YouTube2Text

Demis Hassabis Reveals the Future of AI - This Cambridge Lecture Is Mind-Blowing @AICreatorForge — Transcript

by AI Creator Forge · 11,256 words · 1,727 segments · language en · Watch on YouTube

Full transcript

  1. 0:00[applause]
  2. 0:03>> Thanks Alister for that lovely
  3. 0:05introduction and it's so great to be
  4. 0:07back at Cambridge. I always have a warm
  5. 0:09feeling when I sort of my homecoming
  6. 0:12back to Cambridge and specifically this
  7. 0:14lecture hall as Alister reminded me I
  8. 0:15think it is the first lecture hall I was
  9. 0:17in it's always been my favorite lecture
  10. 0:18hall.
  11. 0:19I remember telling and I see a lot of my
  12. 0:21old friends here from my Cambridge days
  13. 0:24Aaron I think about that one day maybe
  14. 0:27I'd come back to give a lecture in here
  15. 0:30and talk about announcing AGI and maybe
  16. 0:32a robot would walk on and
  17. 0:34astound everyone. I'm not going to do
  18. 0:36that today
  19. 0:37to to disappoint you but maybe in a few
  20. 0:40years time I'll come back again and I'll
  21. 0:41give that lecture.
  22. 0:43But it you know this is amazing place
  23. 0:45it's such an inspiring place and I'm
  24. 0:46going to talk a little bit about how
  25. 0:48Cambridge has inspired my whole career
  26. 0:50actually and hopefully it's going to do
  27. 0:51the same for many of you the students in
  28. 0:54in the room.
  29. 0:56So for me it started
  30. 0:58get my journey on AI started with games
  31. 1:01and specifically chess as Alister
  32. 1:03mentioned. So I was playing chess from
  33. 1:04the age of four years old and very
  34. 1:07seriously for the England junior teams
  35. 1:08and things like that and it got me
  36. 1:10thinking about thinking itself. You know
  37. 1:12how does our mind come up with these
  38. 1:14plans with these ideas how do we problem
  39. 1:17solve and how can we improve? Obviously
  40. 1:19when you're playing chess at a young age
  41. 1:21you're trying to play competitively
  42. 1:22you're trying to improve that process.
  43. 1:24And it was fascinating to me perhaps
  44. 1:26more fascinating than even the games I
  45. 1:28was playing was the actual mental
  46. 1:30processes behind it.
  47. 1:32And in fact AI
  48. 1:34and can computers I came across
  49. 1:36computers and AI for the first time in
  50. 1:38the context of chess and trying to use
  51. 1:40very early chess computers like the one
  52. 1:42on the right here I think this was my
  53. 1:44first ever chess computer there were
  54. 1:46physical boards where you had to
  55. 1:47actually had to press the squares down
  56. 1:49to make move the pieces
  57. 1:51and of course we were supposed to be
  58. 1:52using these chess computers to train
  59. 1:54opening theory and and learn more about
  60. 1:56chess but I remember being fascinated by
  61. 1:59the fact that someone had programmed
  62. 2:01this lump of inanimate plastic to
  63. 2:03actually play chess really well against
  64. 2:06you. And I was sort of really fascinated
  65. 2:08by how that was done and how how someone
  66. 2:11could program something like that. And I
  67. 2:13ended up experimenting myself in my
  68. 2:15early teenage years with the Amiga 500
  69. 2:18computer, amazing home computer back in
  70. 2:20the in the late '80s and early '90s, and
  71. 2:23building those kinds of AI programs
  72. 2:25myself to play games like Othello. And
  73. 2:27really that was the first my first taste
  74. 2:29of AI and and I was hooked from then on
  75. 2:31and that's, you know, I decided from
  76. 2:32very early on that I would spend my my
  77. 2:35entire career trying to push the
  78. 2:37frontiers of of this technology.
  79. 2:41So then that led me to
  80. 2:43to Cambridge
  81. 2:44and which was really my three years here
  82. 2:46were incredibly formative for me.
  83. 2:49And I went to a school in comprehensive
  84. 2:52school in North London. No one had ever
  85. 2:54got to sort of no one had gone to
  86. 2:56Oxbridge for in sort of living memory,
  87. 2:58but and the reason I wanted to come to
  88. 3:00Cambridge was all these inspiring
  89. 3:02stories that I'd heard about
  90. 3:04what happened at Cambridge, all these
  91. 3:06amazing people that I used to read about
  92. 3:07their biographies and the work they'd
  93. 3:09done. Especially people like Crick and
  94. 3:12Watson in the in the top left there. And
  95. 3:14I remember a particularly a film The
  96. 3:16Race for the Double Helix, an amazing
  97. 3:18film from the '80s if you haven't seen
  98. 3:20it with Jeff Goldblum, one of his early
  99. 3:22parts so he you know with all the
  100. 3:23enthusiasm that he plays all of his
  101. 3:25parts as he was Watson. And they were
  102. 3:28having just such an amazing time
  103. 3:30discovering, roaming around Cambridge,
  104. 3:33you know, working on things like DNA and
  105. 3:34I thought like that's what I want a
  106. 3:36piece of that. And I want to feel what
  107. 3:38it's like to be at the frontier of
  108. 3:41discovery and what could be more
  109. 3:42exhilarating. And and that film actually
  110. 3:44really brought it to life what that
  111. 3:47might be like.
  112. 3:49And then of course all of my my my
  113. 3:51heroes, my scientific heroes, a lot of
  114. 3:53them had gone through Cambridge. People
  115. 3:55like Alan Turing and Charles Babbage, of
  116. 3:57course, and in the lecture hall that we
  117. 3:59we now sit in.
  118. 4:01Um and and and even places like the
  119. 4:03Eagle pub where if you start at Queen's
  120. 4:05College, one of the tours they give you
  121. 4:07on the first day is to go and show you
  122. 4:09the the the reputed sort of
  123. 4:11table that they were discussing and and
  124. 4:13and at the DNA structure around. And you
  125. 4:17can't help but be inspired by that and
  126. 4:19walking down King's Parade, um and I
  127. 4:22almost felt like the the the
  128. 4:24intellectual giants of the past were
  129. 4:26were almost speaking to you from the
  130. 4:27stones.
  131. 4:28Um and and that's how I felt, you know,
  132. 4:30going for a late night burger at
  133. 4:32Gardenia's.
  134. 4:33Um but that was what was inspiring me
  135. 4:35around, like all of these amazing people
  136. 4:37that had had walked those same steps
  137. 4:39over hundreds of years. And and that's
  138. 4:41the history that is sort of unrivaled
  139. 4:43here at Cambridge that I think we can
  140. 4:45still draw on and take inspiration from
  141. 4:47today. Uh and then there's a picture of
  142. 4:50me and Aaron there, one of my best
  143. 4:51friends from Queen's, you know,
  144. 4:53obviously on the mathematical bridge
  145. 4:55there.
  146. 4:56And then finally, you know, what when we
  147. 4:58Alister mentioned obviously the Nobel
  148. 5:00Prize and it was a honor of a lifetime
  149. 5:02to go and collect that in in Stockholm
  150. 5:05um in December. Um amazing one week of
  151. 5:08of activities, but my favorite activity
  152. 5:10was when you get to sign the Nobel book
  153. 5:12in the Nobel Foundation. And and and
  154. 5:15that's the book there and one of my
  155. 5:17pictures of and I started sort of
  156. 5:18leafing through the book. You saw your
  157. 5:20name and then you leaf back and you
  158. 5:21wonder is you know, is Crick in there?
  159. 5:23And of course he is. And then you go
  160. 5:25back further and then Einstein's
  161. 5:26signature's there and it's just
  162. 5:27mind-blowing, really. And I started to
  163. 5:29spend an hour just photoing every page
  164. 5:31of the of the book.
  165. 5:34Um
  166. 5:34so, you know, it's full circle for me of
  167. 5:36of of that picture and and then really
  168. 5:39seeing that film in the in the in the in
  169. 5:41the late '80s.
  170. 5:44So, then
  171. 5:46you know, in 2010 we started DeepMind in
  172. 5:48London as really at the time it was a
  173. 5:51kind of an Apollo program effort is the
  174. 5:53way we thought of it for trying to build
  175. 5:55artificial general intelligence. You
  176. 5:57know, AI that was truly general and
  177. 6:00could perform all the cognitive
  178. 6:02capabilities that that humans are
  179. 6:04capable of. So, it would be a truly
  180. 6:06general AI system. In fact, you know,
  181. 6:09the idea for that really comes from
  182. 6:10Turing and Turing machines. So,
  183. 6:12something that's able to compute
  184. 6:15anything that is computable as Turing
  185. 6:17showed with his Turing machines. And
  186. 6:19really that's been the foundation for me
  187. 6:22of and one of the main things that I
  188. 6:24carried with me from the lectures here
  189. 6:26at Cambridge was all these theoretical
  190. 6:28underpinnings of computer science and
  191. 6:30computation theory that that people like
  192. 6:32Turing and Shannon famously did in the
  193. 6:3540s and the 50s.
  194. 6:38So, we started in 2010 and it's amazing
  195. 6:40like it's 15 years ago, which in some
  196. 6:42ways isn't that long ago, but when we
  197. 6:44started out DeepMind, almost nobody was
  198. 6:46working on AI, which is hard to believe
  199. 6:48today given that almost everyone seems
  200. 6:50to be working on AI today
  201. 6:52in just a matter of just over a decade,
  202. 6:54things have gone, you know, accelerated
  203. 6:56incredibly. And obviously, we've been
  204. 6:57part of that very exciting journey.
  205. 7:01So, our mission at DeepMind from the
  206. 7:02beginning was
  207. 7:05we talk about building AI responsibly to
  208. 7:06benefit humanity, but the way we used to
  209. 7:08articulate it when we started out was in
  210. 7:11a two-step process. Step one, solve
  211. 7:13intelligence. Step two, use it to solve
  212. 7:15everything else. And and this seemed
  213. 7:18very outlandish at the time in 2010 and
  214. 7:20you can imagine trying to pitch a
  215. 7:22venture capitalist on that on the basis
  216. 7:24of that mission. It seemed pretty crazy,
  217. 7:26but I I still fundamentally believe in
  218. 7:28that today and I think more and more
  219. 7:30people are realizing that AI
  220. 7:32built in this general way could have
  221. 7:34these kind of profound and
  222. 7:36transformative impact on almost any
  223. 7:38field,
  224. 7:39which is obviously the second part of
  225. 7:41that mission statement, and I think
  226. 7:43for me that that involves accelerating
  227. 7:46scientific discovery itself and medicine
  228. 7:49and advancing our understanding of the
  229. 7:51universe around us.
  230. 7:55So back when we started out, there were
  231. 7:56basically two ways, and in fact when I
  232. 7:58was studying here in the '90s, there
  233. 8:00were two ways to build AI, broadly
  234. 8:02speaking. There's the expert system way,
  235. 8:04which is you kind of pre-program an
  236. 8:06expert system directly with the
  237. 8:08solution.
  238. 8:10Uh things like Deep Blue that beat Garry
  239. 8:11Kasparov at chess very famously in the
  240. 8:13'90s, actually while while I was
  241. 8:15studying here.
  242. 8:16That would be the pinnacle example
  243. 8:18probably of an expert system. But the
  244. 8:20problem with these expert systems was,
  245. 8:22and why they never really scaled to full
  246. 8:24general intelligence, is they can't deal
  247. 8:26with the unexpected.
  248. 8:29If you you know, if something unexpected
  249. 8:30happens that you didn't already cater
  250. 8:32for, there's nothing in the system that
  251. 8:34will allow it to to deal with that.
  252. 8:37And they're sort of were inspired by
  253. 8:38logic systems,
  254. 8:40and they were quite rigid and and
  255. 8:41fragile and brittle because of that.
  256. 8:44Whereas the modern-day approaches
  257. 8:46are built on learning systems. So these
  258. 8:47are systems that are able to learn for
  259. 8:49themselves and learn directly from
  260. 8:50experience or data from first
  261. 8:52principles,
  262. 8:53and really inspired by more neuroscience
  263. 8:55ideas.
  264. 8:57And the obviously the promise of these
  265. 8:58systems that we have today is that they
  266. 9:00can go beyond the knowledge potentially
  267. 9:03that we as the programmers or the system
  268. 9:05designers already know how to solve. And
  269. 9:07of course that's extremely valuable in
  270. 9:09areas like
  271. 9:11scientific discovery.
  272. 9:14So we started in the in the early 2010s
  273. 9:16with with games, of course, and I've
  274. 9:18used games many times in my life. First
  275. 9:20of all, to train my own mind, then I
  276. 9:22used to build games and AI for computer
  277. 9:25games, and then finally in a third way
  278. 9:27to train up our AI systems. And games
  279. 9:29are the perfect proving ground for AI
  280. 9:31systems.
  281. 9:32You can start with very simple games
  282. 9:34like Atari games from the '70s,
  283. 9:36and really at this this this system DQN
  284. 9:39was the first time anyone had built an
  285. 9:41end-to-end learning system that could
  286. 9:43learn directly from raw data. So, in
  287. 9:45this case, the raw pixels on the screen,
  288. 9:48and it's it's not told anything about
  289. 9:49the games or anything about what it's
  290. 9:51controlling. It's just told to maximize
  291. 9:53the score based on this
  292. 9:56video stream input, pixel stream input.
  293. 10:00So, we were able to master all different
  294. 10:02Atari games
  295. 10:04in sort of around 2013. Then we took
  296. 10:06these
  297. 10:07systems and we scaled them up to really
  298. 10:09the I would say the grand challenge of
  299. 10:12games AI, which is can you you you
  300. 10:15create systems that can play the game of
  301. 10:17Go
  302. 10:18at world champion level or beyond. And
  303. 10:21Go, of course, is probably the most
  304. 10:23complex game that humans have ever
  305. 10:25invented. It's thousands of years old,
  306. 10:27so it's also the oldest game and one of
  307. 10:29the most elegant games.
  308. 10:31But, one of the ways you can just see
  309. 10:33the complexity of Go is that there are
  310. 10:3410 to the power 170 possible positions
  311. 10:37in Go.
  312. 10:38So, that's more way more than there are
  313. 10:40atoms in the observable universe. And
  314. 10:43the important reason point about that is
  315. 10:45that you cannot come up with a strategy
  316. 10:47in Go using brute force techniques. It
  317. 10:49would be impossible. It'll be totally
  318. 10:51intractable. So, you have to do
  319. 10:52something much smarter. And we famously
  320. 10:55in basically 2016 won a million-dollar
  321. 10:58challenge match against 10-times world
  322. 11:01champion Lee Sedol and one of the
  323. 11:03legends of the game as South Korean
  324. 11:05grandmaster and was watched by 200
  325. 11:07million people around the world. And not
  326. 11:09only did AlphaGo, our system, win that
  327. 11:12match, importantly it actually came up
  328. 11:14with new, original Go strategies, even
  329. 11:17though
  330. 11:18we've played Go for thousands of years
  331. 11:21and professionally for hundreds of
  332. 11:22years, it was still able to find
  333. 11:24never-seen-before strategies. Most
  334. 11:26famously, this move 37 move here in red
  335. 11:31uh game two, which um if you watch the
  336. 11:33documentary on this, you'll see uh which
  337. 11:35is on YouTube, you'll see uh that how
  338. 11:37surprised the the best players in the
  339. 11:39world were. They were commentating on
  340. 11:41the game uh about this move. It was sort
  341. 11:43of an unthinkable move, and yet it
  342. 11:45decided um the game this game two in
  343. 11:48favor of AlphaGo um
  344. 11:50100 moves later.
  345. 11:52So, again, that so told me about the
  346. 11:54potential for these types of systems to
  347. 11:56invent and discover new knowledge. So,
  348. 11:59here, of course, we're just talking
  349. 12:00about game knowledge, but um obviously,
  350. 12:03my dream was to generalize this to all
  351. 12:05areas of um scientific discovery.
  352. 12:10So, how do these systems work? Um we we
  353. 12:13basically train up these neural networks
  354. 12:15through a system of self-play. So, this
  355. 12:17is actually AlphaGo and also um
  356. 12:20subsequent systems uh that AlphaGo Zero
  357. 12:23and AlphaZero that generalized what
  358. 12:25we've done for Go to play any two-player
  359. 12:27uh game from scratch. And you start off
  360. 12:30with a version one of the system that
  361. 12:32doesn't really know anything about the
  362. 12:33game, just the rules, and it plays
  363. 12:35randomly, and you play I'd say 100,000
  364. 12:38games against itself uh uh of this
  365. 12:40system, and that creates a new database
  366. 12:43of um game positions from those 100,000
  367. 12:46uh games. And from that, you train a
  368. 12:49second version a a slightly better
  369. 12:50version of the model, version two, uh
  370. 12:53that's trained to predict um what the
  371. 12:55likely moves are to be played in any one
  372. 12:58position, and also who is more likely to
  373. 13:00win, which side, black or white, is
  374. 13:02likely to win from that position, uh and
  375. 13:05what percentage chance do they have of
  376. 13:06winning.
  377. 13:07Uh and then you can use that version two
  378. 13:10to um you play it against version one uh
  379. 13:13and in a 100-game match-off, and then if
  380. 13:16it wins by significant uh margin, so in
  381. 13:18this case 55% win rate, you replace the
  382. 13:21version one with version two, and you
  383. 13:24create a new uh database of games that
  384. 13:27are slightly higher quality and then you
  385. 13:29learn a version three system. And if you
  386. 13:31do this and you repeat it around 17, 18
  387. 13:34times, you go from playing randomly in
  388. 13:36the morning to 24 hours or less later,
  389. 13:39you're you're you're by the version 17
  390. 13:41or 18, you're stronger than the world
  391. 13:43champion level. So, it's quite an
  392. 13:45incredible thing process to see this
  393. 13:47this self-improvement process playing
  394. 13:50out in a very, very short amount of
  395. 13:52time.
  396. 13:54So, if we think about what these neural
  397. 13:56networks are doing is you're kind of
  398. 13:58reducing down this intractable search
  399. 14:01space of, you know, 10 to the power 170
  400. 14:03possibilities down to something that's
  401. 14:05much more tractable in a few minutes of
  402. 14:08compute time. And it's doing this by
  403. 14:11narrowing down using the neural network
  404. 14:14to efficiently guide the search
  405. 14:16mechanism. So, if you think about this
  406. 14:18tree of possibilities as and each of the
  407. 14:21nodes in this tree is a go position,
  408. 14:24then instead of having to to look at
  409. 14:26sort of every possibility, you can
  410. 14:28actually use the neural network to guide
  411. 14:30you just down the most the most
  412. 14:33most interesting and most useful lines
  413. 14:36to examine. So, in this case, the ones
  414. 14:38in blue. And then
  415. 14:40after you've run out of thinking time,
  416. 14:42you pick the best line, the most
  417. 14:43promising line that you've seen thus
  418. 14:45far. So, I I in this case, this this
  419. 14:48particular line in purple.
  420. 14:52So, this then leads to, you know, we we
  421. 14:53then play not just go, but any
  422. 14:56two-player perfect information game and
  423. 14:58it was even able to discover new
  424. 15:00strategies and new styles of playing
  425. 15:02chess,
  426. 15:03which which is kind of extraordinary but
  427. 15:05given that chess computers were so
  428. 15:07strong already. So, programs like
  429. 15:09Stockfish. And AlphaZero was able to
  430. 15:12beat Stockfish at the time
  431. 15:15in at chess, which is almost sort of
  432. 15:17impossible to do. And not only did it
  433. 15:19beat Stockfish, so AlphaZero here is in
  434. 15:21in white uh playing against Stockfish
  435. 15:23who's black. But, in this particular
  436. 15:25position, one of the most famous uh
  437. 15:27games that Alpha Zero played is called
  438. 15:28the immortal Zugzwang game. Um
  439. 15:31white is winning here um uh because it
  440. 15:34favors mobility over material. So, most
  441. 15:36chess computers favor material, and
  442. 15:38you'll see that black Those of you who
  443. 15:40play chess, you'll see that black has
  444. 15:41more material, um but actually can't
  445. 15:43move any of its pieces. They're all
  446. 15:45stuck in the corner. And this is what
  447. 15:47Alpha Zero sacrificed material for this
  448. 15:49mobility. And actually, for human
  449. 15:52grandmasters uh and top chess players,
  450. 15:54this is not only very effective style,
  451. 15:56it's very beautiful aesthetic style uh
  452. 15:58to and uh to play chess in. So, um it's
  453. 16:01kind of amazing that Alpha Zero was able
  454. 16:03to sort of discover this new way, this
  455. 16:06new dynamic way of playing. And in fact,
  456. 16:08some of the top chess players in the
  457. 16:09world commentated about this. So, Garry
  458. 16:11Kasparov, my all-time sort of favorite
  459. 16:13chess player, he said, "The programs
  460. 16:15usually reflect priorities and
  461. 16:16prejudices of the programmer, but
  462. 16:18because Alpha Zero learns for itself, I
  463. 16:20would say that its style reflects the
  464. 16:22truth." And then, the current world
  465. 16:24champion at the time, Magnus Carlsen,
  466. 16:26said, uh you know, read uh and looked at
  467. 16:28these games and and read about the books
  468. 16:30that were written about Alpha Zero, and
  469. 16:32I said, "I've been influenced by one of
  470. 16:34my heroes recently, one of which is
  471. 16:36Alpha Zero." So, he actually
  472. 16:37incorporated a lot of these ideas uh
  473. 16:40into his own game to dominate the chess
  474. 16:42scene for, you know, almost a decade
  475. 16:44now.
  476. 16:47So, we did all these landmark
  477. 16:48breakthroughs in games AI and over the
  478. 16:50first, I guess, sort of decade of of of
  479. 16:53DeepMind's existence, but of course,
  480. 16:55these were just the training ground for
  481. 16:56what we wanted to do and uh was just a a
  482. 16:59means to an end. It wasn't the end in
  483. 17:01itself to play these games, much as I
  484. 17:03love games. Um it was to create these
  485. 17:05algorithms that could be generally
  486. 17:07useful for tackling real-world problems.
  487. 17:12So, um what we look for in real-world
  488. 17:14problems, not only scientific problems,
  489. 17:16uh but actually uh industrial problems
  490. 17:18as well. And we look for sort of three
  491. 17:21different criteria that make it suitable
  492. 17:23the problem suitable to be tackled by
  493. 17:26these types of AI systems and and ideas
  494. 17:29and algorithms that we developed for
  495. 17:30playing games. Number one is we look for
  496. 17:33problems that can be described as
  497. 17:35massive combinatorial search spaces. So
  498. 17:39So usually far too complex, far too many
  499. 17:41combinations to brute force the
  500. 17:43solution. So but maybe there's some kind
  501. 17:46of structure that we can learn about
  502. 17:48with our neural networks that can guide
  503. 17:50that search very efficiently.
  504. 17:52Secondly, we look for problems that can
  505. 17:54be described with a clear objective
  506. 17:56function or some sort of metric that you
  507. 17:58can optimize against. So in games that's
  508. 18:01very easy. It's things like maximizing
  509. 18:03the score or winning the game. But
  510. 18:06actually there's a lot of real world
  511. 18:07problems that you can boil down to a few
  512. 18:10metrics or few objective functions that
  513. 18:12you're trying to maximize.
  514. 18:14And then finally, of course you need
  515. 18:15quite a lot of data or experience to
  516. 18:18learn from and or ideally it's and a
  517. 18:21kind of accurate and efficient simulator
  518. 18:23so you can generate more synthetic data
  519. 18:26to augment the real data that you have.
  520. 18:30And it turns out that there are a lot of
  521. 18:32problems that can be couched in these
  522. 18:34terms if you're looking at the problem
  523. 18:37from this angle.
  524. 18:38And including many many problems
  525. 18:40important problems in science.
  526. 18:42And the one that I always had in mind
  527. 18:44actually from my days of first seeing
  528. 18:46coming across the problem here at
  529. 18:47Cambridge as an undergrad is the protein
  530. 18:49folding problem which I'll just quickly
  531. 18:51describe to you for those that don't
  532. 18:53know about biology and proteins. So
  533. 18:55proteins are incredibly important.
  534. 18:58They're they're the building blocks of
  535. 18:59life. Pretty much every every
  536. 19:02function in the living body depends on
  537. 19:04proteins from your neurons firing to
  538. 19:07your muscle fibers twitching. So really
  539. 19:10proteins are uh makes life possible.
  540. 19:14And the protein folding problem then is
  541. 19:15is really easy to describe. It's
  542. 19:17basically a protein is is defined by its
  543. 19:20uh it's
  544. 19:22defined by its gene sequence, its
  545. 19:23genetic sequence, uh which then
  546. 19:25specifies an amino acid sequence, uh
  547. 19:28which uh in nature then folds up
  548. 19:30spontaneously into a usually a very
  549. 19:33beautiful uh protein structure. So, you
  550. 19:36go from this genetic sequence to a
  551. 19:37protein structure. And the reason the
  552. 19:39protein structure, the 3D structure is
  553. 19:41very important is it goes a long way to
  554. 19:44defining what function it has, what it
  555. 19:46does uh in the body. So, um it's not
  556. 19:49doesn't totally describe the function,
  557. 19:51but it it has a big part to play in what
  558. 19:53what it actually does uh in nature. So,
  559. 19:56the protein folding problem then is is
  560. 19:57this problem of can you predict the
  561. 19:59protein structure directly from this
  562. 20:02one-dimensional amino acid sequence? Can
  563. 20:05you predict that computationally, that
  564. 20:06incredible 3D structure uh from that
  565. 20:09sequence?
  566. 20:12So, why is this such a hard problem?
  567. 20:14Well, um
  568. 20:15Levinthal is a famous researcher,
  569. 20:17protein researcher in the in the '60s
  570. 20:19described this uh
  571. 20:21conjecture that that that became called
  572. 20:24it became known as the Levinthal's
  573. 20:25paradox, which is that he calculated
  574. 20:27there roughly, you know, 10 to the 300
  575. 20:30possible shapes that an average protein
  576. 20:33can take. Um and yet somehow in nature
  577. 20:36and in the body, these proteins fold up
  578. 20:38spontaneously in a matter of
  579. 20:40milliseconds. So, that's the paradox.
  580. 20:43Like, if there's so many possibilities,
  581. 20:45how does nature do this, right? And
  582. 20:47basically, how does physics uh achieve
  583. 20:49this?
  584. 20:50Um and that gives you hope that this
  585. 20:52must be tractable computationally in
  586. 20:54some reasonable amount of time because
  587. 20:56uh physics does solve this problem uh um
  588. 20:59you know, uh billions of times a second
  589. 21:01in the body.
  590. 21:03And um furthermore, what attracted me to
  591. 21:04this um problem was that there was a
  592. 21:07bi-annual competition called the CASP
  593. 21:08competition, which is like you can think
  594. 21:10of it as like the Olympics for protein
  595. 21:12folding. And it happens every 2 years
  596. 21:14and it's run by some amazing people led
  597. 21:16by Professor John Moult,
  598. 21:18University of Maryland. And it's been
  599. 21:21running for since 1994 and it's a great
  600. 21:23competition because they work with
  601. 21:26experimentalists and who who
  602. 21:28painstakingly find these structures
  603. 21:30using very exotic and
  604. 21:33expensive equipment like electron
  605. 21:35microscopes. And and they newly
  606. 21:38discovered structures that haven't been
  607. 21:39published yet, they put it into the
  608. 21:41competition. So you actually the the the
  609. 21:43the competition organizers know what the
  610. 21:46ground truth is, but the computational
  611. 21:48teams, you know, hundreds of teams enter
  612. 21:50every year every every competition every
  613. 21:52couple of years, and they try with their
  614. 21:54computational methods to predict that
  615. 21:56those structures. And there's usually
  616. 21:57around 100 proteins that are in the
  617. 21:59competition. And then at the end of the
  618. 22:01summer, they reveal what the true
  619. 22:03structures are and you can compare the
  620. 22:05predicted ones and their distances, the
  621. 22:07error in the predictions, to to the real
  622. 22:10structures.
  623. 22:13So we entered AlphaFold 1 actually for
  624. 22:16the first time in 2018 and we started
  625. 22:19this AlphaFold project in 2016, actually
  626. 22:22pretty much the day after we we got back
  627. 22:25from the AlphaGo match in Seoul for in
  628. 22:27Korea. We felt that and I felt that we
  629. 22:29were ready, we had the techniques that
  630. 22:31were mature enough and ready to now be
  631. 22:33applied outside of games and to try and
  632. 22:36tackle really meaningful
  633. 22:39problems. We call them kind of root node
  634. 22:40problems because they open if they could
  635. 22:42be solved, they they open up whole new
  636. 22:44branches and avenues of discovery that
  637. 22:46can be built on top. And protein folding
  638. 22:48was was a prime example of that. So we
  639. 22:51started working on it 2016. We we
  640. 22:53AlphaFold 1 was ready after a couple of
  641. 22:55years and we entered it into the CASP 13
  642. 22:57competition. And you can see for the
  643. 22:59decade prior,
  644. 23:01uh bar charts are showing the winning
  645. 23:03score of the winning team
  646. 23:05uh uh in the hardest category, actually,
  647. 23:08the hardest proteins that were being
  648. 23:09predicted. And you can sort of think of
  649. 23:11it as a percentage accuracy of how many
  650. 23:14of the atomic how many of these amino
  651. 23:16acids have you got in the right position
  652. 23:17within a certain tolerance?
  653. 23:19Uh within the sort of width of an atom
  654. 23:21you need to predict it within. And um
  655. 23:24you can see there was sort of not much
  656. 23:25progress for a decade and we we were
  657. 23:27stuck at this um 60 points level. And
  658. 23:31you we for the
  659. 23:37which is effectively if you got to 90,
  660. 23:39you would be within the width of an
  661. 23:41atom. So, you'd be atomic accuracy. And
  662. 23:43that's what we were told by
  663. 23:44experimentalists was the accuracy you'd
  664. 23:46had to reach so that it was competitive
  665. 23:48with experimental methods so that
  666. 23:50experimentalists could actually rely on
  667. 23:53these predictions rather than having to
  668. 23:55necessarily do the laborious,
  669. 23:56painstaking uh work to um uh to find
  670. 24:00that structure. And uh just as a rule of
  671. 24:02thumb, my biologist friends would always
  672. 24:04tell me that, you know, it it it take a
  673. 24:06a a PhD student their entire PhD, so 4-5
  674. 24:09years, to fold uh to find the structure
  675. 24:11of just one protein. Um and um and and
  676. 24:15there are 200 million proteins uh known
  677. 24:17to science and 20,000 uh uh proteins in
  678. 24:21the human uh uh proteome.
  679. 24:23So, uh so with AlphaFold 1, we were able
  680. 24:26to uh win this competition and and sort
  681. 24:29of um be better by almost 50% than the
  682. 24:31next best system. And AlphaFold 1 for
  683. 24:33the first time introduced machine
  684. 24:35learning techniques as the main
  685. 24:37component of the system uh for the first
  686. 24:39time. Uh but it was not enough to to
  687. 24:42reach this atomic accuracy. We actually
  688. 24:43had to go back to the drawing board with
  689. 24:45what we'd learned and re-architected uh
  690. 24:48for AlphaFold 2 from scratch uh
  691. 24:51using all the learnings we had from
  692. 24:52AlphaFold 1 to finally reach this atomic
  693. 24:54accuracy. And that led the organizers to
  694. 24:57declare that the problem had been solved
  695. 24:59in uh the end of 2020.
  696. 25:02So, this is an example of how AlphaFold
  697. 25:04works visually. So, you can see on the
  698. 25:06left-hand side here is a very complex
  699. 25:08protein. The ground truth is in green,
  700. 25:11the predicted structure is in blue, and
  701. 25:13you can see how closely the blue
  702. 25:15overlaps the green.
  703. 25:17And then on the right-hand side, you can
  704. 25:19see how AlphaFold 2 works. It sort of
  705. 25:21builds up that structure in an iterative
  706. 25:23process. It kind of recycles itself
  707. 25:25actually over 192 steps,
  708. 25:27and then builds out starts as a as a
  709. 25:29scrunched ball of of protein matter,
  710. 25:32amino acids, and then it builds out more
  711. 25:34and more plausible structure.
  712. 25:36At the end, it sort of refines the last
  713. 25:38parts of it until it has the finished
  714. 25:40prediction.
  715. 25:42We we immediately then tried to because
  716. 25:46AlphaFold is so accurate, but it's not
  717. 25:48just accurate, it's also extremely fast.
  718. 25:50It's able to fold proteins in a matter
  719. 25:52of seconds. An average protein, we
  720. 25:54realized quite quickly that we could
  721. 25:56actually fold all 200 million proteins
  722. 25:59known to science. Over the course of a
  723. 26:02year, we used a lot of computers on the
  724. 26:04Google Cloud to to fold all of them and
  725. 26:06then put them out freely
  726. 26:08on a database with our colleagues at
  727. 26:10EMBL-EBI just up the road at the Sanger
  728. 26:13Center just outside of Cambridge. And we
  729. 26:16provided that for free unrestricted
  730. 26:18access to anyone in the world to use it.
  731. 26:21And that 200 million proteins, if you
  732. 26:23think about how long it takes to do that
  733. 26:25experimentally, 4-5 years, it's kind of
  734. 26:28like a billion years of PhD time done in
  735. 26:301 year. So,
  736. 26:33you know, it's it's kind of amazing to
  737. 26:34think about how much science could be
  738. 26:36accelerated. And it opened up whole new
  739. 26:39avenues of exploration because many of
  740. 26:42these structures, especially for the
  741. 26:44less well-studied organisms like some
  742. 26:46certain types of plants, they're very
  743. 26:48important for science and for
  744. 26:50agriculture um, and agricultural
  745. 26:53research, but that there's there's
  746. 26:54there's that you wouldn't have any
  747. 26:56almost none of those structures will
  748. 26:58have been found and be available. So,
  749. 27:00now those are all available and also 200
  750. 27:02million, you can look at them at an
  751. 27:04aggregate level and sort of look at
  752. 27:06structural uh structures, um, across
  753. 27:08species, uh, and, uh, kind of meta
  754. 27:11structures and see what the
  755. 27:13commonalities are, uh, through
  756. 27:14evolution. So, there's really
  757. 27:16interesting new avenues of of uh
  758. 27:18branches of of structural biology now
  759. 27:20that are being explored.
  760. 27:22And of course we thought about safety
  761. 27:23from the beginning, uh, and we take our
  762. 27:25responsibility very seriously at the
  763. 27:27forefront of AI and in this case we, you
  764. 27:30know, consulted with over 30 biosecurity
  765. 27:32and bioethics experts to make sure that,
  766. 27:35um, what we're putting out into the
  767. 27:36world, um, what the benefits of that far
  768. 27:38outweighed any any, uh, risks associated
  769. 27:41with it. And now, uh, I'm very proud to
  770. 27:44say that, you know, over 2 million
  771. 27:45researchers are using it, um, from
  772. 27:47pretty much every country in the world,
  773. 27:50um, and, uh, and it's been sort of sort
  774. 27:52of cited over 30,000 times now and it's
  775. 27:53become a standard tool, uh, in in in
  776. 27:57biology research now and many of you in
  777. 27:59the audience who are PhD students,
  778. 28:00hopefully you're using it and making use
  779. 28:02of it and it's just, um, part of the
  780. 28:04standard canon now that is used for
  781. 28:06biology research.
  782. 28:09Uh, it's been amazing to see what other
  783. 28:11researchers have done uh with all of
  784. 28:13this uh technology and all these
  785. 28:14structures. Um,
  786. 28:16I've just called out uh six of my
  787. 28:18favorite uh uh examples. Uh, people at
  788. 28:21uh University of Portsmouth are using it
  789. 28:23a research group is using it to tackle
  790. 28:25plastic pollution in the environment,
  791. 28:27trying to design new enzymes which are
  792. 28:29types of proteins uh that can digest
  793. 28:31plastic. Um,
  794. 28:33we're working with the Fleming Center on
  795. 28:35antibiotic resistance, um,
  796. 28:38neglected diseases, uh, like tropical
  797. 28:40diseases that affect the poorer parts of
  798. 28:42the world. We work with uh Drugs for
  799. 28:44Neglected Diseases Institute. Um, and
  800. 28:46here's a good example of where we can
  801. 28:48accelerate research in those areas
  802. 28:50where, you know, whether it's malaria,
  803. 28:52leishmaniasis,
  804. 28:53um, Zika virus, uh, a lot of those
  805. 28:55structures are not known. Uh, but now
  806. 28:57they can go straight to drug discovery
  807. 28:59because, uh, they have a lot of the
  808. 29:01information about the structures for
  809. 29:03those viruses and and bacteria.
  810. 29:06Uh, and then there's been a lot of
  811. 29:07fundamental research being done on
  812. 29:09things like, uh, finding the the
  813. 29:11structure of the nano pore complex,
  814. 29:13which is a very important protein that
  815. 29:15that that lets in nutrients in and out
  816. 29:17of the nuclear pore of the cell. Um,
  817. 29:20as amazing work at the Broad Institute
  818. 29:21done with drug delivery, sort of
  819. 29:23designing molecular syringes,
  820. 29:25redesigning, uh, proteins that can
  821. 29:27deliver drugs targeted to particular
  822. 29:30part of the body. Uh, and it's even
  823. 29:32being used in in things like, uh,
  824. 29:33looking at mechanisms of of fertility.
  825. 29:36So, the the amount of things that it's
  826. 29:38being used for is sort of almost every
  827. 29:40area of biology and medical research now
  828. 29:42is is, uh, making use of AlphaFold.
  829. 29:47We've continued in the last few years to
  830. 29:49develop, uh, more struct more, uh, uh,
  831. 29:51uh,
  832. 29:52developments and improving the systems.
  833. 29:54We released AlphaFold 3, uh, earlier
  834. 29:56this year, uh, uh, for academics to use,
  835. 29:59and we've extended now AlphaFold 3 to to
  836. 30:02deal with, um, interactions. So, um, you
  837. 30:05can think of AlphaFold 2 as a picture of
  838. 30:07the the static, uh, protein structure,
  839. 30:10but really biology is a dynamic process.
  840. 30:12So, you need to understand how how, um,
  841. 30:15different biological elements interact
  842. 30:16with each other. So, of course, proteins
  843. 30:18with other proteins, but also proteins
  844. 30:20with others other molecules uh,
  845. 30:23important to life, things like DNA and
  846. 30:25RNA, and also ligands. So, uh, small
  847. 30:28molecules which, uh, you know, things
  848. 30:30like drug compounds. How does the
  849. 30:32protein bind with, um, with that
  850. 30:35compound?
  851. 30:37And then we have a separate set of work,
  852. 30:38AlphaProteo, which is doing the reverse
  853. 30:41of AlphaFold, uh, but making use of
  854. 30:43still of of the AlphaFold techniques
  855. 30:45where if you want to design a novel
  856. 30:47protein, maybe that doesn't exist in
  857. 30:49nature for a particular job, a
  858. 30:51particular function, what is the amino
  859. 30:53acid sequence and the genetic sequence
  860. 30:56that will give you that structure. So,
  861. 30:58um so it's kind of like running it in
  862. 31:00reverse and trying to design new
  863. 31:02structures that will do novel things um
  864. 31:05and again could be extremely useful uh
  865. 31:07in in for for designing drugs and things
  866. 31:09like that. Antibiotics and antibodies.
  867. 31:15So, taking a step back then, what is uh
  868. 31:17or you know, if I look at all the work
  869. 31:18we've done in the last 15 years, um
  870. 31:21what's the implications uh for science
  871. 31:23and and also machine learning? Um and if
  872. 31:26you think about what we've done with
  873. 31:27first of all our games work uh and then
  874. 31:29now with the scientific work that we've
  875. 31:31been working on of which AlphaFold is
  876. 31:33our best example, um it's all about
  877. 31:35making this search tractable. You have
  878. 31:38this incredibly complex problem, there's
  879. 31:40many, many possible solutions to the
  880. 31:42problem, and you've got to find uh the
  881. 31:44optimal solution, kind of needle in the
  882. 31:46haystack of that uh enormous
  883. 31:48combinatorial search space. And you
  884. 31:51can't do it by brute force, so you have
  885. 31:53to learn this neural network model um so
  886. 31:56it sort of learns about the topology of
  887. 31:58the problem uh so that you can
  888. 32:00efficiently guide the search to reach
  889. 32:03your uh to maximize uh or find the
  890. 32:05optimal solution to the objective that
  891. 32:07you have in mind. And I think this is an
  892. 32:10incredibly general way uh
  893. 32:12uh and general solution, an incredibly
  894. 32:14general way to approach um a whole
  895. 32:16myriad of problems.
  896. 32:17And so, we think about back to the Go
  897. 32:20example, so we're trying to use these
  898. 32:21systems to find the best Go move, but
  899. 32:24you could also change those nodes to be
  900. 32:27uh chemical compounds, and now you're
  901. 32:29trying to find the best molecule in
  902. 32:31chemistry space, in chemical space. And
  903. 32:34the best molecule, you know, and this
  904. 32:36this is the beginning of drug design.
  905. 32:38And and that will bind specifically to
  906. 32:40the target you're interested in, but
  907. 32:42nothing else. So, reduces the side
  908. 32:44effects and the toxicity of that
  909. 32:46compound. Um and it's a very very
  910. 32:48similar um techniques that we're using
  911. 32:51in order to design these molecules now
  912. 32:53as the next steps as we as we move more
  913. 32:55and more into drug discovery.
  914. 32:58So, I think in biology at least, um I
  915. 33:01feel like we're entering a new era now
  916. 33:03of what I like to call digital biology.
  917. 33:05So, you know, I think of biology in its
  918. 33:07most fundamental level as an information
  919. 33:09processing system, you know, that's
  920. 33:11trying to resist entropy around it. And
  921. 33:14I think that's basically what life is.
  922. 33:16Um of course, it's a phenomenally
  923. 33:17complex and emergent information
  924. 33:20processing system.
  925. 33:21Uh and I think that's where AI comes in.
  926. 33:23Just like maths uh and the maths that I
  927. 33:25learned in this room was the perfect
  928. 33:27description language for physics um and
  929. 33:30physic physic physical phenomena, I
  930. 33:32think that AI is potentially the perfect
  931. 33:34description language for biology. Um
  932. 33:36it's perfect for dealing with the
  933. 33:37complexities of the emergent uh
  934. 33:39behaviors and interactions that you get
  935. 33:41in a dynamic system like biology.
  936. 33:45And I think AlphaFold is a proof point
  937. 33:47of that. Um and I hope when we look back
  938. 33:49in 10 years' time, it won't be an
  939. 33:51isolated breakthrough, but it actually
  940. 33:53um have sort of heralded in this new era
  941. 33:56um golden era of digital biology. And
  942. 33:59we're trying to progress that ourselves.
  943. 34:01We started a new spin-out company
  944. 34:02Isomorphic Labs to build on our
  945. 34:04AlphaFold technology and move more into
  946. 34:07the chemistry space that I was just
  947. 34:08talking about and actually try and
  948. 34:10reimagine drug discovery from first
  949. 34:12principles with AI. And right now it
  950. 34:14takes an average of 10 years um for a
  951. 34:17drug to be developed um and it's
  952. 34:19extraordinarily expensive. It costs
  953. 34:21billions and billions of dollars. And
  954. 34:22so, I'm thinking, why can't we use these
  955. 34:24techniques to reduce that down from
  956. 34:26years to months, maybe even one day
  957. 34:29weeks, just like we reduced down the
  958. 34:32discovery of protein structures from
  959. 34:34potentially years down to now minutes
  960. 34:37and seconds.
  961. 34:38And so we think of this as a kind of
  962. 34:41doing science at digital speed. So
  963. 34:44trying to bring the best of what we do
  964. 34:46in the technology area to the
  965. 34:49natural sciences.
  966. 34:52And my dream one day is to be able to
  967. 34:54create maybe a kind of virtual cell,
  968. 34:57computational cell perhaps of something
  969. 34:59very simple like a yeast cell
  970. 35:01and that you can actually do run
  971. 35:03experiments in silico on it and the
  972. 35:05predictions that you get out of the
  973. 35:06virtual cell will actually tell you and
  974. 35:09form your your real world experiments in
  975. 35:11the lab and you can reduce down a lot of
  976. 35:13the search that's done in the wet lab
  977. 35:15and actually more use the wet lab for
  978. 35:18validation steps rather than the very
  979. 35:20expensive
  980. 35:22and slow search process.
  981. 35:26Of course AI and we've been using AI not
  982. 35:28just in biology, but it can be used for
  983. 35:30science, mathematics, medicine more
  984. 35:33generally
  985. 35:34and we've had a whole range of
  986. 35:35breakthroughs not just in the biological
  987. 35:37sciences, but from things in health like
  988. 35:40identifying eye disease from retinal
  989. 35:41scans, discovering new materials,
  990. 35:44helping with plasma containing fusion
  991. 35:46reactors, faster algorithms
  992. 35:49so AI discovering better algorithms for
  993. 35:51itself like faster matrix
  994. 35:53multiplication, doing weather prediction
  995. 35:55and even helping with quantum computers
  996. 35:57and error correction in quantum
  997. 35:59computing. And that's just a small
  998. 36:00example of some of the work we've been
  999. 36:02doing in the last two three years and I
  1000. 36:05think AI will be almost be applicable to
  1001. 36:07pretty much every field and I always
  1002. 36:10encourage actually universities to start
  1003. 36:12thinking very seriously about
  1004. 36:14multi-disciplinary work where you apply
  1005. 36:16AI to the right questions in a
  1006. 36:19particular specialist field and I think
  1007. 36:21there's many many advances to be made
  1008. 36:23over the next five to 10 years by doing
  1009. 36:25that.
  1010. 36:28So I'll just end then with a little bit
  1011. 36:30of
  1012. 36:31a more general view about not just AI
  1013. 36:34for science, but the path to AGI and how
  1014. 36:36close we are to that and our more
  1015. 36:38general work on the original mission of
  1016. 36:40of AGI.
  1017. 36:41And we've been making a lot of advances
  1018. 36:43in all areas of general understanding of
  1019. 36:46the world. We sometimes call them world
  1020. 36:48models.
  1021. 36:49So, we're particularly proud of our new
  1022. 36:51video model called VO2, which was just
  1023. 36:53released at the end of last year. It's
  1024. 36:55it's state of the art video generation
  1025. 36:58and it's able to generate these videos
  1026. 37:01just from a text description, right? Or
  1027. 37:04a single static image. So, you know, and
  1028. 37:07actually although some of these videos
  1029. 37:08may not seem that impressive, if you
  1030. 37:10think about this chopping the tomato
  1031. 37:13one, this is like the Turing test for
  1032. 37:14video models because usually you get,
  1033. 37:17you know, the tomato comes magically
  1034. 37:18back together or you're chopping through
  1035. 37:20the fingers or the knife moves off
  1036. 37:21somewhere and it's actually if you think
  1037. 37:23about what the systems had to do to
  1038. 37:25really understand the physics of the
  1039. 37:27world or this or the bubbles around this
  1040. 37:29blueberry here, right? It's just
  1041. 37:30generating that from text, you know,
  1042. 37:32blueberries dropping into a glass of
  1043. 37:34water and it's it's doing all the
  1044. 37:36physics of this correctly or the motion
  1045. 37:38of
  1046. 37:39you know, these these little cartoon
  1047. 37:41characters or the bee here. It's kind of
  1048. 37:44mind-blowing really and I think even if
  1049. 37:45you told me 5 years ago
  1050. 37:48that this would be possible without sort
  1051. 37:50of building in some special
  1052. 37:51understanding of physics or something, I
  1053. 37:53would have told you that that's seems
  1054. 37:55unlikely that that would be possible.
  1055. 37:57Then yet somehow these learning systems
  1056. 37:59are able to learn about real-world
  1057. 38:01physics just from watching, you know,
  1058. 38:03many many YouTube videos. And
  1059. 38:06it's it's pretty crazy that that's
  1060. 38:07possible. So,
  1061. 38:09we've done that. We've gone a step
  1062. 38:11further with Genie 2, which of course
  1063. 38:12bringing my games
  1064. 38:14hat back in here and this is taking
  1065. 38:16those video models a step further. So,
  1066. 38:18now with a text instruction, you can
  1067. 38:20generate a whole game. So, you know,
  1068. 38:23here we at the bottom here we said
  1069. 38:24generate a playable world as a robot in
  1070. 38:26a futuristic city, and it just comes up
  1071. 38:28with this, and you can control it with
  1072. 38:30the, you know, QWE keys and the arrow
  1073. 38:32keys. Um at the moment it's only
  1074. 38:34consistent for a few seconds, um but
  1075. 38:36we're we're we're working to extend that
  1076. 38:38to so that the consistency of the game
  1077. 38:41world lasts for many minutes.
  1078. 38:43Um and so then you've really got what I
  1079. 38:44would call a world model, a really an
  1080. 38:47understanding of the real world. Um and
  1081. 38:49and how interactions in that real world
  1082. 38:51work and the physics of the real world
  1083. 38:53work.
  1084. 38:55Um of course we've been working very
  1085. 38:56hard on the safety aspects of this. Uh
  1086. 39:00and from about very from the very
  1087. 39:01beginning in 2010, we were working on um
  1088. 39:05uh planning for success even though
  1089. 39:06almost nobody was working on AI back
  1090. 39:08then. Um we we imagined that it would be
  1091. 39:10a 20-year mission, and actually
  1092. 39:12amazingly we're sort of on track 15
  1093. 39:14years in. And um we we were sort of
  1094. 39:17planning for success. If we were to
  1095. 39:19build these kinds of transformative
  1096. 39:20systems and technologies, um it would
  1097. 39:23come with a lot of uh uh responsibility
  1098. 39:25as well to make sure they get deployed
  1099. 39:27in a safe uh and responsible way. And
  1100. 39:30one of the uh systems we technology we
  1101. 39:32built is called SynthID, which um
  1102. 39:35invisibly watermarks uh actually using
  1103. 39:37an AI system and adversarial AI system
  1104. 39:40uh slightly adjust the pixels or the
  1105. 39:42text or the or the audio um
  1106. 39:45imperceptibly to the human ear or eye,
  1107. 39:47um but uh it can be detected by a
  1108. 39:49detection system that these were
  1109. 39:51synthetically generated images, whether
  1110. 39:53that's audio, um image, or video. And of
  1111. 39:57course it's going to become increasingly
  1112. 39:58important um as uh as as these
  1113. 40:01technologies become widely deployed that
  1114. 40:03we're able to easily distinguish between
  1115. 40:05synthetically generated images and real
  1116. 40:07images.
  1117. 40:09So, AI's been, you know, this incredible
  1118. 40:11potential uh to to help with our
  1119. 40:14greatest challenges, from climate to to
  1120. 40:17health. but obviously this is going to
  1121. 40:19affect everyone.
  1122. 40:20Um so I think it's really important that
  1123. 40:22we engage not the just it's not just the
  1124. 40:24technologists deciding this, but that we
  1125. 40:26engage with a wide range of stakeholders
  1126. 40:28from society. So I've been really
  1127. 40:30pleased in the last couple of years that
  1128. 40:32one of the consequences of AI becoming
  1129. 40:34mainstream is that many governments have
  1130. 40:36gotten interested in it and many parts
  1131. 40:38of all parts of society. And I think
  1132. 40:39it's been great to see these
  1133. 40:40international summits. Actually the UK
  1134. 40:42hosted the first one in Bletchley Park
  1135. 40:44couple of years ago bringing together
  1136. 40:46heads of government with academia and
  1137. 40:48civil society to discuss uh these
  1138. 40:51technologies, um how to put guard the
  1139. 40:53right guardrails on it, how to make sure
  1140. 40:55we embrace the opportunities, um but we
  1141. 40:58mitigate the risks um that are coming
  1142. 41:00down the line. And uh I think that's
  1143. 41:02going to become increasingly important
  1144. 41:04um given the exponential improvement uh
  1145. 41:06that we're seeing with these
  1146. 41:07technologies. So my my my short answer
  1147. 41:10for this is to say, you know, with with
  1148. 41:12a lot of Silicon Valley tries to the
  1149. 41:14kind of mantra in Silicon Valley is is
  1150. 41:16like move fast and break things. And of
  1151. 41:18course that's that's created a lot of
  1152. 41:19advances, a lot of of the technologies
  1153. 41:22we all use every day today, but I think
  1154. 41:24it's not appropriate in my opinion for
  1155. 41:26this type of transformative technology.
  1156. 41:28I think instead, you know, we should be
  1157. 41:30trying to use uh uh uh the scientific
  1158. 41:33method and approach it with a kind of
  1159. 41:35humility and respect that this kind of
  1160. 41:37technology deserves. And you know, not
  1161. 41:39we we don't know a lot of things. There
  1162. 41:40are a lot of unknowns around how this
  1163. 41:42technology is going to develop. It's so
  1164. 41:44new. Um um and I think with exceptional
  1165. 41:47sort of care and foresight we can get
  1166. 41:49all the benefits
  1167. 41:50uh and minimize uh the downsides of
  1168. 41:52this, but I think only if we um
  1169. 41:55start the research and the debate about
  1170. 41:57that now.
  1171. 42:00So just to end then, we're now building
  1172. 42:02our own big multimodal models that try
  1173. 42:04and take the best of all of these
  1174. 42:05different models I've shown you and put
  1175. 42:07it into one system. Uh we call it the
  1176. 42:09Gemini series. Our latest one's Gemini
  1177. 42:112.0 that some of you may have tried,
  1178. 42:13which is state-of-the-art across many
  1179. 42:15leading benchmarks. Um we're using it to
  1180. 42:18further that I'm very excited about the
  1181. 42:20next generation of assistance. I call it
  1182. 42:22universal assistance. We call it Project
  1183. 42:24Astra, where actually you have it on
  1184. 42:26your phone or some other devices, maybe
  1185. 42:28glasses, and it starts is being an
  1186. 42:30assistant that you can take around with
  1187. 42:31you in the in the real world, and it
  1188. 42:33helps you in everyday life to um to to
  1189. 42:36get to enrich your life or to make you
  1190. 42:38more productive.
  1191. 42:41And the next step then in AI is
  1192. 42:42combining what I've shown you with
  1193. 42:44AlphaGo, these kinds of um agent-based
  1194. 42:47models that are able to um effec-
  1195. 42:50efficiently search through and find a
  1196. 42:52good solution uh to a problem in a in a
  1197. 42:55limited domain, in this case in games.
  1198. 42:57But we want to actually build those
  1199. 42:59types of search systems and planning
  1200. 43:01systems on top of much more general
  1201. 43:03models like Gemini, these world models
  1202. 43:05that understand uh how the real world
  1203. 43:07works, and then can then plan and and
  1204. 43:10achieve things in the real world. And of
  1205. 43:12course that's key to things like
  1206. 43:14robotics uh working, which I think in
  1207. 43:15the next 2 3 years is going to be a huge
  1208. 43:18area um that's going to going to have a
  1209. 43:21a massive advances in it.
  1210. 43:24So I'll just finish then by just having
  1211. 43:26a a slight conjecture about what does
  1212. 43:29this all mean if we think back to Turing
  1213. 43:31and all of the work that he did to lay
  1214. 43:33down the foundations uh of computer
  1215. 43:36science. And um
  1216. 43:38and I think that uh if you see the work
  1217. 43:40that we've done, I see myself as a kind
  1218. 43:42of Turing's champion in a way. Like he
  1219. 43:44how far can the Turing machines and this
  1220. 43:46idea of classical computing go? And I
  1221. 43:49feel like um probably one of the
  1222. 43:50lectures that I took in this room was
  1223. 43:52one of my favo- favorite things to think
  1224. 43:54about is the P equals NP problem, which
  1225. 43:57is uh a famous problem in computer
  1226. 43:59science of, you know, what sorts of
  1227. 44:00problems are tractable on classical
  1228. 44:02systems. And um and there's obviously a
  1229. 44:05lot of great work going on in quantum
  1230. 44:06computing systems, uh many of that here
  1231. 44:09in Cambridge and also Google we have one
  1232. 44:11of the top computer quantum computing
  1233. 44:13groups in the world. And there's a lot
  1234. 44:15of things that are thought to require
  1235. 44:17quantum computing to solve. A lot of
  1236. 44:19real world systems that we would like to
  1237. 44:22understand and model. And my conjecture
  1238. 44:24is that actually classical Turing
  1239. 44:26machines, basically classical machines
  1240. 44:28that these types of AI systems are built
  1241. 44:30on, can do a lot more than we previously
  1242. 44:33gave them credit for. And if you think
  1243. 44:34about AlphaFold and protein folding, you
  1244. 44:37know, proteins are quantum systems, you
  1245. 44:39know, they operate at the atomic scale
  1246. 44:41and one might think you need quantum
  1247. 44:43simulations to actually be able to find
  1248. 44:46the structures of proteins. And yet we
  1249. 44:48were able to approximate those solutions
  1250. 44:51with our neural networks.
  1251. 44:53And so I think, you know, one one
  1252. 44:55potential idea here is that any pattern
  1253. 44:58that can be generated or found in
  1254. 45:00nature, so I it has some real structure,
  1255. 45:02physical structure, can be efficiently
  1256. 45:05discovered and modeled by one of these
  1257. 45:07classical learning algorithms like
  1258. 45:09AlphaFold.
  1259. 45:10And if that turns out to be true, I
  1260. 45:11think that has all sorts of implications
  1261. 45:14for quantum mechanics and and actually
  1262. 45:16fundamental physics,
  1263. 45:17which, you know, is something that I
  1264. 45:19hope to explore and many of my
  1265. 45:21colleagues hope to explore maybe with
  1266. 45:22the help of these classical systems as
  1267. 45:24well to help us uncover what the true
  1268. 45:26nature of reality might be.
  1269. 45:29And that leads me back to the whole
  1270. 45:31reason why I started my path on AI many
  1271. 45:33many years ago is that I always believed
  1272. 45:36that AGI built in this way could be the
  1273. 45:39ultimate general purpose tool to
  1274. 45:40understand the universe around us and
  1275. 45:42our place in it. Thank you.
  1276. 45:53>> [applause]
  1277. 45:59>> Great, we have a time for some questions
  1278. 46:01if people have questions. A first hand
  1279. 46:03shot up just here.
  1280. 46:04Hi. Thank you.
  1281. 46:07Thank you. Thank you for the great talk.
  1282. 46:09Because you have a background in
  1283. 46:11neuroscience,
  1284. 46:12and you really like to think in terms of
  1285. 46:14root note problems, was there ever a
  1286. 46:16root note problem you came across in
  1287. 46:18neuroscience that you thought was worth
  1288. 46:20tackling and still worth tackling to
  1289. 46:23understand biological and artificial
  1290. 46:25intelligence better?
  1291. 46:27>> Yeah, there's many. I mean, I that's
  1292. 46:29what I studied for my PhD actually was
  1293. 46:32memory, but also imagination. So, future
  1294. 46:35thinking, kind of planning. So, I really
  1295. 46:37wanted to understand
  1296. 46:39how the brain did that. And it turns out
  1297. 46:41the hippocampus is involved in both, so
  1298. 46:44that we could maybe mimic that with with
  1299. 46:46some of these algorithms. So, I think
  1300. 46:48there are many key things in that. Of
  1301. 46:50course,
  1302. 46:51there are all the big questions around,
  1303. 46:53you know, creativity, dreaming,
  1304. 46:55consciousness, all these big questions
  1305. 46:57that I think building AI and then
  1306. 47:01comparing it to the human mind is one of
  1307. 47:03the best ways I think we'll make
  1308. 47:04progress with those sorts of root note
  1309. 47:07problems, like, you know, what is the
  1310. 47:08nature of consciousness and is there
  1311. 47:09something special about the
  1312. 47:11instantiation of the substrate of the
  1313. 47:13brain versus
  1314. 47:15algorithmically mimicking that in
  1315. 47:17silicon.
  1316. 47:20Great.
  1317. 47:21Got a question just here.
  1318. 47:23Hi.
  1319. 47:25I have two questions actually. So, since
  1320. 47:27DeepMind was founded before the deep
  1321. 47:30learning revolution, I wanted to know
  1322. 47:32what your state of mind was had deep
  1323. 47:35learning not picked up, or how were you
  1324. 47:38going to progress?
  1325. 47:40That's the first question. The second
  1326. 47:42question is,
  1327. 47:43since you've had intimate experiences
  1328. 47:46with such challenging problems, such
  1329. 47:48high-dimensional problems,
  1330. 47:50and we know that gradient descent and
  1331. 47:53its variants can't
  1332. 47:55uh, of converge to optimum solution,
  1333. 47:58only locally optimum solutions. Were you
  1334. 48:01surprised that anything works at all in
  1335. 48:03these systems at every
  1336. 48:06at every every point in time? And
  1337. 48:08do you think that most of the nature is
  1338. 48:11kind of suboptimal? And so we can
  1339. 48:14potentially build a more optimal nature.
  1340. 48:16So, look, I I think they're both great
  1341. 48:18questions. So, the first the first one,
  1342. 48:21look, that's why we called it DeepMind,
  1343. 48:23partly because the deep refers to deep
  1344. 48:25learning, right? So, we we deep learning
  1345. 48:28had Well, the early parts of the It
  1346. 48:30wasn't called deep learning then, but it
  1347. 48:31was it was sort of just becoming
  1348. 48:33common. There was these Boltzmann
  1349. 48:34machines and things that Geoffrey Hinton
  1350. 48:36had invented and
  1351. 48:38in just a few years before in 2006,
  1352. 48:402005, these hierarchical neural
  1353. 48:42networks. And it seemed like a super
  1354. 48:45promising idea even back then to us that
  1355. 48:47come across it in academia. And
  1356. 48:51then the other thing we bet on was
  1357. 48:52reinforcement learning and the
  1358. 48:54combination of that, right? Which again
  1359. 48:56is coming back into vogue, but it was
  1360. 48:58also important for us to solve something
  1361. 49:00like AlphaGo. So, you need both parts of
  1362. 49:02that. You need to the deep learning to
  1363. 49:03kind of model the environment and the
  1364. 49:05world, and then you need the
  1365. 49:06reinforcement learning to make the plans
  1366. 49:08and and and the solutions and take
  1367. 49:10action in the world. And
  1368. 49:12so there was two reasons we bet on that
  1369. 49:14even when it it was just sort of the
  1370. 49:15beginning of it is that we we knew that
  1371. 49:18the classical methods, these expert
  1372. 49:20systems, would not scale. And actually
  1373. 49:22that's again one of the things I learned
  1374. 49:24here and also at my post doc at MIT was
  1375. 49:26that was that they were the kind of
  1376. 49:28churches, if you like, of the classical
  1377. 49:30methods, these expert systems. And
  1378. 49:32actually that's something else you can
  1379. 49:33learn here is like, you know, in your
  1380. 49:35university courses is not just what to
  1381. 49:37do, but also what not to do. And why it
  1382. 49:39may not like like I sort of thought
  1383. 49:41about it and I felt it would never scale
  1384. 49:43to
  1385. 49:44the kinds of problems that I wanted to
  1386. 49:46solve with AI. Whereas the learning
  1387. 49:48system seemed to have unlimited
  1388. 49:50potential, although they were a lot
  1389. 49:51harder to get anything to do
  1390. 49:53significant to do at the beginning.
  1391. 49:55Right, and this was the problem cuz they
  1392. 49:56weren't scaled up enough. And the other
  1393. 49:58the the other reason we started DeepMind
  1394. 49:592010 is we could also see the computing
  1395. 50:02paradigm was shifting on the hardware
  1396. 50:04side, GPUs and other things
  1397. 50:07which were of course also invented for
  1398. 50:08gaming. And turns out everything's a
  1399. 50:10matrix multiplication, right?
  1400. 50:12Intelligence and gaming and computer
  1401. 50:14graphics. So,
  1402. 50:16and so all of those different influences
  1403. 50:18came together and the new the you know,
  1404. 50:20the understanding of neuroscience and
  1405. 50:22fMRI machines and neuroscience had
  1406. 50:24advanced a lot in that previous 10
  1407. 50:26years, too. So, I felt it was the
  1408. 50:27perfect time to sort of bring all of
  1409. 50:29that together
  1410. 50:31back in 2010. And we were betting on
  1411. 50:33that, not necessarily because we knew it
  1412. 50:34would work, but we knew that the other
  1413. 50:36methods we were pretty confident would
  1414. 50:38not work.
  1415. 50:39And that's what the AI winters were
  1416. 50:40about, basically with people trying to
  1417. 50:42push those expert systems. And then the
  1418. 50:44second question, you know,
  1419. 50:46um
  1420. 50:47I think I wouldn't say that well, first
  1421. 50:50of all, it is surprising that some of
  1422. 50:52these things converge and actually we
  1423. 50:54weren't sure. So, you know, that Atari
  1424. 50:55stuff I showed you. So, for the first
  1425. 50:57couple of years, nothing worked. Right?
  1426. 50:59So, we couldn't even get one point on if
  1427. 51:01some of you remember Pong, the the you
  1428. 51:03know, one of the first computer games,
  1429. 51:04we couldn't this sort of tennis bat and
  1430. 51:06ball game, simplest game you could
  1431. 51:07imagine, we weren't able to get a point
  1432. 51:09on it. So, we were wondering whether
  1433. 51:11were we like 10, 20 years too early?
  1434. 51:14Like like Babbage turned out to be with
  1435. 51:15his difference engine, right? Amazing
  1436. 51:17ideas, it worked, but he was just in the
  1437. 51:18end 50 years or 100 years too early. And
  1438. 51:22and I always say you want to be like 5
  1439. 51:23years ahead of your time, not 50 years
  1440. 51:25ahead. Otherwise, you know, you're
  1441. 51:27you're you're in for a lot of pain like
  1442. 51:29like Babbage was. And
  1443. 51:31and and so we were worried about that,
  1444. 51:33but then it did converge and then that
  1445. 51:35gave us confidence to to tackle the
  1446. 51:36harder problems. And I think if you ask
  1447. 51:38him, I think the last part of your
  1448. 51:39question was about the things in nature.
  1449. 51:41Well, the things there, what I'm
  1450. 51:42thinking is they're not suboptimal,
  1451. 51:44they're actually probably pretty optimal
  1452. 51:45because they've they've gone through an
  1453. 51:47evolutionary process, not just life with
  1454. 51:49biology, but actually geologically and
  1455. 51:51physically. You know, asteroids and
  1456. 51:54physical phenomena combined together,
  1457. 51:56they survive some amount of time because
  1458. 51:58they're stable over time, right? And if
  1459. 52:00they're stable over time, then there's
  1460. 52:02probably some structure that's
  1461. 52:03learnable. That That would be my
  1462. 52:05conjecture.
  1463. 52:07Great. There's a question down here.
  1464. 52:12What do you think about building
  1465. 52:15high-bandwidth brain-machine interfaces
  1466. 52:17and implantable memory and reasoning
  1467. 52:19modules so that humans can be further
  1468. 52:21empowered to make discoveries
  1469. 52:22autonomously as opposed to only talking
  1470. 52:24to AI in the cloud? Yeah, I I I I'm I I
  1471. 52:27love that area and I've
  1472. 52:30carefully followed it and and help
  1473. 52:31people building like EEG caps and
  1474. 52:34and things. Of course, the problem is
  1475. 52:36the resolution of these things to get
  1476. 52:38the readouts from the brain. And also,
  1477. 52:40ideally, you'd want it read and write.
  1478. 52:42But I'm very fascinated by projects like
  1479. 52:44Neuralink or, you know, chips in the
  1480. 52:46brain. Obviously, right now,
  1481. 52:48that's for,
  1482. 52:50you know, veterans and people to get
  1483. 52:51function back in their body. I think
  1484. 52:52it's going to be amazing things on that
  1485. 52:54where I think people will be able to
  1486. 52:55gain
  1487. 52:56be able to walk again if they've broken
  1488. 52:58their back and things like that. I think
  1489. 53:00there's going to be some incredible
  1490. 53:01advances in medical sciences that will
  1491. 53:02be amazing. But then beyond that, maybe
  1492. 53:06if things, you know, if it becomes
  1493. 53:07routine and it's surgery safe and
  1494. 53:09there's safe ways of doing this, I could
  1495. 53:11imagine
  1496. 53:13that would be one way for us to keep up
  1497. 53:14with the with the technology. And in
  1498. 53:16some senses, it's no different to what
  1499. 53:18we already have today with our
  1500. 53:19technologies all around us. We all have
  1501. 53:21our phones all, you know, 24/7 with us
  1502. 53:23and computers and other things. So,
  1503. 53:25we're already symbiotic almost with our
  1504. 53:27technology. It would be one step
  1505. 53:29further, of course, to have that. But
  1506. 53:30I'm not sure I mean, maybe that's for
  1507. 53:31the for the philosophers in the room to
  1508. 53:33answer what the difference what if
  1509. 53:35whether there's a hard boundary there if
  1510. 53:37if with the technology if it's attached
  1511. 53:39to you versus, you know, it's just
  1512. 53:41something you carry around with you, um,
  1513. 53:43all the time.
  1514. 53:45Great. Some question there for you.
  1515. 53:47That's one just
  1516. 53:49Sure, then.
  1517. 53:51Hi.
  1518. 53:55Are you sure? Oh.
  1519. 53:57Hello. Ah, there we go. There we go.
  1520. 54:00Ah.
  1521. 54:02What do you think about the speed at
  1522. 54:03which artificial intelligence is
  1523. 54:05developing and its effects on sort of
  1524. 54:07economic developments? There are a lot
  1525. 54:09of people out there who are deciding
  1526. 54:10careers right now who given the
  1527. 54:13sort of rapid change in the landscape,
  1528. 54:16make it really difficult to kind of
  1529. 54:17predict what they should go into.
  1530. 54:19Yeah, so it's it's it's, um, it's very
  1531. 54:21complicated one because, as you say,
  1532. 54:23things are changing at kind of lightning
  1533. 54:25speed.
  1534. 54:26Um, we were just discussing it with
  1535. 54:28Alice actually earlier. Even designing
  1536. 54:29three-year computer science courses is
  1537. 54:31quite difficult given that this the
  1538. 54:33underlying material changes in in we we
  1539. 54:35we, you know, less than three years. Um,
  1540. 54:38I think the only thing we can say for
  1541. 54:39sure is there's going to be a lot of
  1542. 54:40change, but I think that brings with it
  1543. 54:43disruption and opportunity. So, um, I
  1544. 54:46mean, I just give you an example on
  1545. 54:47coding, which I don't know if you're a
  1546. 54:48computer scientist, but you know, so I
  1547. 54:51still would recommend that you get good
  1548. 54:52at coding, uh, and math because I think
  1549. 54:55you'll be able to use these new tools in
  1550. 54:57a much deeper way if you have an
  1551. 54:58understanding of how they're built. But
  1552. 55:00on the other hand, um,
  1553. 55:02I think coding is going to be more
  1554. 55:04available to many, many more types of
  1555. 55:06people,
  1556. 55:07uh, because of the way that AI you'll be
  1557. 55:09able to program sort of a natural
  1558. 55:10language, probably, rather than, uh,
  1559. 55:13quite a complicated, uh, computer
  1560. 55:15language. And so, that will open up
  1561. 55:18field fields for creative people to, you
  1562. 55:21know, build games, make films, make
  1563. 55:23applications that, um,
  1564. 55:26maybe is more the kind of balance of
  1565. 55:28that is more on the creative side than
  1566. 55:30it is on the engineering side. But I
  1567. 55:32also think that will enhance engineers
  1568. 55:34to be able to do 10x of what they can do
  1569. 55:35today. So, I still I think it's
  1570. 55:37difficult to know, um but what I would
  1571. 55:39say is just uh to um focus on uh kind of
  1572. 55:44embracing those tools in your spare time
  1573. 55:47and and
  1574. 55:49making training yourself to be really
  1575. 55:50good at picking up new information
  1576. 55:52extremely quickly,
  1577. 55:52>> [clears throat]
  1578. 55:53>> cuz I think that's that's basically
  1579. 55:54what's going to happen in the next 10
  1580. 55:56years.
  1581. 55:57Okay. Um we've got one question just in
  1582. 55:59the here, the yellow and black top.
  1583. 56:02Um is this Yeah, do you think that there
  1584. 56:05are any biological processes or
  1585. 56:07behaviors or patterns which can't be
  1586. 56:10modeled with existing um deep learning
  1587. 56:13techniques? I'm not saying like throw
  1588. 56:16models computer to this until it works
  1589. 56:17and just make a bigger and bigger model.
  1590. 56:19Do you think there are some processes
  1591. 56:20which physically cannot be modeled with
  1592. 56:21the architecture we have?
  1593. 56:23Um I mean, there's certainly lots of
  1594. 56:25processes that can't be modeled uh today
  1595. 56:27with but but but again, this is sort of
  1596. 56:29goes back to what I said at the end of
  1597. 56:30the talk. I'm not sure in the limit uh
  1598. 56:33that there are. I think in the end uh if
  1599. 56:35physics can can sort of solve it and
  1600. 56:38there's some structure there to be
  1601. 56:39learned,
  1602. 56:40uh probably with enough examples, one
  1603. 56:42could reverse engineer a model of that.
  1604. 56:45Uh and then um I don't see any
  1605. 56:47theoretical reason why
  1606. 56:49uh a a classical system um or be a very
  1607. 56:52complex one could not uh make
  1608. 56:54predictions or or simulations of that
  1609. 56:57biological system. So, I I don't really
  1610. 56:59see
  1611. 57:00uh any in the limit uh what what the
  1612. 57:02limit of that could be. I mean, there
  1613. 57:04are lots of abstract things like
  1614. 57:06factorizing large numbers, uh
  1615. 57:08cryptography where it's sort of it's
  1616. 57:10sort of human-made systems, right? Where
  1617. 57:12there may not be any structure. I mean,
  1618. 57:13there may be structure in them in the in
  1619. 57:15the natural numbers. Lots of people
  1620. 57:17conjecture there is. If there is, then
  1621. 57:19that will also be learnable. If there
  1622. 57:20isn't and it's kind of uniform
  1623. 57:22distribution, then you would need uh
  1624. 57:24quantum computer to to you know, crack
  1625. 57:27uh cryptography, things like that. So,
  1626. 57:29those are open conjectures, but I think
  1627. 57:31most things in nature have evolved and
  1628. 57:34um over geological or biological physics
  1629. 57:37time. And so then that suggests to me
  1630. 57:40there is some structure to learn. So
  1631. 57:42that makes the the search or the
  1632. 57:44prediction potentially tractable.
  1633. 57:47Okay, and uh
  1634. 57:48last question then
  1635. 57:50we'll have
  1636. 57:51person in the pink shirt.
  1637. 57:54This question comes from the behalf of
  1638. 57:56the Cambridge University Game
  1639. 57:57Development Society.
  1640. 57:59So you mentioned
  1641. 58:00you mentioned about the Genie 2 model
  1642. 58:02and how currently it's stable for a few
  1643. 58:05seconds consistency and you hope to
  1644. 58:08eventually have that at at a few
  1645. 58:09minutes. But I suppose the question that
  1646. 58:11our society has is
  1647. 58:13games that we actually play have
  1648. 58:15consistency that's indefinite. You know,
  1649. 58:17when you're playing Minecraft you expect
  1650. 58:18to turn around and the village is still
  1651. 58:20there, right?
  1652. 58:20>> Yes. So do you see your current model
  1653. 58:23being integrated into a workflow or what
  1654. 58:26exactly how do you see AI and your model
  1655. 58:29and what you're working on?
  1656. 58:30How do you see it integrating into game
  1657. 58:32development in the coming decades?
  1658. 58:35Yeah, well look
  1659. 58:37I think there's many ways AI is going to
  1660. 58:39come in. So
  1661. 58:40one is the tools to build the assets you
  1662. 58:43need for games. So 3D models,
  1663. 58:46animations. I think that's all going to
  1664. 58:48come in in the next you know couple of
  1665. 58:50years.
  1666. 58:52I think you can think of AI as well for
  1667. 58:54game balancing. So you imagine you
  1668. 58:56design a game and overnight it could
  1669. 58:59play a million playthroughs of that game
  1670. 59:01and then in the morning you could just
  1671. 59:02get a report as a game designer of like
  1672. 59:04these things are unbalanced, right? Or
  1673. 59:06reduce the power of this unit or
  1674. 59:08whatever it is.
  1675. 59:09I think also bug testing for open world
  1676. 59:11games. So I used to make I used to make
  1677. 59:14simulation games, open world games.
  1678. 59:16They're a nightmare to bug test because
  1679. 59:18the whole point of them is that the
  1680. 59:19player can almost do almost anything and
  1681. 59:21then the game will react to you. So how
  1682. 59:22do you test you know 10 million people
  1683. 59:25having their own unique journey through
  1684. 59:26your game, well actually having AI
  1685. 59:29players play it before you release it
  1686. 59:32could help you solve a lot of those
  1687. 59:33bugs. And then finally, I mean
  1688. 59:36I think excitingly as well is like AI
  1689. 59:38characters that are much more lifelike
  1690. 59:39that move the storyline on. Used to
  1691. 59:41dream about that massive multiplayer
  1692. 59:43worlds where, you know, actually the AI
  1693. 59:45characters were intelligent and actually
  1694. 59:47updated their beliefs and
  1695. 59:49their storylines based on what the
  1696. 59:50players were doing. So it felt like a
  1697. 59:52much more
  1698. 59:53living realistic world. Um and I think
  1699. 59:56we're on the cusp of having or building
  1700. 59:58those types of of of games. And then
  1701. 1:00:01finally, the world model we're building
  1702. 1:00:02that's more about general AI and and can
  1703. 1:00:05you actually model the world It's an
  1704. 1:00:07expression of being able to understand
  1705. 1:00:08the world. Does your model understand
  1706. 1:00:10the world? Well, if it can if it can
  1707. 1:00:11generate it for some amount of time,
  1708. 1:00:14then obviously it must be in some sense
  1709. 1:00:16understanding It's a kind of empirical
  1710. 1:00:18evidence that it's understanding
  1711. 1:00:19something about the underlying physics.
  1712. 1:00:21So um so that's more for the general
  1713. 1:00:23intelligence rather than I think we're
  1714. 1:00:25going to maybe one day we'll have this
  1715. 1:00:26holodeck thing that you can just imagine
  1716. 1:00:29and it's just all there around you. Um
  1717. 1:00:31probably we will be able to have that
  1718. 1:00:33once we have AGI, but um I think that's
  1719. 1:00:35still a ways off.
  1720. 1:00:37That's great. Uh well, it seems like a
  1721. 1:00:39nice place to finish. A question on
  1722. 1:00:40games returning back to games, but uh
  1723. 1:00:42thank you all so much for coming and a
  1724. 1:00:44particular special thank you for for
  1725. 1:00:46Demis for coming in and talk to us
  1726. 1:00:47today. So thank you very
  1727. 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.