really wanted to understand um how the brain did that and it turns out the hippocampus is involved in both uh so that we could maybe mimic that with with some of these algorithms. So, I think there are many key things in that. Of course um there are all the big questions around you know creativity, dreaming, consciousness, all these big
50:00questions that I think um building AI uh and then comparing it to the human mind is one of the best ways I think we'll make progress with those sorts of root no problems like you know what is the nature of consciousness and is there something special about the instantiation of the substrate of the brain versus um uh algorithmically uh mimicking that in silicon. Great. Um, got a question just here.
50:27Uh, hi. Um, I have two questions actually. So, since Deep Mind was founded before the deep learning revolution, I wanted to know what your state of mind was had deep learning not picked up or how were you going to progress? Uh, that's the first question.
50:45And second question is since you've um had intimate experiences with um such challenging problems such high dimensional problems and we know that um gradient descent and its variants can't uh sort of converse to the optimum solution only locally optimum solutions.
51:04Were you surprised that anything works at all in these systems at every s at any every point in time and um do you think that most of the nature is kind of suboptimal and so we can potentially build a more optimal nature?
51:20So look I I think um they're both great questions. So the f the first one uh look that's why we called it deep mind partly because the deep refers to deep learning right? So we we deep learning had well the early parts of the it wasn't called deep learning then but it was it was sort of just becoming um common there was these Boltzman machines and things that Jeffrey Hinton had invented and uh uh in just a few years before in 2006 2005 these hierarchical neural networks and it seemed like a
51:48super promising idea even back then uh to us that come across it in academia and um and then the other thing we bet on was reinforcement learning and the combination of that right which uh again is coming back into vogue but it was also important for us to solve something like Alph Go. So you need both parts of that. You need the deep learning to kind of model the environment and the world and then you need the reinforcement learning to make the plans and and and the solutions and take action in the world and um so there was two reasons we
52:17bet on that even when it it was just sort of the beginning of it is that we we knew that the classical methods these expert systems would not scale and actually that's again one of the things I learned here and also my posttock at MIT was that was that they were the kind of churches if you like of the classical methods these expert systems And actually that's something else you can learn here is like you know in your university courses is not just what to do but also what not to do and why it may not like like I sort of thought about it and I felt it would never scale
52:46to uh uh the kinds of problems that I wanted to solve with AI. Um whereas the learning system seemed to have unlimited potential although they were a lot harder to get anything to do uh significant to do at the beginning right and this was the problem because they weren't scaled up enough. And the other the the other reason we started Deepmind 2010 is we could also see the computing uh paradigm was shifting on the hardware side GPUs and other things uh uh which of course also invented for gaming and turns out everything's a matrix multiplication right intelligence and
53:16gaming and computer graphics. So um and so all of those different influences came together and the you know the understanding of neuroscience and fMRI machines and neuroscience had advanced a lot in that previous 10 years too. So I felt it was the perfect time to sort of bring all of that together uh back in 2010. And we were betting on that not necessarily because we knew it would work but we knew that the other methods we were pretty confident would not work.
53:42Uh and that's what the AI winters were about basically with people trying to push those expert systems. Um and then the second question you know um I think I wouldn't say that well first of all it is surprising that some of these things converge and actually we weren't sure.
53:58So, you know that Atari stuff I showed you. So, for the first couple of years, nothing worked, right? So, we couldn't even get one point on, if some of you remember Pong, the the you know, one of the first computer games, we couldn't this sort of tennis bat and ball game, simplest game you could imagine. We weren't able to get a point on it. So, we were wondering whether were we like 10 20 years too early like like Babage turned out to be with his difference engine, right? Amazing ideas, it worked, but he was just in the end 50 years or 100 years too early. And uh and I always say you want to be like 5 years ahead of
54:28your time, not 50 years ahead. Otherwise, you know, you you you'll be in for a lot of pain like like Babish was. And um and and so we were worried about that, but then it did converge.
54:38And then that gave us confidence to to tackle the harder problems. And I think if you ask him, I think the last part of your question was about that things in nature. Well, the things there what I'm thinking is they're not sub-optimal.
54:48They're actually probably pretty optimal because they've they've gone through an evolutionary process. Not just life with biology, but actually geologically and physically. You know, asteroids and physical phenomena combined together, they survive some amount of time because they're stable over time, right? And if they're stable over time, then there's probably some structure that's learnable. That that would be my conjecture.
55:10Great. Uh there's a question down here.
55:16What do you think about uh building high bandwidth brain machine interfaces and implantable memory and reasoning modules so that humans can be further empowered to make discoveries autonomously as opposed to only talking to AI in the cloud?
55:29Yeah, I I I love that area and I've um carefully followed it and and help people building like EEG caps and and things. Um, of course the problem is on resolution of these things um to get the readouts from the brain and also ideally you'd want it read and write. Um, but I'm very fascinated by projects like Neurolink or you know chips in the brain. Obviously right now um that's for uh you know veterans and people to get function back in their body. I think it's going to be amazing things on that
55:58where I think people will be able to gain uh be able to walk again if they've broken their back and things like that.
56:03I think there's going to be some credible advances in medical sciences that will be amazing. But then beyond that, maybe uh if things, you know, if becomes routine and it's surgery safe and there's safe ways of doing this, I could imagine um that would be one way for us to keep up with the with the technology. And in some senses, it's no different to what we already have today with our technologies all around us. We all have our phones all, you know, 247 with us and computers and other things.
56:28So, we're already symbiotic almost with our technology. It would be one step further, of course, to have that, but I'm not sure. I mean maybe that's for the philosophers in the room to answer what the difference what whether there's a howard boundary there if if with the technology if it's attached to you versus you know it's just something you carry around with you um all the time.
56:48Great. Some question over here that's one just Sure.
57:00Oh hello. Ah there we there we go. Uh what do you think about the speed at which artificial intelligence is developing and its effects on sort of economic developments? There are a lot of people out there who are deciding careers right now who given the sort of rapid change in the landscape make it really difficult to kind of predict what they should go into.
57:22Yes, it's it's it's um it's a very complicated one because as you say things are changing at kind of lightning speed. Um we were just discussing it with Alistister actually earlier even designing three-year computer science courses is quite difficult given that this the underlying material changes in with you know less than 3 years. Um I think the only thing we can say for sure is there's going to be a lot of change but I think that brings with it disruption and opportunity. So um I mean I'll just give you an example on coding
57:51which I don't know if you're a computer scientist but you know so I still would recommend that you get good at coding uh and math because I think you'll be able to use these new tools in a much deeper way if you have an understanding of how they're built. But on the other hand, um I think coding is going to be more available to [clears throat] many many more types of people uh because of the way that AI you'll be able to program sort of in natural language probably rather than quite a complicated uh computer language. And so that will open
58:21up field fields for creative people to you know build games, make films, make applications that um maybe is more the kind of balance of that is more on the creative side than it is on the engineering side. But I also think that will enhance engineers to be able to do 10x of what they can do today. So I still I think it's difficult to know.
58:42Um, but what I would say is just uh to um focus on uh kind of embracing those tools in your spare time and and making training yourself to be really good at picking up new information extremely quickly [clears throat] because I think that's that's basically what's going to happen in the next 10 years.
59:00Okay. Um we've got one question just on the here the yellow and black top.
59:06Um is this Yeah. uh do you think that there are any biological processes or behaviors or patterns which can't be modeled with existing um deep learning techniques? I'm not saying like throw more computer at it until it works and just make a bigger and bigger model. Do you think there are some processes which physically cannot be modeled with the architecture we have?
59:27Um I mean there's certainly lots of processes that can't be modeled uh today with but but but again this is sort of goes back to what I said at the end of the talk. I'm not sure in the limit uh that there are I think in the end uh if physics can can sort of solve it and there's some structure there to be learned uh probably with enough examples one could reverse engineer a model of that uh and then um I don't see any theoretical reason why uh a classical system um or be a very complex one could
59:57not make predictions or or simulations of that biological system. So I I don't really see uh any in the limit uh what what the limit of that could be. I mean there are lots of abstract things like factorizing large numbers uh cryptography where it's sort of it's sort of human-made systems right where there may not be any structure. I mean there may be structure in the in the in the natural numbers. Lots of people conjecture there is. If there is then that will also be learnable. If there isn't and it's kind of uniform
1:00:26distribution then you would need uh quantum computer to to you know crack uh cryptography things like that. So those are open conjectures but I think most things in nature have evolved and um over geological or biological physics time and so then that suggests to me there is some structure to learn. So that makes the the search or the prediction potentially tractable.
1:00:51Okay. in uh la last question then um we'll have the person in the pink shirt.
1:00:57Thank you. This question comes from the behalf of the Cambridge University Game Development Society. Uh so you mentioned uh you mentioned about the Genie2 model and how currently it's uh stable for a few seconds of consistency and you hope to eventually have that a few at a few minutes. Uh but I suppose the question that our society has is um games that we actually play have consistency that's indefinite. you know, when you're playing Minecraft, you expect to turn around and the village is still there, right? Yes.
1:01:24Um, so do you see your current model being integrated into a workflow or what exactly how do you see AI and your model and what you're working on? Uh, how do you see it integrating into game development in the coming decades?
1:01:39Yeah. Well, look, um, I I think there's many ways AI is going to come in. So, uh, one is, uh, the tools to build the assets you need for games. So 3D models, animations, I think that's all uh going to come in in the next um you know couple of years. Um I think you can think of AI as well for game balancing.
1:01:59So you imagine you design a game and overnight it could play a million playthroughs of that game and then in the morning you could just get a report as a game designer of like these things are unbalanced, right? Or uh reduce the power of this unit or whatever it is. Um I think also bug testing for open world games. So, I used to make I used to make simulation games, open world games.
1:02:20They're a nightmare to bug test because the whole point of them is that the player can almost do almost anything and then the game will react to you. So, how do you test, you know, 10 million people having their own unique journey through your game while actually uh having AI players play it before you release it um could help you solve a lot of those bugs. And then finally, I mean, I think excitingly as well is like AI characters that are much more lifelike that move the story line on. He used to dream about that massive multiplayer worlds where, you know, actually the AI
1:02:48characters were intelligent and actually updated their beliefs and um their storylines based on what the players were doing. So, it felt like a much more uh living realistic world. Um, and I think we're on the cusp of having of building those types of of of games. And then finally, the world model we're building that's more about general AI and and can you actually model the it's an expression of being able to understand the world. Does your model understand the world? Well, if it can if it can generate it for some amount of
1:03:17time, then obviously it must be in some sense understanding it's called an empirical evidence that it's understanding something about the underlying physics. So, um so that's more for the general intelligence rather than I think we're going to maybe one day we'll have this holiday thing that you can just imagine and it's just all there around you. Um probably we will be able to have that once we have AGI but um I think that's still a ways off.
1:03:41Great. Uh well it seems like a nice place to finish a question on games returning back to games but uh thank you all so much for coming and a particular special thank you to Deis coming and talk to us today. So thank you [applause]
