Demis Hassabis Reveals the Future of AI - This Cambridge Lecture Is Mind-Blowing @AICreatorForge — Transcript
Full transcript
- 0:00[applause]
- 0:03>> Thanks Alister for that lovely
- 0:05introduction and it's so great to be
- 0:07back at Cambridge. I always have a warm
- 0:09feeling when I sort of my homecoming
- 0:12back to Cambridge and specifically this
- 0:14lecture hall as Alister reminded me I
- 0:15think it is the first lecture hall I was
- 0:17in it's always been my favorite lecture
- 0:18hall.
- 0:19I remember telling and I see a lot of my
- 0:21old friends here from my Cambridge days
- 0:24Aaron I think about that one day maybe
- 0:27I'd come back to give a lecture in here
- 0:30and talk about announcing AGI and maybe
- 0:32a robot would walk on and
- 0:34astound everyone. I'm not going to do
- 0:36that today
- 0:37to to disappoint you but maybe in a few
- 0:40years time I'll come back again and I'll
- 0:41give that lecture.
- 0:43But it you know this is amazing place
- 0:45it's such an inspiring place and I'm
- 0:46going to talk a little bit about how
- 0:48Cambridge has inspired my whole career
- 0:50actually and hopefully it's going to do
- 0:51the same for many of you the students in
- 0:54in the room.
- 0:56So for me it started
- 0:58get my journey on AI started with games
- 1:01and specifically chess as Alister
- 1:03mentioned. So I was playing chess from
- 1:04the age of four years old and very
- 1:07seriously for the England junior teams
- 1:08and things like that and it got me
- 1:10thinking about thinking itself. You know
- 1:12how does our mind come up with these
- 1:14plans with these ideas how do we problem
- 1:17solve and how can we improve? Obviously
- 1:19when you're playing chess at a young age
- 1:21you're trying to play competitively
- 1:22you're trying to improve that process.
- 1:24And it was fascinating to me perhaps
- 1:26more fascinating than even the games I
- 1:28was playing was the actual mental
- 1:30processes behind it.
- 1:32And in fact AI
- 1:34and can computers I came across
- 1:36computers and AI for the first time in
- 1:38the context of chess and trying to use
- 1:40very early chess computers like the one
- 1:42on the right here I think this was my
- 1:44first ever chess computer there were
- 1:46physical boards where you had to
- 1:47actually had to press the squares down
- 1:49to make move the pieces
- 1:51and of course we were supposed to be
- 1:52using these chess computers to train
- 1:54opening theory and and learn more about
- 1:56chess but I remember being fascinated by
- 1:59the fact that someone had programmed
- 2:01this lump of inanimate plastic to
- 2:03actually play chess really well against
- 2:06you. And I was sort of really fascinated
- 2:08by how that was done and how how someone
- 2:11could program something like that. And I
- 2:13ended up experimenting myself in my
- 2:15early teenage years with the Amiga 500
- 2:18computer, amazing home computer back in
- 2:20the in the late '80s and early '90s, and
- 2:23building those kinds of AI programs
- 2:25myself to play games like Othello. And
- 2:27really that was the first my first taste
- 2:29of AI and and I was hooked from then on
- 2:31and that's, you know, I decided from
- 2:32very early on that I would spend my my
- 2:35entire career trying to push the
- 2:37frontiers of of this technology.
- 2:41So then that led me to
- 2:43to Cambridge
- 2:44and which was really my three years here
- 2:46were incredibly formative for me.
- 2:49And I went to a school in comprehensive
- 2:52school in North London. No one had ever
- 2:54got to sort of no one had gone to
- 2:56Oxbridge for in sort of living memory,
- 2:58but and the reason I wanted to come to
- 3:00Cambridge was all these inspiring
- 3:02stories that I'd heard about
- 3:04what happened at Cambridge, all these
- 3:06amazing people that I used to read about
- 3:07their biographies and the work they'd
- 3:09done. Especially people like Crick and
- 3:12Watson in the in the top left there. And
- 3:14I remember a particularly a film The
- 3:16Race for the Double Helix, an amazing
- 3:18film from the '80s if you haven't seen
- 3:20it with Jeff Goldblum, one of his early
- 3:22parts so he you know with all the
- 3:23enthusiasm that he plays all of his
- 3:25parts as he was Watson. And they were
- 3:28having just such an amazing time
- 3:30discovering, roaming around Cambridge,
- 3:33you know, working on things like DNA and
- 3:34I thought like that's what I want a
- 3:36piece of that. And I want to feel what
- 3:38it's like to be at the frontier of
- 3:41discovery and what could be more
- 3:42exhilarating. And and that film actually
- 3:44really brought it to life what that
- 3:47might be like.
- 3:49And then of course all of my my my
- 3:51heroes, my scientific heroes, a lot of
- 3:53them had gone through Cambridge. People
- 3:55like Alan Turing and Charles Babbage, of
- 3:57course, and in the lecture hall that we
- 3:59we now sit in.
- 4:01Um and and and even places like the
- 4:03Eagle pub where if you start at Queen's
- 4:05College, one of the tours they give you
- 4:07on the first day is to go and show you
- 4:09the the the reputed sort of
- 4:11table that they were discussing and and
- 4:13and at the DNA structure around. And you
- 4:17can't help but be inspired by that and
- 4:19walking down King's Parade, um and I
- 4:22almost felt like the the the
- 4:24intellectual giants of the past were
- 4:26were almost speaking to you from the
- 4:27stones.
- 4:28Um and and that's how I felt, you know,
- 4:30going for a late night burger at
- 4:32Gardenia's.
- 4:33Um but that was what was inspiring me
- 4:35around, like all of these amazing people
- 4:37that had had walked those same steps
- 4:39over hundreds of years. And and that's
- 4:41the history that is sort of unrivaled
- 4:43here at Cambridge that I think we can
- 4:45still draw on and take inspiration from
- 4:47today. Uh and then there's a picture of
- 4:50me and Aaron there, one of my best
- 4:51friends from Queen's, you know,
- 4:53obviously on the mathematical bridge
- 4:55there.
- 4:56And then finally, you know, what when we
- 4:58Alister mentioned obviously the Nobel
- 5:00Prize and it was a honor of a lifetime
- 5:02to go and collect that in in Stockholm
- 5:05um in December. Um amazing one week of
- 5:08of activities, but my favorite activity
- 5:10was when you get to sign the Nobel book
- 5:12in the Nobel Foundation. And and and
- 5:15that's the book there and one of my
- 5:17pictures of and I started sort of
- 5:18leafing through the book. You saw your
- 5:20name and then you leaf back and you
- 5:21wonder is you know, is Crick in there?
- 5:23And of course he is. And then you go
- 5:25back further and then Einstein's
- 5:26signature's there and it's just
- 5:27mind-blowing, really. And I started to
- 5:29spend an hour just photoing every page
- 5:31of the of the book.
- 5:34Um
- 5:34so, you know, it's full circle for me of
- 5:36of of that picture and and then really
- 5:39seeing that film in the in the in the in
- 5:41the late '80s.
- 5:44So, then
- 5:46you know, in 2010 we started DeepMind in
- 5:48London as really at the time it was a
- 5:51kind of an Apollo program effort is the
- 5:53way we thought of it for trying to build
- 5:55artificial general intelligence. You
- 5:57know, AI that was truly general and
- 6:00could perform all the cognitive
- 6:02capabilities that that humans are
- 6:04capable of. So, it would be a truly
- 6:06general AI system. In fact, you know,
- 6:09the idea for that really comes from
- 6:10Turing and Turing machines. So,
- 6:12something that's able to compute
- 6:15anything that is computable as Turing
- 6:17showed with his Turing machines. And
- 6:19really that's been the foundation for me
- 6:22of and one of the main things that I
- 6:24carried with me from the lectures here
- 6:26at Cambridge was all these theoretical
- 6:28underpinnings of computer science and
- 6:30computation theory that that people like
- 6:32Turing and Shannon famously did in the
- 6:3540s and the 50s.
- 6:38So, we started in 2010 and it's amazing
- 6:40like it's 15 years ago, which in some
- 6:42ways isn't that long ago, but when we
- 6:44started out DeepMind, almost nobody was
- 6:46working on AI, which is hard to believe
- 6:48today given that almost everyone seems
- 6:50to be working on AI today
- 6:52in just a matter of just over a decade,
- 6:54things have gone, you know, accelerated
- 6:56incredibly. And obviously, we've been
- 6:57part of that very exciting journey.
- 7:01So, our mission at DeepMind from the
- 7:02beginning was
- 7:05we talk about building AI responsibly to
- 7:06benefit humanity, but the way we used to
- 7:08articulate it when we started out was in
- 7:11a two-step process. Step one, solve
- 7:13intelligence. Step two, use it to solve
- 7:15everything else. And and this seemed
- 7:18very outlandish at the time in 2010 and
- 7:20you can imagine trying to pitch a
- 7:22venture capitalist on that on the basis
- 7:24of that mission. It seemed pretty crazy,
- 7:26but I I still fundamentally believe in
- 7:28that today and I think more and more
- 7:30people are realizing that AI
- 7:32built in this general way could have
- 7:34these kind of profound and
- 7:36transformative impact on almost any
- 7:38field,
- 7:39which is obviously the second part of
- 7:41that mission statement, and I think
- 7:43for me that that involves accelerating
- 7:46scientific discovery itself and medicine
- 7:49and advancing our understanding of the
- 7:51universe around us.
- 7:55So back when we started out, there were
- 7:56basically two ways, and in fact when I
- 7:58was studying here in the '90s, there
- 8:00were two ways to build AI, broadly
- 8:02speaking. There's the expert system way,
- 8:04which is you kind of pre-program an
- 8:06expert system directly with the
- 8:08solution.
- 8:10Uh things like Deep Blue that beat Garry
- 8:11Kasparov at chess very famously in the
- 8:13'90s, actually while while I was
- 8:15studying here.
- 8:16That would be the pinnacle example
- 8:18probably of an expert system. But the
- 8:20problem with these expert systems was,
- 8:22and why they never really scaled to full
- 8:24general intelligence, is they can't deal
- 8:26with the unexpected.
- 8:29If you you know, if something unexpected
- 8:30happens that you didn't already cater
- 8:32for, there's nothing in the system that
- 8:34will allow it to to deal with that.
- 8:37And they're sort of were inspired by
- 8:38logic systems,
- 8:40and they were quite rigid and and
- 8:41fragile and brittle because of that.
- 8:44Whereas the modern-day approaches
- 8:46are built on learning systems. So these
- 8:47are systems that are able to learn for
- 8:49themselves and learn directly from
- 8:50experience or data from first
- 8:52principles,
- 8:53and really inspired by more neuroscience
- 8:55ideas.
- 8:57And the obviously the promise of these
- 8:58systems that we have today is that they
- 9:00can go beyond the knowledge potentially
- 9:03that we as the programmers or the system
- 9:05designers already know how to solve. And
- 9:07of course that's extremely valuable in
- 9:09areas like
- 9:11scientific discovery.
- 9:14So we started in the in the early 2010s
- 9:16with with games, of course, and I've
- 9:18used games many times in my life. First
- 9:20of all, to train my own mind, then I
- 9:22used to build games and AI for computer
- 9:25games, and then finally in a third way
- 9:27to train up our AI systems. And games
- 9:29are the perfect proving ground for AI
- 9:31systems.
- 9:32You can start with very simple games
- 9:34like Atari games from the '70s,
- 9:36and really at this this this system DQN
- 9:39was the first time anyone had built an
- 9:41end-to-end learning system that could
- 9:43learn directly from raw data. So, in
- 9:45this case, the raw pixels on the screen,
- 9:48and it's it's not told anything about
- 9:49the games or anything about what it's
- 9:51controlling. It's just told to maximize
- 9:53the score based on this
- 9:56video stream input, pixel stream input.
- 10:00So, we were able to master all different
- 10:02Atari games
- 10:04in sort of around 2013. Then we took
- 10:06these
- 10:07systems and we scaled them up to really
- 10:09the I would say the grand challenge of
- 10:12games AI, which is can you you you
- 10:15create systems that can play the game of
- 10:17Go
- 10:18at world champion level or beyond. And
- 10:21Go, of course, is probably the most
- 10:23complex game that humans have ever
- 10:25invented. It's thousands of years old,
- 10:27so it's also the oldest game and one of
- 10:29the most elegant games.
- 10:31But, one of the ways you can just see
- 10:33the complexity of Go is that there are
- 10:3410 to the power 170 possible positions
- 10:37in Go.
- 10:38So, that's more way more than there are
- 10:40atoms in the observable universe. And
- 10:43the important reason point about that is
- 10:45that you cannot come up with a strategy
- 10:47in Go using brute force techniques. It
- 10:49would be impossible. It'll be totally
- 10:51intractable. So, you have to do
- 10:52something much smarter. And we famously
- 10:55in basically 2016 won a million-dollar
- 10:58challenge match against 10-times world
- 11:01champion Lee Sedol and one of the
- 11:03legends of the game as South Korean
- 11:05grandmaster and was watched by 200
- 11:07million people around the world. And not
- 11:09only did AlphaGo, our system, win that
- 11:12match, importantly it actually came up
- 11:14with new, original Go strategies, even
- 11:17though
- 11:18we've played Go for thousands of years
- 11:21and professionally for hundreds of
- 11:22years, it was still able to find
- 11:24never-seen-before strategies. Most
- 11:26famously, this move 37 move here in red
- 11:31uh game two, which um if you watch the
- 11:33documentary on this, you'll see uh which
- 11:35is on YouTube, you'll see uh that how
- 11:37surprised the the best players in the
- 11:39world were. They were commentating on
- 11:41the game uh about this move. It was sort
- 11:43of an unthinkable move, and yet it
- 11:45decided um the game this game two in
- 11:48favor of AlphaGo um
- 11:50100 moves later.
- 11:52So, again, that so told me about the
- 11:54potential for these types of systems to
- 11:56invent and discover new knowledge. So,
- 11:59here, of course, we're just talking
- 12:00about game knowledge, but um obviously,
- 12:03my dream was to generalize this to all
- 12:05areas of um scientific discovery.
- 12:10So, how do these systems work? Um we we
- 12:13basically train up these neural networks
- 12:15through a system of self-play. So, this
- 12:17is actually AlphaGo and also um
- 12:20subsequent systems uh that AlphaGo Zero
- 12:23and AlphaZero that generalized what
- 12:25we've done for Go to play any two-player
- 12:27uh game from scratch. And you start off
- 12:30with a version one of the system that
- 12:32doesn't really know anything about the
- 12:33game, just the rules, and it plays
- 12:35randomly, and you play I'd say 100,000
- 12:38games against itself uh uh of this
- 12:40system, and that creates a new database
- 12:43of um game positions from those 100,000
- 12:46uh games. And from that, you train a
- 12:49second version a a slightly better
- 12:50version of the model, version two, uh
- 12:53that's trained to predict um what the
- 12:55likely moves are to be played in any one
- 12:58position, and also who is more likely to
- 13:00win, which side, black or white, is
- 13:02likely to win from that position, uh and
- 13:05what percentage chance do they have of
- 13:06winning.
- 13:07Uh and then you can use that version two
- 13:10to um you play it against version one uh
- 13:13and in a 100-game match-off, and then if
- 13:16it wins by significant uh margin, so in
- 13:18this case 55% win rate, you replace the
- 13:21version one with version two, and you
- 13:24create a new uh database of games that
- 13:27are slightly higher quality and then you
- 13:29learn a version three system. And if you
- 13:31do this and you repeat it around 17, 18
- 13:34times, you go from playing randomly in
- 13:36the morning to 24 hours or less later,
- 13:39you're you're you're by the version 17
- 13:41or 18, you're stronger than the world
- 13:43champion level. So, it's quite an
- 13:45incredible thing process to see this
- 13:47this self-improvement process playing
- 13:50out in a very, very short amount of
- 13:52time.
- 13:54So, if we think about what these neural
- 13:56networks are doing is you're kind of
- 13:58reducing down this intractable search
- 14:01space of, you know, 10 to the power 170
- 14:03possibilities down to something that's
- 14:05much more tractable in a few minutes of
- 14:08compute time. And it's doing this by
- 14:11narrowing down using the neural network
- 14:14to efficiently guide the search
- 14:16mechanism. So, if you think about this
- 14:18tree of possibilities as and each of the
- 14:21nodes in this tree is a go position,
- 14:24then instead of having to to look at
- 14:26sort of every possibility, you can
- 14:28actually use the neural network to guide
- 14:30you just down the most the most
- 14:33most interesting and most useful lines
- 14:36to examine. So, in this case, the ones
- 14:38in blue. And then
- 14:40after you've run out of thinking time,
- 14:42you pick the best line, the most
- 14:43promising line that you've seen thus
- 14:45far. So, I I in this case, this this
- 14:48particular line in purple.
- 14:52So, this then leads to, you know, we we
- 14:53then play not just go, but any
- 14:56two-player perfect information game and
- 14:58it was even able to discover new
- 15:00strategies and new styles of playing
- 15:02chess,
- 15:03which which is kind of extraordinary but
- 15:05given that chess computers were so
- 15:07strong already. So, programs like
- 15:09Stockfish. And AlphaZero was able to
- 15:12beat Stockfish at the time
- 15:15in at chess, which is almost sort of
- 15:17impossible to do. And not only did it
- 15:19beat Stockfish, so AlphaZero here is in
- 15:21in white uh playing against Stockfish
- 15:23who's black. But, in this particular
- 15:25position, one of the most famous uh
- 15:27games that Alpha Zero played is called
- 15:28the immortal Zugzwang game. Um
- 15:31white is winning here um uh because it
- 15:34favors mobility over material. So, most
- 15:36chess computers favor material, and
- 15:38you'll see that black Those of you who
- 15:40play chess, you'll see that black has
- 15:41more material, um but actually can't
- 15:43move any of its pieces. They're all
- 15:45stuck in the corner. And this is what
- 15:47Alpha Zero sacrificed material for this
- 15:49mobility. And actually, for human
- 15:52grandmasters uh and top chess players,
- 15:54this is not only very effective style,
- 15:56it's very beautiful aesthetic style uh
- 15:58to and uh to play chess in. So, um it's
- 16:01kind of amazing that Alpha Zero was able
- 16:03to sort of discover this new way, this
- 16:06new dynamic way of playing. And in fact,
- 16:08some of the top chess players in the
- 16:09world commentated about this. So, Garry
- 16:11Kasparov, my all-time sort of favorite
- 16:13chess player, he said, "The programs
- 16:15usually reflect priorities and
- 16:16prejudices of the programmer, but
- 16:18because Alpha Zero learns for itself, I
- 16:20would say that its style reflects the
- 16:22truth." And then, the current world
- 16:24champion at the time, Magnus Carlsen,
- 16:26said, uh you know, read uh and looked at
- 16:28these games and and read about the books
- 16:30that were written about Alpha Zero, and
- 16:32I said, "I've been influenced by one of
- 16:34my heroes recently, one of which is
- 16:36Alpha Zero." So, he actually
- 16:37incorporated a lot of these ideas uh
- 16:40into his own game to dominate the chess
- 16:42scene for, you know, almost a decade
- 16:44now.
- 16:47So, we did all these landmark
- 16:48breakthroughs in games AI and over the
- 16:50first, I guess, sort of decade of of of
- 16:53DeepMind's existence, but of course,
- 16:55these were just the training ground for
- 16:56what we wanted to do and uh was just a a
- 16:59means to an end. It wasn't the end in
- 17:01itself to play these games, much as I
- 17:03love games. Um it was to create these
- 17:05algorithms that could be generally
- 17:07useful for tackling real-world problems.
- 17:12So, um what we look for in real-world
- 17:14problems, not only scientific problems,
- 17:16uh but actually uh industrial problems
- 17:18as well. And we look for sort of three
- 17:21different criteria that make it suitable
- 17:23the problem suitable to be tackled by
- 17:26these types of AI systems and and ideas
- 17:29and algorithms that we developed for
- 17:30playing games. Number one is we look for
- 17:33problems that can be described as
- 17:35massive combinatorial search spaces. So
- 17:39So usually far too complex, far too many
- 17:41combinations to brute force the
- 17:43solution. So but maybe there's some kind
- 17:46of structure that we can learn about
- 17:48with our neural networks that can guide
- 17:50that search very efficiently.
- 17:52Secondly, we look for problems that can
- 17:54be described with a clear objective
- 17:56function or some sort of metric that you
- 17:58can optimize against. So in games that's
- 18:01very easy. It's things like maximizing
- 18:03the score or winning the game. But
- 18:06actually there's a lot of real world
- 18:07problems that you can boil down to a few
- 18:10metrics or few objective functions that
- 18:12you're trying to maximize.
- 18:14And then finally, of course you need
- 18:15quite a lot of data or experience to
- 18:18learn from and or ideally it's and a
- 18:21kind of accurate and efficient simulator
- 18:23so you can generate more synthetic data
- 18:26to augment the real data that you have.
- 18:30And it turns out that there are a lot of
- 18:32problems that can be couched in these
- 18:34terms if you're looking at the problem
- 18:37from this angle.
- 18:38And including many many problems
- 18:40important problems in science.
- 18:42And the one that I always had in mind
- 18:44actually from my days of first seeing
- 18:46coming across the problem here at
- 18:47Cambridge as an undergrad is the protein
- 18:49folding problem which I'll just quickly
- 18:51describe to you for those that don't
- 18:53know about biology and proteins. So
- 18:55proteins are incredibly important.
- 18:58They're they're the building blocks of
- 18:59life. Pretty much every every
- 19:02function in the living body depends on
- 19:04proteins from your neurons firing to
- 19:07your muscle fibers twitching. So really
- 19:10proteins are uh makes life possible.
- 19:14And the protein folding problem then is
- 19:15is really easy to describe. It's
- 19:17basically a protein is is defined by its
- 19:20uh it's
- 19:22defined by its gene sequence, its
- 19:23genetic sequence, uh which then
- 19:25specifies an amino acid sequence, uh
- 19:28which uh in nature then folds up
- 19:30spontaneously into a usually a very
- 19:33beautiful uh protein structure. So, you
- 19:36go from this genetic sequence to a
- 19:37protein structure. And the reason the
- 19:39protein structure, the 3D structure is
- 19:41very important is it goes a long way to
- 19:44defining what function it has, what it
- 19:46does uh in the body. So, um it's not
- 19:49doesn't totally describe the function,
- 19:51but it it has a big part to play in what
- 19:53what it actually does uh in nature. So,
- 19:56the protein folding problem then is is
- 19:57this problem of can you predict the
- 19:59protein structure directly from this
- 20:02one-dimensional amino acid sequence? Can
- 20:05you predict that computationally, that
- 20:06incredible 3D structure uh from that
- 20:09sequence?
- 20:12So, why is this such a hard problem?
- 20:14Well, um
- 20:15Levinthal is a famous researcher,
- 20:17protein researcher in the in the '60s
- 20:19described this uh
- 20:21conjecture that that that became called
- 20:24it became known as the Levinthal's
- 20:25paradox, which is that he calculated
- 20:27there roughly, you know, 10 to the 300
- 20:30possible shapes that an average protein
- 20:33can take. Um and yet somehow in nature
- 20:36and in the body, these proteins fold up
- 20:38spontaneously in a matter of
- 20:40milliseconds. So, that's the paradox.
- 20:43Like, if there's so many possibilities,
- 20:45how does nature do this, right? And
- 20:47basically, how does physics uh achieve
- 20:49this?
- 20:50Um and that gives you hope that this
- 20:52must be tractable computationally in
- 20:54some reasonable amount of time because
- 20:56uh physics does solve this problem uh um
- 20:59you know, uh billions of times a second
- 21:01in the body.
- 21:03And um furthermore, what attracted me to
- 21:04this um problem was that there was a
- 21:07bi-annual competition called the CASP
- 21:08competition, which is like you can think
- 21:10of it as like the Olympics for protein
- 21:12folding. And it happens every 2 years
- 21:14and it's run by some amazing people led
- 21:16by Professor John Moult,
- 21:18University of Maryland. And it's been
- 21:21running for since 1994 and it's a great
- 21:23competition because they work with
- 21:26experimentalists and who who
- 21:28painstakingly find these structures
- 21:30using very exotic and
- 21:33expensive equipment like electron
- 21:35microscopes. And and they newly
- 21:38discovered structures that haven't been
- 21:39published yet, they put it into the
- 21:41competition. So you actually the the the
- 21:43the competition organizers know what the
- 21:46ground truth is, but the computational
- 21:48teams, you know, hundreds of teams enter
- 21:50every year every every competition every
- 21:52couple of years, and they try with their
- 21:54computational methods to predict that
- 21:56those structures. And there's usually
- 21:57around 100 proteins that are in the
- 21:59competition. And then at the end of the
- 22:01summer, they reveal what the true
- 22:03structures are and you can compare the
- 22:05predicted ones and their distances, the
- 22:07error in the predictions, to to the real
- 22:10structures.
- 22:13So we entered AlphaFold 1 actually for
- 22:16the first time in 2018 and we started
- 22:19this AlphaFold project in 2016, actually
- 22:22pretty much the day after we we got back
- 22:25from the AlphaGo match in Seoul for in
- 22:27Korea. We felt that and I felt that we
- 22:29were ready, we had the techniques that
- 22:31were mature enough and ready to now be
- 22:33applied outside of games and to try and
- 22:36tackle really meaningful
- 22:39problems. We call them kind of root node
- 22:40problems because they open if they could
- 22:42be solved, they they open up whole new
- 22:44branches and avenues of discovery that
- 22:46can be built on top. And protein folding
- 22:48was was a prime example of that. So we
- 22:51started working on it 2016. We we
- 22:53AlphaFold 1 was ready after a couple of
- 22:55years and we entered it into the CASP 13
- 22:57competition. And you can see for the
- 22:59decade prior,
- 23:01uh bar charts are showing the winning
- 23:03score of the winning team
- 23:05uh uh in the hardest category, actually,
- 23:08the hardest proteins that were being
- 23:09predicted. And you can sort of think of
- 23:11it as a percentage accuracy of how many
- 23:14of the atomic how many of these amino
- 23:16acids have you got in the right position
- 23:17within a certain tolerance?
- 23:19Uh within the sort of width of an atom
- 23:21you need to predict it within. And um
- 23:24you can see there was sort of not much
- 23:25progress for a decade and we we were
- 23:27stuck at this um 60 points level. And
- 23:31you we for the
- 23:37which is effectively if you got to 90,
- 23:39you would be within the width of an
- 23:41atom. So, you'd be atomic accuracy. And
- 23:43that's what we were told by
- 23:44experimentalists was the accuracy you'd
- 23:46had to reach so that it was competitive
- 23:48with experimental methods so that
- 23:50experimentalists could actually rely on
- 23:53these predictions rather than having to
- 23:55necessarily do the laborious,
- 23:56painstaking uh work to um uh to find
- 24:00that structure. And uh just as a rule of
- 24:02thumb, my biologist friends would always
- 24:04tell me that, you know, it it it take a
- 24:06a a PhD student their entire PhD, so 4-5
- 24:09years, to fold uh to find the structure
- 24:11of just one protein. Um and um and and
- 24:15there are 200 million proteins uh known
- 24:17to science and 20,000 uh uh proteins in
- 24:21the human uh uh proteome.
- 24:23So, uh so with AlphaFold 1, we were able
- 24:26to uh win this competition and and sort
- 24:29of um be better by almost 50% than the
- 24:31next best system. And AlphaFold 1 for
- 24:33the first time introduced machine
- 24:35learning techniques as the main
- 24:37component of the system uh for the first
- 24:39time. Uh but it was not enough to to
- 24:42reach this atomic accuracy. We actually
- 24:43had to go back to the drawing board with
- 24:45what we'd learned and re-architected uh
- 24:48for AlphaFold 2 from scratch uh
- 24:51using all the learnings we had from
- 24:52AlphaFold 1 to finally reach this atomic
- 24:54accuracy. And that led the organizers to
- 24:57declare that the problem had been solved
- 24:59in uh the end of 2020.
- 25:02So, this is an example of how AlphaFold
- 25:04works visually. So, you can see on the
- 25:06left-hand side here is a very complex
- 25:08protein. The ground truth is in green,
- 25:11the predicted structure is in blue, and
- 25:13you can see how closely the blue
- 25:15overlaps the green.
- 25:17And then on the right-hand side, you can
- 25:19see how AlphaFold 2 works. It sort of
- 25:21builds up that structure in an iterative
- 25:23process. It kind of recycles itself
- 25:25actually over 192 steps,
- 25:27and then builds out starts as a as a
- 25:29scrunched ball of of protein matter,
- 25:32amino acids, and then it builds out more
- 25:34and more plausible structure.
- 25:36At the end, it sort of refines the last
- 25:38parts of it until it has the finished
- 25:40prediction.
- 25:42We we immediately then tried to because
- 25:46AlphaFold is so accurate, but it's not
- 25:48just accurate, it's also extremely fast.
- 25:50It's able to fold proteins in a matter
- 25:52of seconds. An average protein, we
- 25:54realized quite quickly that we could
- 25:56actually fold all 200 million proteins
- 25:59known to science. Over the course of a
- 26:02year, we used a lot of computers on the
- 26:04Google Cloud to to fold all of them and
- 26:06then put them out freely
- 26:08on a database with our colleagues at
- 26:10EMBL-EBI just up the road at the Sanger
- 26:13Center just outside of Cambridge. And we
- 26:16provided that for free unrestricted
- 26:18access to anyone in the world to use it.
- 26:21And that 200 million proteins, if you
- 26:23think about how long it takes to do that
- 26:25experimentally, 4-5 years, it's kind of
- 26:28like a billion years of PhD time done in
- 26:301 year. So,
- 26:33you know, it's it's kind of amazing to
- 26:34think about how much science could be
- 26:36accelerated. And it opened up whole new
- 26:39avenues of exploration because many of
- 26:42these structures, especially for the
- 26:44less well-studied organisms like some
- 26:46certain types of plants, they're very
- 26:48important for science and for
- 26:50agriculture um, and agricultural
- 26:53research, but that there's there's
- 26:54there's that you wouldn't have any
- 26:56almost none of those structures will
- 26:58have been found and be available. So,
- 27:00now those are all available and also 200
- 27:02million, you can look at them at an
- 27:04aggregate level and sort of look at
- 27:06structural uh structures, um, across
- 27:08species, uh, and, uh, kind of meta
- 27:11structures and see what the
- 27:13commonalities are, uh, through
- 27:14evolution. So, there's really
- 27:16interesting new avenues of of uh
- 27:18branches of of structural biology now
- 27:20that are being explored.
- 27:22And of course we thought about safety
- 27:23from the beginning, uh, and we take our
- 27:25responsibility very seriously at the
- 27:27forefront of AI and in this case we, you
- 27:30know, consulted with over 30 biosecurity
- 27:32and bioethics experts to make sure that,
- 27:35um, what we're putting out into the
- 27:36world, um, what the benefits of that far
- 27:38outweighed any any, uh, risks associated
- 27:41with it. And now, uh, I'm very proud to
- 27:44say that, you know, over 2 million
- 27:45researchers are using it, um, from
- 27:47pretty much every country in the world,
- 27:50um, and, uh, and it's been sort of sort
- 27:52of cited over 30,000 times now and it's
- 27:53become a standard tool, uh, in in in
- 27:57biology research now and many of you in
- 27:59the audience who are PhD students,
- 28:00hopefully you're using it and making use
- 28:02of it and it's just, um, part of the
- 28:04standard canon now that is used for
- 28:06biology research.
- 28:09Uh, it's been amazing to see what other
- 28:11researchers have done uh with all of
- 28:13this uh technology and all these
- 28:14structures. Um,
- 28:16I've just called out uh six of my
- 28:18favorite uh uh examples. Uh, people at
- 28:21uh University of Portsmouth are using it
- 28:23a research group is using it to tackle
- 28:25plastic pollution in the environment,
- 28:27trying to design new enzymes which are
- 28:29types of proteins uh that can digest
- 28:31plastic. Um,
- 28:33we're working with the Fleming Center on
- 28:35antibiotic resistance, um,
- 28:38neglected diseases, uh, like tropical
- 28:40diseases that affect the poorer parts of
- 28:42the world. We work with uh Drugs for
- 28:44Neglected Diseases Institute. Um, and
- 28:46here's a good example of where we can
- 28:48accelerate research in those areas
- 28:50where, you know, whether it's malaria,
- 28:52leishmaniasis,
- 28:53um, Zika virus, uh, a lot of those
- 28:55structures are not known. Uh, but now
- 28:57they can go straight to drug discovery
- 28:59because, uh, they have a lot of the
- 29:01information about the structures for
- 29:03those viruses and and bacteria.
- 29:06Uh, and then there's been a lot of
- 29:07fundamental research being done on
- 29:09things like, uh, finding the the
- 29:11structure of the nano pore complex,
- 29:13which is a very important protein that
- 29:15that that lets in nutrients in and out
- 29:17of the nuclear pore of the cell. Um,
- 29:20as amazing work at the Broad Institute
- 29:21done with drug delivery, sort of
- 29:23designing molecular syringes,
- 29:25redesigning, uh, proteins that can
- 29:27deliver drugs targeted to particular
- 29:30part of the body. Uh, and it's even
- 29:32being used in in things like, uh,
- 29:33looking at mechanisms of of fertility.
- 29:36So, the the amount of things that it's
- 29:38being used for is sort of almost every
- 29:40area of biology and medical research now
- 29:42is is, uh, making use of AlphaFold.
- 29:47We've continued in the last few years to
- 29:49develop, uh, more struct more, uh, uh,
- 29:51uh,
- 29:52developments and improving the systems.
- 29:54We released AlphaFold 3, uh, earlier
- 29:56this year, uh, uh, for academics to use,
- 29:59and we've extended now AlphaFold 3 to to
- 30:02deal with, um, interactions. So, um, you
- 30:05can think of AlphaFold 2 as a picture of
- 30:07the the static, uh, protein structure,
- 30:10but really biology is a dynamic process.
- 30:12So, you need to understand how how, um,
- 30:15different biological elements interact
- 30:16with each other. So, of course, proteins
- 30:18with other proteins, but also proteins
- 30:20with others other molecules uh,
- 30:23important to life, things like DNA and
- 30:25RNA, and also ligands. So, uh, small
- 30:28molecules which, uh, you know, things
- 30:30like drug compounds. How does the
- 30:32protein bind with, um, with that
- 30:35compound?
- 30:37And then we have a separate set of work,
- 30:38AlphaProteo, which is doing the reverse
- 30:41of AlphaFold, uh, but making use of
- 30:43still of of the AlphaFold techniques
- 30:45where if you want to design a novel
- 30:47protein, maybe that doesn't exist in
- 30:49nature for a particular job, a
- 30:51particular function, what is the amino
- 30:53acid sequence and the genetic sequence
- 30:56that will give you that structure. So,
- 30:58um so it's kind of like running it in
- 31:00reverse and trying to design new
- 31:02structures that will do novel things um
- 31:05and again could be extremely useful uh
- 31:07in in for for designing drugs and things
- 31:09like that. Antibiotics and antibodies.
- 31:15So, taking a step back then, what is uh
- 31:17or you know, if I look at all the work
- 31:18we've done in the last 15 years, um
- 31:21what's the implications uh for science
- 31:23and and also machine learning? Um and if
- 31:26you think about what we've done with
- 31:27first of all our games work uh and then
- 31:29now with the scientific work that we've
- 31:31been working on of which AlphaFold is
- 31:33our best example, um it's all about
- 31:35making this search tractable. You have
- 31:38this incredibly complex problem, there's
- 31:40many, many possible solutions to the
- 31:42problem, and you've got to find uh the
- 31:44optimal solution, kind of needle in the
- 31:46haystack of that uh enormous
- 31:48combinatorial search space. And you
- 31:51can't do it by brute force, so you have
- 31:53to learn this neural network model um so
- 31:56it sort of learns about the topology of
- 31:58the problem uh so that you can
- 32:00efficiently guide the search to reach
- 32:03your uh to maximize uh or find the
- 32:05optimal solution to the objective that
- 32:07you have in mind. And I think this is an
- 32:10incredibly general way uh
- 32:12uh and general solution, an incredibly
- 32:14general way to approach um a whole
- 32:16myriad of problems.
- 32:17And so, we think about back to the Go
- 32:20example, so we're trying to use these
- 32:21systems to find the best Go move, but
- 32:24you could also change those nodes to be
- 32:27uh chemical compounds, and now you're
- 32:29trying to find the best molecule in
- 32:31chemistry space, in chemical space. And
- 32:34the best molecule, you know, and this
- 32:36this is the beginning of drug design.
- 32:38And and that will bind specifically to
- 32:40the target you're interested in, but
- 32:42nothing else. So, reduces the side
- 32:44effects and the toxicity of that
- 32:46compound. Um and it's a very very
- 32:48similar um techniques that we're using
- 32:51in order to design these molecules now
- 32:53as the next steps as we as we move more
- 32:55and more into drug discovery.
- 32:58So, I think in biology at least, um I
- 33:01feel like we're entering a new era now
- 33:03of what I like to call digital biology.
- 33:05So, you know, I think of biology in its
- 33:07most fundamental level as an information
- 33:09processing system, you know, that's
- 33:11trying to resist entropy around it. And
- 33:14I think that's basically what life is.
- 33:16Um of course, it's a phenomenally
- 33:17complex and emergent information
- 33:20processing system.
- 33:21Uh and I think that's where AI comes in.
- 33:23Just like maths uh and the maths that I
- 33:25learned in this room was the perfect
- 33:27description language for physics um and
- 33:30physic physic physical phenomena, I
- 33:32think that AI is potentially the perfect
- 33:34description language for biology. Um
- 33:36it's perfect for dealing with the
- 33:37complexities of the emergent uh
- 33:39behaviors and interactions that you get
- 33:41in a dynamic system like biology.
- 33:45And I think AlphaFold is a proof point
- 33:47of that. Um and I hope when we look back
- 33:49in 10 years' time, it won't be an
- 33:51isolated breakthrough, but it actually
- 33:53um have sort of heralded in this new era
- 33:56um golden era of digital biology. And
- 33:59we're trying to progress that ourselves.
- 34:01We started a new spin-out company
- 34:02Isomorphic Labs to build on our
- 34:04AlphaFold technology and move more into
- 34:07the chemistry space that I was just
- 34:08talking about and actually try and
- 34:10reimagine drug discovery from first
- 34:12principles with AI. And right now it
- 34:14takes an average of 10 years um for a
- 34:17drug to be developed um and it's
- 34:19extraordinarily expensive. It costs
- 34:21billions and billions of dollars. And
- 34:22so, I'm thinking, why can't we use these
- 34:24techniques to reduce that down from
- 34:26years to months, maybe even one day
- 34:29weeks, just like we reduced down the
- 34:32discovery of protein structures from
- 34:34potentially years down to now minutes
- 34:37and seconds.
- 34:38And so we think of this as a kind of
- 34:41doing science at digital speed. So
- 34:44trying to bring the best of what we do
- 34:46in the technology area to the
- 34:49natural sciences.
- 34:52And my dream one day is to be able to
- 34:54create maybe a kind of virtual cell,
- 34:57computational cell perhaps of something
- 34:59very simple like a yeast cell
- 35:01and that you can actually do run
- 35:03experiments in silico on it and the
- 35:05predictions that you get out of the
- 35:06virtual cell will actually tell you and
- 35:09form your your real world experiments in
- 35:11the lab and you can reduce down a lot of
- 35:13the search that's done in the wet lab
- 35:15and actually more use the wet lab for
- 35:18validation steps rather than the very
- 35:20expensive
- 35:22and slow search process.
- 35:26Of course AI and we've been using AI not
- 35:28just in biology, but it can be used for
- 35:30science, mathematics, medicine more
- 35:33generally
- 35:34and we've had a whole range of
- 35:35breakthroughs not just in the biological
- 35:37sciences, but from things in health like
- 35:40identifying eye disease from retinal
- 35:41scans, discovering new materials,
- 35:44helping with plasma containing fusion
- 35:46reactors, faster algorithms
- 35:49so AI discovering better algorithms for
- 35:51itself like faster matrix
- 35:53multiplication, doing weather prediction
- 35:55and even helping with quantum computers
- 35:57and error correction in quantum
- 35:59computing. And that's just a small
- 36:00example of some of the work we've been
- 36:02doing in the last two three years and I
- 36:05think AI will be almost be applicable to
- 36:07pretty much every field and I always
- 36:10encourage actually universities to start
- 36:12thinking very seriously about
- 36:14multi-disciplinary work where you apply
- 36:16AI to the right questions in a
- 36:19particular specialist field and I think
- 36:21there's many many advances to be made
- 36:23over the next five to 10 years by doing
- 36:25that.
- 36:28So I'll just end then with a little bit
- 36:30of
- 36:31a more general view about not just AI
- 36:34for science, but the path to AGI and how
- 36:36close we are to that and our more
- 36:38general work on the original mission of
- 36:40of AGI.
- 36:41And we've been making a lot of advances
- 36:43in all areas of general understanding of
- 36:46the world. We sometimes call them world
- 36:48models.
- 36:49So, we're particularly proud of our new
- 36:51video model called VO2, which was just
- 36:53released at the end of last year. It's
- 36:55it's state of the art video generation
- 36:58and it's able to generate these videos
- 37:01just from a text description, right? Or
- 37:04a single static image. So, you know, and
- 37:07actually although some of these videos
- 37:08may not seem that impressive, if you
- 37:10think about this chopping the tomato
- 37:13one, this is like the Turing test for
- 37:14video models because usually you get,
- 37:17you know, the tomato comes magically
- 37:18back together or you're chopping through
- 37:20the fingers or the knife moves off
- 37:21somewhere and it's actually if you think
- 37:23about what the systems had to do to
- 37:25really understand the physics of the
- 37:27world or this or the bubbles around this
- 37:29blueberry here, right? It's just
- 37:30generating that from text, you know,
- 37:32blueberries dropping into a glass of
- 37:34water and it's it's doing all the
- 37:36physics of this correctly or the motion
- 37:38of
- 37:39you know, these these little cartoon
- 37:41characters or the bee here. It's kind of
- 37:44mind-blowing really and I think even if
- 37:45you told me 5 years ago
- 37:48that this would be possible without sort
- 37:50of building in some special
- 37:51understanding of physics or something, I
- 37:53would have told you that that's seems
- 37:55unlikely that that would be possible.
- 37:57Then yet somehow these learning systems
- 37:59are able to learn about real-world
- 38:01physics just from watching, you know,
- 38:03many many YouTube videos. And
- 38:06it's it's pretty crazy that that's
- 38:07possible. So,
- 38:09we've done that. We've gone a step
- 38:11further with Genie 2, which of course
- 38:12bringing my games
- 38:14hat back in here and this is taking
- 38:16those video models a step further. So,
- 38:18now with a text instruction, you can
- 38:20generate a whole game. So, you know,
- 38:23here we at the bottom here we said
- 38:24generate a playable world as a robot in
- 38:26a futuristic city, and it just comes up
- 38:28with this, and you can control it with
- 38:30the, you know, QWE keys and the arrow
- 38:32keys. Um at the moment it's only
- 38:34consistent for a few seconds, um but
- 38:36we're we're we're working to extend that
- 38:38to so that the consistency of the game
- 38:41world lasts for many minutes.
- 38:43Um and so then you've really got what I
- 38:44would call a world model, a really an
- 38:47understanding of the real world. Um and
- 38:49and how interactions in that real world
- 38:51work and the physics of the real world
- 38:53work.
- 38:55Um of course we've been working very
- 38:56hard on the safety aspects of this. Uh
- 39:00and from about very from the very
- 39:01beginning in 2010, we were working on um
- 39:05uh planning for success even though
- 39:06almost nobody was working on AI back
- 39:08then. Um we we imagined that it would be
- 39:10a 20-year mission, and actually
- 39:12amazingly we're sort of on track 15
- 39:14years in. And um we we were sort of
- 39:17planning for success. If we were to
- 39:19build these kinds of transformative
- 39:20systems and technologies, um it would
- 39:23come with a lot of uh uh responsibility
- 39:25as well to make sure they get deployed
- 39:27in a safe uh and responsible way. And
- 39:30one of the uh systems we technology we
- 39:32built is called SynthID, which um
- 39:35invisibly watermarks uh actually using
- 39:37an AI system and adversarial AI system
- 39:40uh slightly adjust the pixels or the
- 39:42text or the or the audio um
- 39:45imperceptibly to the human ear or eye,
- 39:47um but uh it can be detected by a
- 39:49detection system that these were
- 39:51synthetically generated images, whether
- 39:53that's audio, um image, or video. And of
- 39:57course it's going to become increasingly
- 39:58important um as uh as as these
- 40:01technologies become widely deployed that
- 40:03we're able to easily distinguish between
- 40:05synthetically generated images and real
- 40:07images.
- 40:09So, AI's been, you know, this incredible
- 40:11potential uh to to help with our
- 40:14greatest challenges, from climate to to
- 40:17health. but obviously this is going to
- 40:19affect everyone.
- 40:20Um so I think it's really important that
- 40:22we engage not the just it's not just the
- 40:24technologists deciding this, but that we
- 40:26engage with a wide range of stakeholders
- 40:28from society. So I've been really
- 40:30pleased in the last couple of years that
- 40:32one of the consequences of AI becoming
- 40:34mainstream is that many governments have
- 40:36gotten interested in it and many parts
- 40:38of all parts of society. And I think
- 40:39it's been great to see these
- 40:40international summits. Actually the UK
- 40:42hosted the first one in Bletchley Park
- 40:44couple of years ago bringing together
- 40:46heads of government with academia and
- 40:48civil society to discuss uh these
- 40:51technologies, um how to put guard the
- 40:53right guardrails on it, how to make sure
- 40:55we embrace the opportunities, um but we
- 40:58mitigate the risks um that are coming
- 41:00down the line. And uh I think that's
- 41:02going to become increasingly important
- 41:04um given the exponential improvement uh
- 41:06that we're seeing with these
- 41:07technologies. So my my my short answer
- 41:10for this is to say, you know, with with
- 41:12a lot of Silicon Valley tries to the
- 41:14kind of mantra in Silicon Valley is is
- 41:16like move fast and break things. And of
- 41:18course that's that's created a lot of
- 41:19advances, a lot of of the technologies
- 41:22we all use every day today, but I think
- 41:24it's not appropriate in my opinion for
- 41:26this type of transformative technology.
- 41:28I think instead, you know, we should be
- 41:30trying to use uh uh uh the scientific
- 41:33method and approach it with a kind of
- 41:35humility and respect that this kind of
- 41:37technology deserves. And you know, not
- 41:39we we don't know a lot of things. There
- 41:40are a lot of unknowns around how this
- 41:42technology is going to develop. It's so
- 41:44new. Um um and I think with exceptional
- 41:47sort of care and foresight we can get
- 41:49all the benefits
- 41:50uh and minimize uh the downsides of
- 41:52this, but I think only if we um
- 41:55start the research and the debate about
- 41:57that now.
- 42:00So just to end then, we're now building
- 42:02our own big multimodal models that try
- 42:04and take the best of all of these
- 42:05different models I've shown you and put
- 42:07it into one system. Uh we call it the
- 42:09Gemini series. Our latest one's Gemini
- 42:112.0 that some of you may have tried,
- 42:13which is state-of-the-art across many
- 42:15leading benchmarks. Um we're using it to
- 42:18further that I'm very excited about the
- 42:20next generation of assistance. I call it
- 42:22universal assistance. We call it Project
- 42:24Astra, where actually you have it on
- 42:26your phone or some other devices, maybe
- 42:28glasses, and it starts is being an
- 42:30assistant that you can take around with
- 42:31you in the in the real world, and it
- 42:33helps you in everyday life to um to to
- 42:36get to enrich your life or to make you
- 42:38more productive.
- 42:41And the next step then in AI is
- 42:42combining what I've shown you with
- 42:44AlphaGo, these kinds of um agent-based
- 42:47models that are able to um effec-
- 42:50efficiently search through and find a
- 42:52good solution uh to a problem in a in a
- 42:55limited domain, in this case in games.
- 42:57But we want to actually build those
- 42:59types of search systems and planning
- 43:01systems on top of much more general
- 43:03models like Gemini, these world models
- 43:05that understand uh how the real world
- 43:07works, and then can then plan and and
- 43:10achieve things in the real world. And of
- 43:12course that's key to things like
- 43:14robotics uh working, which I think in
- 43:15the next 2 3 years is going to be a huge
- 43:18area um that's going to going to have a
- 43:21a massive advances in it.
- 43:24So I'll just finish then by just having
- 43:26a a slight conjecture about what does
- 43:29this all mean if we think back to Turing
- 43:31and all of the work that he did to lay
- 43:33down the foundations uh of computer
- 43:36science. And um
- 43:38and I think that uh if you see the work
- 43:40that we've done, I see myself as a kind
- 43:42of Turing's champion in a way. Like he
- 43:44how far can the Turing machines and this
- 43:46idea of classical computing go? And I
- 43:49feel like um probably one of the
- 43:50lectures that I took in this room was
- 43:52one of my favo- favorite things to think
- 43:54about is the P equals NP problem, which
- 43:57is uh a famous problem in computer
- 43:59science of, you know, what sorts of
- 44:00problems are tractable on classical
- 44:02systems. And um and there's obviously a
- 44:05lot of great work going on in quantum
- 44:06computing systems, uh many of that here
- 44:09in Cambridge and also Google we have one
- 44:11of the top computer quantum computing
- 44:13groups in the world. And there's a lot
- 44:15of things that are thought to require
- 44:17quantum computing to solve. A lot of
- 44:19real world systems that we would like to
- 44:22understand and model. And my conjecture
- 44:24is that actually classical Turing
- 44:26machines, basically classical machines
- 44:28that these types of AI systems are built
- 44:30on, can do a lot more than we previously
- 44:33gave them credit for. And if you think
- 44:34about AlphaFold and protein folding, you
- 44:37know, proteins are quantum systems, you
- 44:39know, they operate at the atomic scale
- 44:41and one might think you need quantum
- 44:43simulations to actually be able to find
- 44:46the structures of proteins. And yet we
- 44:48were able to approximate those solutions
- 44:51with our neural networks.
- 44:53And so I think, you know, one one
- 44:55potential idea here is that any pattern
- 44:58that can be generated or found in
- 45:00nature, so I it has some real structure,
- 45:02physical structure, can be efficiently
- 45:05discovered and modeled by one of these
- 45:07classical learning algorithms like
- 45:09AlphaFold.
- 45:10And if that turns out to be true, I
- 45:11think that has all sorts of implications
- 45:14for quantum mechanics and and actually
- 45:16fundamental physics,
- 45:17which, you know, is something that I
- 45:19hope to explore and many of my
- 45:21colleagues hope to explore maybe with
- 45:22the help of these classical systems as
- 45:24well to help us uncover what the true
- 45:26nature of reality might be.
- 45:29And that leads me back to the whole
- 45:31reason why I started my path on AI many
- 45:33many years ago is that I always believed
- 45:36that AGI built in this way could be the
- 45:39ultimate general purpose tool to
- 45:40understand the universe around us and
- 45:42our place in it. Thank you.
- 45:53>> [applause]
- 45:59>> Great, we have a time for some questions
- 46:01if people have questions. A first hand
- 46:03shot up just here.
- 46:04Hi. Thank you.
- 46:07Thank you. Thank you for the great talk.
- 46:09Because you have a background in
- 46:11neuroscience,
- 46:12and you really like to think in terms of
- 46:14root note problems, was there ever a
- 46:16root note problem you came across in
- 46:18neuroscience that you thought was worth
- 46:20tackling and still worth tackling to
- 46:23understand biological and artificial
- 46:25intelligence better?
- 46:27>> Yeah, there's many. I mean, I that's
- 46:29what I studied for my PhD actually was
- 46:32memory, but also imagination. So, future
- 46:35thinking, kind of planning. So, I really
- 46:37wanted to understand
- 46:39how the brain did that. And it turns out
- 46:41the hippocampus is involved in both, so
- 46:44that we could maybe mimic that with with
- 46:46some of these algorithms. So, I think
- 46:48there are many key things in that. Of
- 46:50course,
- 46:51there are all the big questions around,
- 46:53you know, creativity, dreaming,
- 46:55consciousness, all these big questions
- 46:57that I think building AI and then
- 47:01comparing it to the human mind is one of
- 47:03the best ways I think we'll make
- 47:04progress with those sorts of root note
- 47:07problems, like, you know, what is the
- 47:08nature of consciousness and is there
- 47:09something special about the
- 47:11instantiation of the substrate of the
- 47:13brain versus
- 47:15algorithmically mimicking that in
- 47:17silicon.
- 47:20Great.
- 47:21Got a question just here.
- 47:23Hi.
- 47:25I have two questions actually. So, since
- 47:27DeepMind was founded before the deep
- 47:30learning revolution, I wanted to know
- 47:32what your state of mind was had deep
- 47:35learning not picked up, or how were you
- 47:38going to progress?
- 47:40That's the first question. The second
- 47:42question is,
- 47:43since you've had intimate experiences
- 47:46with such challenging problems, such
- 47:48high-dimensional problems,
- 47:50and we know that gradient descent and
- 47:53its variants can't
- 47:55uh, of converge to optimum solution,
- 47:58only locally optimum solutions. Were you
- 48:01surprised that anything works at all in
- 48:03these systems at every
- 48:06at every every point in time? And
- 48:08do you think that most of the nature is
- 48:11kind of suboptimal? And so we can
- 48:14potentially build a more optimal nature.
- 48:16So, look, I I think they're both great
- 48:18questions. So, the first the first one,
- 48:21look, that's why we called it DeepMind,
- 48:23partly because the deep refers to deep
- 48:25learning, right? So, we we deep learning
- 48:28had Well, the early parts of the It
- 48:30wasn't called deep learning then, but it
- 48:31was it was sort of just becoming
- 48:33common. There was these Boltzmann
- 48:34machines and things that Geoffrey Hinton
- 48:36had invented and
- 48:38in just a few years before in 2006,
- 48:402005, these hierarchical neural
- 48:42networks. And it seemed like a super
- 48:45promising idea even back then to us that
- 48:47come across it in academia. And
- 48:51then the other thing we bet on was
- 48:52reinforcement learning and the
- 48:54combination of that, right? Which again
- 48:56is coming back into vogue, but it was
- 48:58also important for us to solve something
- 49:00like AlphaGo. So, you need both parts of
- 49:02that. You need to the deep learning to
- 49:03kind of model the environment and the
- 49:05world, and then you need the
- 49:06reinforcement learning to make the plans
- 49:08and and and the solutions and take
- 49:10action in the world. And
- 49:12so there was two reasons we bet on that
- 49:14even when it it was just sort of the
- 49:15beginning of it is that we we knew that
- 49:18the classical methods, these expert
- 49:20systems, would not scale. And actually
- 49:22that's again one of the things I learned
- 49:24here and also at my post doc at MIT was
- 49:26that was that they were the kind of
- 49:28churches, if you like, of the classical
- 49:30methods, these expert systems. And
- 49:32actually that's something else you can
- 49:33learn here is like, you know, in your
- 49:35university courses is not just what to
- 49:37do, but also what not to do. And why it
- 49:39may not like like I sort of thought
- 49:41about it and I felt it would never scale
- 49:43to
- 49:44the kinds of problems that I wanted to
- 49:46solve with AI. Whereas the learning
- 49:48system seemed to have unlimited
- 49:50potential, although they were a lot
- 49:51harder to get anything to do
- 49:53significant to do at the beginning.
- 49:55Right, and this was the problem cuz they
- 49:56weren't scaled up enough. And the other
- 49:58the the other reason we started DeepMind
- 49:592010 is we could also see the computing
- 50:02paradigm was shifting on the hardware
- 50:04side, GPUs and other things
- 50:07which were of course also invented for
- 50:08gaming. And turns out everything's a
- 50:10matrix multiplication, right?
- 50:12Intelligence and gaming and computer
- 50:14graphics. So,
- 50:16and so all of those different influences
- 50:18came together and the new the you know,
- 50:20the understanding of neuroscience and
- 50:22fMRI machines and neuroscience had
- 50:24advanced a lot in that previous 10
- 50:26years, too. So, I felt it was the
- 50:27perfect time to sort of bring all of
- 50:29that together
- 50:31back in 2010. And we were betting on
- 50:33that, not necessarily because we knew it
- 50:34would work, but we knew that the other
- 50:36methods we were pretty confident would
- 50:38not work.
- 50:39And that's what the AI winters were
- 50:40about, basically with people trying to
- 50:42push those expert systems. And then the
- 50:44second question, you know,
- 50:46um
- 50:47I think I wouldn't say that well, first
- 50:50of all, it is surprising that some of
- 50:52these things converge and actually we
- 50:54weren't sure. So, you know, that Atari
- 50:55stuff I showed you. So, for the first
- 50:57couple of years, nothing worked. Right?
- 50:59So, we couldn't even get one point on if
- 51:01some of you remember Pong, the the you
- 51:03know, one of the first computer games,
- 51:04we couldn't this sort of tennis bat and
- 51:06ball game, simplest game you could
- 51:07imagine, we weren't able to get a point
- 51:09on it. So, we were wondering whether
- 51:11were we like 10, 20 years too early?
- 51:14Like like Babbage turned out to be with
- 51:15his difference engine, right? Amazing
- 51:17ideas, it worked, but he was just in the
- 51:18end 50 years or 100 years too early. And
- 51:22and I always say you want to be like 5
- 51:23years ahead of your time, not 50 years
- 51:25ahead. Otherwise, you know, you're
- 51:27you're you're in for a lot of pain like
- 51:29like Babbage was. And
- 51:31and and so we were worried about that,
- 51:33but then it did converge and then that
- 51:35gave us confidence to to tackle the
- 51:36harder problems. And I think if you ask
- 51:38him, I think the last part of your
- 51:39question was about the things in nature.
- 51:41Well, the things there, what I'm
- 51:42thinking is they're not suboptimal,
- 51:44they're actually probably pretty optimal
- 51:45because they've they've gone through an
- 51:47evolutionary process, not just life with
- 51:49biology, but actually geologically and
- 51:51physically. You know, asteroids and
- 51:54physical phenomena combined together,
- 51:56they survive some amount of time because
- 51:58they're stable over time, right? And if
- 52:00they're stable over time, then there's
- 52:02probably some structure that's
- 52:03learnable. That That would be my
- 52:05conjecture.
- 52:07Great. There's a question down here.
- 52:12What do you think about building
- 52:15high-bandwidth brain-machine interfaces
- 52:17and implantable memory and reasoning
- 52:19modules so that humans can be further
- 52:21empowered to make discoveries
- 52:22autonomously as opposed to only talking
- 52:24to AI in the cloud? Yeah, I I I I'm I I
- 52:27love that area and I've
- 52:30carefully followed it and and help
- 52:31people building like EEG caps and
- 52:34and things. Of course, the problem is
- 52:36the resolution of these things to get
- 52:38the readouts from the brain. And also,
- 52:40ideally, you'd want it read and write.
- 52:42But I'm very fascinated by projects like
- 52:44Neuralink or, you know, chips in the
- 52:46brain. Obviously, right now,
- 52:48that's for,
- 52:50you know, veterans and people to get
- 52:51function back in their body. I think
- 52:52it's going to be amazing things on that
- 52:54where I think people will be able to
- 52:55gain
- 52:56be able to walk again if they've broken
- 52:58their back and things like that. I think
- 53:00there's going to be some incredible
- 53:01advances in medical sciences that will
- 53:02be amazing. But then beyond that, maybe
- 53:06if things, you know, if it becomes
- 53:07routine and it's surgery safe and
- 53:09there's safe ways of doing this, I could
- 53:11imagine
- 53:13that would be one way for us to keep up
- 53:14with the with the technology. And in
- 53:16some senses, it's no different to what
- 53:18we already have today with our
- 53:19technologies all around us. We all have
- 53:21our phones all, you know, 24/7 with us
- 53:23and computers and other things. So,
- 53:25we're already symbiotic almost with our
- 53:27technology. It would be one step
- 53:29further, of course, to have that. But
- 53:30I'm not sure I mean, maybe that's for
- 53:31the for the philosophers in the room to
- 53:33answer what the difference what if
- 53:35whether there's a hard boundary there if
- 53:37if with the technology if it's attached
- 53:39to you versus, you know, it's just
- 53:41something you carry around with you, um,
- 53:43all the time.
- 53:45Great. Some question there for you.
- 53:47That's one just
- 53:49Sure, then.
- 53:51Hi.
- 53:55Are you sure? Oh.
- 53:57Hello. Ah, there we go. There we go.
- 54:00Ah.
- 54:02What do you think about the speed at
- 54:03which artificial intelligence is
- 54:05developing and its effects on sort of
- 54:07economic developments? There are a lot
- 54:09of people out there who are deciding
- 54:10careers right now who given the
- 54:13sort of rapid change in the landscape,
- 54:16make it really difficult to kind of
- 54:17predict what they should go into.
- 54:19Yeah, so it's it's it's, um, it's very
- 54:21complicated one because, as you say,
- 54:23things are changing at kind of lightning
- 54:25speed.
- 54:26Um, we were just discussing it with
- 54:28Alice actually earlier. Even designing
- 54:29three-year computer science courses is
- 54:31quite difficult given that this the
- 54:33underlying material changes in in we we
- 54:35we, you know, less than three years. Um,
- 54:38I think the only thing we can say for
- 54:39sure is there's going to be a lot of
- 54:40change, but I think that brings with it
- 54:43disruption and opportunity. So, um, I
- 54:46mean, I just give you an example on
- 54:47coding, which I don't know if you're a
- 54:48computer scientist, but you know, so I
- 54:51still would recommend that you get good
- 54:52at coding, uh, and math because I think
- 54:55you'll be able to use these new tools in
- 54:57a much deeper way if you have an
- 54:58understanding of how they're built. But
- 55:00on the other hand, um,
- 55:02I think coding is going to be more
- 55:04available to many, many more types of
- 55:06people,
- 55:07uh, because of the way that AI you'll be
- 55:09able to program sort of a natural
- 55:10language, probably, rather than, uh,
- 55:13quite a complicated, uh, computer
- 55:15language. And so, that will open up
- 55:18field fields for creative people to, you
- 55:21know, build games, make films, make
- 55:23applications that, um,
- 55:26maybe is more the kind of balance of
- 55:28that is more on the creative side than
- 55:30it is on the engineering side. But I
- 55:32also think that will enhance engineers
- 55:34to be able to do 10x of what they can do
- 55:35today. So, I still I think it's
- 55:37difficult to know, um but what I would
- 55:39say is just uh to um focus on uh kind of
- 55:44embracing those tools in your spare time
- 55:47and and
- 55:49making training yourself to be really
- 55:50good at picking up new information
- 55:52extremely quickly,
- 55:52>> [clears throat]
- 55:53>> cuz I think that's that's basically
- 55:54what's going to happen in the next 10
- 55:56years.
- 55:57Okay. Um we've got one question just in
- 55:59the here, the yellow and black top.
- 56:02Um is this Yeah, do you think that there
- 56:05are any biological processes or
- 56:07behaviors or patterns which can't be
- 56:10modeled with existing um deep learning
- 56:13techniques? I'm not saying like throw
- 56:16models computer to this until it works
- 56:17and just make a bigger and bigger model.
- 56:19Do you think there are some processes
- 56:20which physically cannot be modeled with
- 56:21the architecture we have?
- 56:23Um I mean, there's certainly lots of
- 56:25processes that can't be modeled uh today
- 56:27with but but but again, this is sort of
- 56:29goes back to what I said at the end of
- 56:30the talk. I'm not sure in the limit uh
- 56:33that there are. I think in the end uh if
- 56:35physics can can sort of solve it and
- 56:38there's some structure there to be
- 56:39learned,
- 56:40uh probably with enough examples, one
- 56:42could reverse engineer a model of that.
- 56:45Uh and then um I don't see any
- 56:47theoretical reason why
- 56:49uh a a classical system um or be a very
- 56:52complex one could not uh make
- 56:54predictions or or simulations of that
- 56:57biological system. So, I I don't really
- 56:59see
- 57:00uh any in the limit uh what what the
- 57:02limit of that could be. I mean, there
- 57:04are lots of abstract things like
- 57:06factorizing large numbers, uh
- 57:08cryptography where it's sort of it's
- 57:10sort of human-made systems, right? Where
- 57:12there may not be any structure. I mean,
- 57:13there may be structure in them in the in
- 57:15the natural numbers. Lots of people
- 57:17conjecture there is. If there is, then
- 57:19that will also be learnable. If there
- 57:20isn't and it's kind of uniform
- 57:22distribution, then you would need uh
- 57:24quantum computer to to you know, crack
- 57:27uh cryptography, things like that. So,
- 57:29those are open conjectures, but I think
- 57:31most things in nature have evolved and
- 57:34um over geological or biological physics
- 57:37time. And so then that suggests to me
- 57:40there is some structure to learn. So
- 57:42that makes the the search or the
- 57:44prediction potentially tractable.
- 57:47Okay, and uh
- 57:48last question then
- 57:50we'll have
- 57:51person in the pink shirt.
- 57:54This question comes from the behalf of
- 57:56the Cambridge University Game
- 57:57Development Society.
- 57:59So you mentioned
- 58:00you mentioned about the Genie 2 model
- 58:02and how currently it's stable for a few
- 58:05seconds consistency and you hope to
- 58:08eventually have that at at a few
- 58:09minutes. But I suppose the question that
- 58:11our society has is
- 58:13games that we actually play have
- 58:15consistency that's indefinite. You know,
- 58:17when you're playing Minecraft you expect
- 58:18to turn around and the village is still
- 58:20there, right?
- 58:20>> Yes. So do you see your current model
- 58:23being integrated into a workflow or what
- 58:26exactly how do you see AI and your model
- 58:29and what you're working on?
- 58:30How do you see it integrating into game
- 58:32development in the coming decades?
- 58:35Yeah, well look
- 58:37I think there's many ways AI is going to
- 58:39come in. So
- 58:40one is the tools to build the assets you
- 58:43need for games. So 3D models,
- 58:46animations. I think that's all going to
- 58:48come in in the next you know couple of
- 58:50years.
- 58:52I think you can think of AI as well for
- 58:54game balancing. So you imagine you
- 58:56design a game and overnight it could
- 58:59play a million playthroughs of that game
- 59:01and then in the morning you could just
- 59:02get a report as a game designer of like
- 59:04these things are unbalanced, right? Or
- 59:06reduce the power of this unit or
- 59:08whatever it is.
- 59:09I think also bug testing for open world
- 59:11games. So I used to make I used to make
- 59:14simulation games, open world games.
- 59:16They're a nightmare to bug test because
- 59:18the whole point of them is that the
- 59:19player can almost do almost anything and
- 59:21then the game will react to you. So how
- 59:22do you test you know 10 million people
- 59:25having their own unique journey through
- 59:26your game, well actually having AI
- 59:29players play it before you release it
- 59:32could help you solve a lot of those
- 59:33bugs. And then finally, I mean
- 59:36I think excitingly as well is like AI
- 59:38characters that are much more lifelike
- 59:39that move the storyline on. Used to
- 59:41dream about that massive multiplayer
- 59:43worlds where, you know, actually the AI
- 59:45characters were intelligent and actually
- 59:47updated their beliefs and
- 59:49their storylines based on what the
- 59:50players were doing. So it felt like a
- 59:52much more
- 59:53living realistic world. Um and I think
- 59:56we're on the cusp of having or building
- 59:58those types of of of games. And then
- 1:00:01finally, the world model we're building
- 1:00:02that's more about general AI and and can
- 1:00:05you actually model the world It's an
- 1:00:07expression of being able to understand
- 1:00:08the world. Does your model understand
- 1:00:10the world? Well, if it can if it can
- 1:00:11generate it for some amount of time,
- 1:00:14then obviously it must be in some sense
- 1:00:16understanding It's a kind of empirical
- 1:00:18evidence that it's understanding
- 1:00:19something about the underlying physics.
- 1:00:21So um so that's more for the general
- 1:00:23intelligence rather than I think we're
- 1:00:25going to maybe one day we'll have this
- 1:00:26holodeck thing that you can just imagine
- 1:00:29and it's just all there around you. Um
- 1:00:31probably we will be able to have that
- 1:00:33once we have AGI, but um I think that's
- 1:00:35still a ways off.
- 1:00:37That's great. Uh well, it seems like a
- 1:00:39nice place to finish. A question on
- 1:00:40games returning back to games, but uh
- 1:00:42thank you all so much for coming and a
- 1:00:44particular special thank you for for
- 1:00:46Demis for coming in and talk to us
- 1:00:47today. So thank you very
- 1:00:58>> [applause]
About this transcript
This page contains the full transcript of Demis Hassabis Reveals the Future of AI - This Cambridge Lecture Is Mind-Blowing @AICreatorForge by AI Creator Forge, generated from the public captions YouTube serves with the video. The transcript has 11,256 words across 1,727 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.