DeepMind's Demis Hassabis on the future of AI | The TED Interview

[00:00:00] TED Audio Collective Hi everyone, it's Cherelle here. We have a special

[00:00:09] treat for you today. It's an episode of the TED Interview, another podcast from the TED

[00:00:14] Audio Collective. The TED interview is back for another season. Chris Anderson returns

[00:00:19] as host to chat with great thinkers about ideas that can change the world. Today, a conversation

[00:00:25] from 2022 with AI Pioneer Demis Huzabiz. If you like it, check out the TED Interview wherever

[00:00:32] you get your podcasts.

[00:00:41] Welcome to the TED interview. I'm your host, Steven Johnson. When future tech historians

[00:00:47] look back at the first few decades of the 21st century, I suspect they will point

[00:00:53] to a day in late 2017 as one of the enduring milestones from that period. The day, the

[00:00:59] deep learning software program Alpha Zero played 44 million games of chess against a duplicate

[00:01:06] version of itself. The software had begun the day preloaded with only the basic rules of

[00:01:12] chess. Ponds can only move straight ahead unless they're capturing a piece, bishops

[00:01:17] move diagonally, you end by checkmating the king and so on. But by the end of those 44

[00:01:23] million games, which unfolded in less than a day, Alpha Zero had become arguably the most

[00:01:30] dominant chess player the world had ever seen. Alpha Zero is one of a number of pioneering

[00:01:37] AI projects created by the UK company DeepMind, founded in 2010 by one of the most fascinating

[00:01:44] minds in the digital world, Demis Hassabas. Now if you want to feel good about your own

[00:01:50] CV, I suggest you cover your ears right now because Hassabas has had a very productive

[00:01:55] career for a guy who was just in his mid 40s. As a child, he was one of the top ranked

[00:02:01] junior chess players in the world. In his mid teens, he talked his way into a job as

[00:02:06] one of the lead designers of a best-selling video game. After getting degrees in neuroscience

[00:02:11] from Cambridge and University College London, he founded DeepMind in his early 30s, selling

[00:02:17] the company to Google only four years after its founding. Now you can probably imagine that

[00:02:24] when we first started sketching out ideas for a series of interviews about the future of

[00:02:28] intelligence, Demis Hassabas was very high on the list of people we wanted to talk to.

[00:02:34] But the strange thing about DeepMind, like a lot of the AI labs at Big Tech companies

[00:02:38] right now, is that while the organization is working on some of the most revolutionary

[00:02:43] and controversial new technology out there, almost none of it is available yet for ordinary

[00:02:49] consumers to interact with. DeepMind is working on neural nets that can predict the shape

[00:02:54] of proteins, which may someday help design a drug that you might take to cure cancer or

[00:03:00] reverse Parkinson's. They're working on an AI that might be able to control nuclear fusion

[00:03:05] reactors, which could one day give us a source of renewable energy at a much lower cost.

[00:03:11] But most of these projects are still behind the curtain or accessible to a small number

[00:03:16] of outside researchers. So for the next hour we're going to ask Demis to give us a bit of a peek

[00:03:23] behind that curtain and talk about where he thinks AI is going to take us in the coming years.

[00:03:29] One of the smartest minds in the world talking about the future of intelligence. That's

[00:03:35] this week's Ted Interview.

[00:03:45] Tired of unnecessary payroll errors? Stop them in their tracks. With PayCom, employees do their

[00:03:51] own payroll. They're able to identify errors and fix them before submission right in the app

[00:03:57] because no one can afford for payroll to be wrong. Not HR and payroll teams, not leaders

[00:04:03] and definitely not employees. Shorted paychecks, time sheet corrections, uninterred sick days,

[00:04:09] missing overtime hours and expense mistakes are well a necessary for everyone. Manage the

[00:04:16] process to make Pay Day right with PayCom. Learn more at paycom.com slash sound rise. That's paycom.com slash

[00:04:25] sound rise. Support for Ted Tech comes from ODO. What is ODO? Well, ODO is an all-in-one

[00:04:32] management software with apps for every business need. ODO has apps for CRM, accounting, sales, HR,

[00:04:40] inventory, manufacturing and everything in between. And they're all in one easy to use software.

[00:04:47] And the best part about ODO? All ODO apps are integrated, helping you get things done faster

[00:04:53] and more efficiently. So when you think about business, think ODO. To learn more, visit odo.com slash

[00:05:00] Ted Tech. That's odo.com slash Ted Tech. The best place to see stars is at home with Prime

[00:05:10] video. Get everything included with Prime like Mr. and Mrs. Smith starring Donald Glover and Maya

[00:05:16] Rent or Buy hits like Mean Girls starring Renee Rap or add-on channels like Max for the HBO

[00:05:23] original Curb Your Enthusiasm with Larry David. You've never seen so many stars in one place.

[00:05:29] Prime Video Find Your Happy Place Restrictions Apply. Prime membership not required to rent or buy.

[00:05:35] Prime membership required for add-on subscriptions. See Amazon.com slash Amazon Prime for details.

[00:05:41] Demis Hassabas welcome to the Ted interview. Thanks for having me. We're really excited to have

[00:05:52] you on the program and we're going to get into some very profound questions about intelligence

[00:05:57] and machine learning in the future of health and creativity. But I wanted to start with video games

[00:06:05] which is a property for deep minds history but also in a way that perhaps some listeners don't know

[00:06:11] very appropriate for your history in that really one of the first jobs you had as a teenager was

[00:06:19] being one of the key designers of a classic 90s simulation game called Theme Park which I played

[00:06:27] back in the day and I also played Black and White which I think you maybe had a hand in as well

[00:06:31] which is a fascinating game and I guess I wanted to start there just on a kind of biographical note

[00:06:36] like how did you get the gig at Theme Park and how did how did that lead into the work you're doing

[00:06:41] with AI? Sure. I mean yeah no games is a good is a good place to start with me. I've been playing

[00:06:48] games and fascinated by games since I can remember starting with chess which I learned to play when

[00:06:53] I was four years old and then I you know that captain many England junior chess teams and actually

[00:07:00] for a while that was what I was going to potentially do be a professional chess player but actually the

[00:07:05] thing that it left on me the imprinted left for me was thinking about thinking so you know as you

[00:07:10] try and improve especially as a junior chess player trying to improve your decision making and

[00:07:14] your planning and all the things that make you good at chess and that chess teaches you including

[00:07:18] things like visualization and imagination and for me at least it made me start thinking about what was

[00:07:23] it about the brain that was coming up either these ideas sometimes mistakes and got me fascinated

[00:07:29] about the brain and neuroscience and intelligence and then I discovered computers a bit later and

[00:07:35] learn how to program and those two different loves of computers and programming and games obviously

[00:07:42] naturally came together in designing and programming video games and I was lucky enough to you know

[00:07:47] come second in some national programming competition when I was around 13 14 and and the winner got

[00:07:53] a job at was what was then the premier software house in Europe called Bullfrog Productions they made

[00:07:59] amazing games some of my favorite games like populace and so on so I've run the CEO up and said can

[00:08:06] I come for work experience and like I just call off the CEO so he was fascinated by what I was doing

[00:08:17] and then rapidly I ended up taking you know some time off between school and universities and I

[00:08:22] used that time to program theme park like you said and actually at the time in the mid 90s it was

[00:08:27] the golden era of of games design and fantastic creativity going on but also a lot of the best technology

[00:08:34] was being developed as part of games graphics technology but also AI and all the games I've written

[00:08:41] including the two you mentioned theme park and black and white have all had AI as the core gameplay

[00:08:46] component so that the game actually sort of reacts to you and the way that you play as an individual

[00:08:52] it's such an interesting history that those simulation games I think when you're dealing with

[00:08:58] you know managing resources trying to set goals for yourself trying to deal with you know multiple

[00:09:03] layers of the of the simulation you know kind of starting with sim city and then going through

[00:09:07] games like theme park and and black and white to me is not even making for many many years that

[00:09:14] those should be taught in schools I mean it's an incredibly rich way of thinking and it's very

[00:09:18] different from the kind of thinking you do when you read a novel or the kind of thinking you do

[00:09:22] when you solve a math problem but it's actually it aligns with a lot of the kind of thinking that

[00:09:27] one has to do in life probably more than some of those other fields. Yeah I totally agree and

[00:09:34] and actually I mean I think first of all chess should be taught as part of the school curriculum

[00:09:38] I think because it teaches you phenomenal skills you don't learn I think that are generalizable

[00:09:42] and transferable to other parts of life like planning and visualization and so on but also I agree

[00:09:48] with you with these types of simulation games you can call them sandboxes even so the idea is

[00:09:52] you know there's almost like a play pen for your creativity as a gamer it's very different from

[00:09:57] normal games where the game leads you by the hand through it and theme park you know the idea behind

[00:10:02] that was you designed your own Disney world and thousands of little people AI people came into your

[00:10:07] theme park and played on the rides and and depending on how well you designed there were theme park

[00:10:12] they were happy or less happy and then of course if they were happy you could charge them more in

[00:10:16] the burger stands and for the coax and and balloons and other things so the whole economics more

[00:10:20] longer there. So yeah it was it was a really interesting informative experience for me I would say

[00:10:26] not only professionally but also demonstrating to me the power of AI and in those days this was just

[00:10:32] sort of fairly sort of traditional AI right but obviously a deployed within a game of finite

[00:10:37] state machines and other things not like the kind of AI we built today but it was still amazing to

[00:10:42] me how much enjoyment people got from interacting with the game like that that had AI is cool.

[00:10:50] We're going to get into this in more detail but deep deep mind has a long history involving

[00:10:55] algorithms that has developed to play games but as far as I know none of them have been simulation

[00:11:03] games right I mean it's kind of space invaders and cuber and starcraft and things like that but

[00:11:08] but there isn't any simulated black and white players in the in in the cannon over there at

[00:11:14] deep mind is there reason for that I mean in a way it's it's it's kind of the the archetypal vision

[00:11:20] of a future AI that we have in our heads that you know we would have some artificial intelligence

[00:11:25] that will manage the city for us very effectively so that presumably is where we want to go but

[00:11:31] but you haven't done that yet right no no you're right that's an interesting observation and actually

[00:11:36] there's almost three chapters in my life of games being important to my my life in career one is

[00:11:41] the chess and the sort of my youth then there was designing and writing professional video games

[00:11:47] and then finally the third chapter of using games at deep mind from the beginning as part of

[00:11:53] the thesis of deep mind and simulations as a training ground for AI systems very convenient training

[00:11:59] down from many reasons obviously you can run millions of simulations at once in the cloud you don't

[00:12:03] have to deal with things like real robotics which is you know often you end up worrying about the

[00:12:08] hardware breaking the motors and other things so it was something that was I thought was the perfect

[00:12:13] training ground for AI systems to make quick progress and of course the other nice thing about

[00:12:18] games is you know game designers and games companies have spent thousands of person years

[00:12:23] making these things and they're challenging for human players right that was obviously

[00:12:26] that's obviously their challenging and fun for human players to play and you can kind of go up

[00:12:30] the stack of difficulty even in computer games so we started kind of famously now with Atari

[00:12:36] games you know probably the earliest computer games that sort of you know became into the mainstream

[00:12:41] from the 70s and 80s space invaders pong these classic games and that was difficult enough already

[00:12:46] for us back in 2013 2012 I remember we were we couldn't we couldn't win a point at pong

[00:12:55] and I remember for six months I think it was something like that and I remember us thinking

[00:12:59] we'll never get it we can you know it was moving the batter around but we couldn't work out for

[00:13:03] ages is it random or you know occasionally get the ball back and it couldn't win a point against

[00:13:08] the in-built obviously built in AI and we was like this is impossible because obviously it was

[00:13:12] learning just from the pixels on the screen and then finally you know it got a point we should have

[00:13:16] recorded that moment actually the first point it ever got a pong and then pretty soon after that

[00:13:21] it won a game you know to 21 points and then very soon after that it was winning 21 nil

[00:13:27] and it couldn't be beaten anymore and that was the first time we saw that kind of you know

[00:13:31] exponential improvement and we would see that many times again so of course we did that with all

[00:13:35] Atari games that was our famous first result I would say and really the birth of deep reinforcement

[00:13:41] learning our new technique that we know we've we know we've we largely pioneered and um and then we

[00:13:46] go to more complex games you know like go uh uh the most complex ball game out there and then things

[00:13:52] like Starcraft which is the most complex real-time strategy game and so we you can sort of pick

[00:13:57] games that are the sweet spot of being not too easy so it's trivial to solve them but not so hard

[00:14:03] you can't detect any progress and and I think the reason we've chosen games that have um a more

[00:14:09] competitive to begin with rather than the sandbox games is that uh it's it's it's it's better it's

[00:14:14] more convenient to have a metric that you can heal climb against right so winning a game gets the

[00:14:19] system of reward because we use reinforcement learning and uh maximizing the point score

[00:14:25] you know in something like space invaders uh so you know it very quickly you can benchmark if

[00:14:29] you're in making improvements and actually use that reward uh uh those metrics to improve your

[00:14:35] algorithms so but having said that I think we are moving now we've we basically one they're all

[00:14:41] games there are so go and Starcraft so we're actually moving more into was these freeform sandbox

[00:14:47] simulations now where the difficulty there is that the AI's in a way has to come up with its own

[00:14:52] goals right like in a Minecraft or you know like a theme park style game but that is actually where

[00:14:58] we're moving into now including building our own simulations internally yeah precisely what

[00:15:04] makes those games interesting intellectually as a human player that you that you set your own goals

[00:15:10] and you decide what kind do you want to build a you know giant you know dense urban metropolis or

[00:15:14] do you want to build a you know suburban paradise all those kind of questions you ask make them harder

[00:15:20] as a measure of progress um when you're in that training mode i a couple more things I want to

[00:15:26] ask about games but I think it's probably useful for our listeners who may not spend as much time

[00:15:31] in this space let's just define deep reinforcement learning and maybe maybe start with pong i mean

[00:15:38] I think that's a great example of the uh you know starting with a very simple task we in a game at

[00:15:44] pong which you know a six year old six year old can do it was hard initially because you started

[00:15:50] from scratch that the computer is nothing about other than the pixels and just walk us through how

[00:15:55] yes exactly so the reason that was hard for these Atari games is all we gave the system was the

[00:15:59] pixels on the screen the raw pixels values the rules or what it was controlling or or how to get

[00:16:05] points or anything like that it would it or what that was on the screen it has to kind of figure

[00:16:11] that out for itself and there are two main technologies that we combined firstly there's deep

[00:16:15] learning which is all the rage right now um and it was in very nascent when we started deep

[00:16:20] mind back in 2010 and the idea there is a sort of hierarchical neural network loosely inspired by

[00:16:26] the architecture of the brain and um and the job of that part of it is to create a model

[00:16:32] of the environment or the data stream that it finds itself in so in the case of Atari you know

[00:16:38] the Atari screen what are the things what are these pixel numbers you know and obviously there's

[00:16:43] correlations and structure in those pixels so it has to figure that out then there's the second part

[00:16:49] which is reinforcement learning which we do a lot of work on and that part is the reward

[00:16:55] maximizing or goal satisfying part of the system so you've got a model you know of the environment

[00:17:02] what do you do with that model well often if you the agent or the system finds itself in some

[00:17:06] environment it has some goal it's trying to achieve you know win a game maximize the points you

[00:17:12] know specified by the designers of the system and yet so it now has this model and it has to figure

[00:17:17] out what are the right actions to take at any moment in time that will best get it towards its

[00:17:23] overall goal and and that part is reinforcement learning and in fact we know that that's how

[00:17:29] the brain works too like uh in humans and primates it's a dopamine system in the brain that implements

[00:17:34] the form of reinforcement learning called TD learning very famous result discovered in the 90s

[00:17:40] and so we combine these two technologies together the deep learning for the modeling

[00:17:45] and then the decision making with the reinforcement learning and you know we we call that deep

[00:17:49] reinforcement learning or deep RL for short as the as the combined technology and it turns out to be

[00:17:56] extremely powerful and it's also what we used in alpha go which i'm sure we're going to talk about

[00:18:02] which was our go program um and it's you know it's very effective because effectively what you can

[00:18:07] think about is the reinforcement learning is like the planning algorithm you know it's like doing

[00:18:11] a search through all possibilities you know where there's a go game or a tary game or whatever that is

[00:18:16] but the problem is is if you just do a naive brute force search and you look at everything there's

[00:18:21] you know you're normally in space as where that's not tractable it's too bigger space the combinatorial

[00:18:26] explosions too big so what you do is you use your model to sort of imagine different paths

[00:18:31] and then the model tells you what will the environment potentially look like if you were to do that action

[00:18:36] and that helps narrow down that search space too so that so that in the end the system only looks at

[00:18:43] useful things you know much like a human chess grandmaster would do they don't look at all possibilities

[00:18:48] they just look at the few that are likely to be good ideas one of the things that i think has

[00:18:53] been so interesting about the convergence of some neuroscience and AI over the last 20 or 30 years

[00:19:00] is is our understanding of that reward mechanism the dopamine mechanism that we talked about in the

[00:19:05] brain you know i think people the popular explanation of it is the dopamine response to reward in the

[00:19:11] external world but in fact it responds to expectations about reward right you're imagining that

[00:19:17] you're going to get you know five dollars and then you get ten dollars and so there's a dopamine

[00:19:22] search because you exceeded expectations by subversa and that turned out to be relevant

[00:19:28] in the world of AI as well there's a kind of expected reward mechanism there as well right that's

[00:19:32] right that's right so it turned out that it's not important actually so much that that you're going to

[00:19:37] get the reward is actually your expectation of whether you're going to get that reward so in a way

[00:19:42] that what these reinforcement learning systems do is train your predictive capability so what's

[00:19:47] important is you know i'm predicting i'm going to get a reward and then i get one that's okay

[00:19:52] right that means my model's good but if i'm not predicting a reward and then i get a reward that's

[00:19:56] really surprising in the good direction so then i need to update my model to figure out so that next

[00:20:02] time i come across that situation it's more likely to predict the correct thing which is there's

[00:20:06] going to be a reward here and in the end a lot of intelligence is about predictive capability can i

[00:20:12] predict what is going to happen next and then use that to inform my planning you alluded to alpha

[00:20:17] go i wanted to turn a little bit to alpha zero that platform that that you developed probably

[00:20:23] the i would say the most celebrated achievement in terms of press the alpha zero success playing

[00:20:30] go and playing chess i mean i remember reading about alpha zero playing it played 44 million

[00:20:36] games of chess against itself and went from knowing nothing other than the rules of chess to being

[00:20:41] the greatest chess player that had ever lived and what's key to that approach is this adversarial

[00:20:47] model where you have two versions of the software playing against itself and and having this kind

[00:20:53] of competition where they ratchet up to to this grandmaster plus status um i guess my question is

[00:21:01] how how applicable is that adversarial model in in non-game situations are you seeing that as a

[00:21:09] strategy that you can use outside of the the game world yes so so you know without a zero and

[00:21:15] i mean it might be worth talking a little bit about their lineage from alpha go to alpha zero

[00:21:19] so the way we you know without a go what we did is set up two reinforcement learning systems to

[00:21:24] challenge each other and sort of ratchet themselves up by trying to beat each other uh and uh in

[00:21:29] we did that with go first and go only an alpha go and then we what we did with alpha zero is remove

[00:21:34] all the go specific things and made it a general games playing system that could play any two player

[00:21:39] game uh a puff you know to better than world champion level and um you know it's it's interesting

[00:21:46] actually to try and uh couch other more general things uh that are not games into this type of

[00:21:52] self-play mechanism and sometimes it can be in not just two opponents but it can also be the

[00:21:57] the system and the environment being the opponent in some sense and actually we extended in other

[00:22:03] ways with our StarCraft program which played this complex real-time strategy game StarCraft

[00:22:07] and actually there we had a league of agents so it wasn't just one versus one we actually had

[00:22:12] you know 20 or 30 in an alpha star league and um they would all be seated with different strategies

[00:22:18] and then you'd have to take kind of like a Nash equilibrium to find out which agent was the best

[00:22:24] out of that pack you're almost setting up a market dynamic in a way right and then you're allowing

[00:22:28] that to shape the agent development so we've taken that in a lot of ways we sometimes call this

[00:22:33] open ended learning where we have environments that are procedurally generated in simulation

[00:22:38] and then games are almost invented algorithmically little mini games of you know tag and hide and

[00:22:44] seek and these kinds of things and the agents have to figure it out for themselves um in that in

[00:22:50] that game and generalize from other other mazes and other other situations they've seen before

[00:22:55] so actually one thing it's worth mentioning is although we started with games of course as a

[00:22:59] convenient testing ground the ultimate aim for deep mind was then our algorithms was to build

[00:23:05] general purpose algorithms so it was always a means to an end you know to win at these games it

[00:23:11] was never an end in itself you know fascinating those things are especially to a games player like

[00:23:15] me in chess and go and they found you know all these fabulous new ideas in these games and change

[00:23:21] those games worlds but uh you know ultimately we wanted to build powerful general purpose algorithms

[00:23:26] that could be transferred to real world problems and real world domains including things like science

[00:23:31] yeah it always occurred to me that the adversarial gaming model one place where it would have an

[00:23:37] obvious parallel would be the immune system right evolving in response to new unanticipated pathogens

[00:23:45] that appears is is that something that you've done work on? We haven't done work on that but it's

[00:23:51] something on our uh to do lists so I agree with you you know immune systems microbiomes this kind of

[00:23:57] thing obviously we've been thinking a lot about that in biology space I agree with you that could

[00:24:01] be pretty interesting adversarial space also uh I think there are applications we're not really

[00:24:06] doing these ourselves probably in finance and for a financial fintech you know where actually

[00:24:12] you can think of the stock market as a huge game in some way right so almost certainly there will be

[00:24:16] applications there I'm sure other people are using our work in that domain I'm sure

[00:24:20] so I think there are quite a few natural places but there's a lot of things that can actually be

[00:24:24] recouched even scientific things into this kind of um two-and-fro set up way this ratcheting happens

[00:24:31] I could imagine a situation where one AI is the environment itself learnt from real data and then

[00:24:37] the other AI is the agent trying to achieve something in that environment and they're almost

[00:24:42] playing again with each other so what one agent is you know actually the one trying to achieve a goal

[00:24:48] and the other one is the adversary once you could argue or you know all the uh it which is

[00:24:53] the environment and but they're both AI systems so I think it's actually pretty general how it can

[00:24:58] extend

[00:25:08] if there's a surefire way to wake up feeling fresh after a night of enjoying alcohol

[00:25:12] it's with z-biotics z-biotics pre alcohol probiotic drink is the world's first genetically

[00:25:19] engineered probiotic it was invented by phg scientist to tackle rough mornings after drinking

[00:25:24] here's how it works when you drink alcohol gets converted into a toxic byproduct in the gut

[00:25:29] it's this byproduct not dehydration that's to blame for your rough next day z-biotics produces

[00:25:36] an enzyme to break this byproduct down just remember to make z-biotics your first drink of the night

[00:25:41] drink responsibly and you'll feel your best tomorrow go to zbiotics.com slash TED tech to get 15

[00:25:48] percent off your first order when you use TED tech at checkout z-biotics is backed with 100%

[00:25:54] money back guarantee so if you're unsatisfied for any reason they'll refund your money no questions

[00:25:59] asked remember to head to zbiotics.com slash TED tech and use the code TED tech at checkout for 15

[00:26:06] percent off thank you z-biotics for sponsoring this episode and our good times.

[00:26:16] when you look at the general landscape right now i mean i think as i alluded to i think a lot of

[00:26:21] us saw alpha go and alpha zero is a major milestone but there have been a couple i think in the

[00:26:27] last few years both inside of deep mind and maybe in the in the broader field are there any other

[00:26:32] kind of landmarks of the last three or four years where you said oh this is big this is this is

[00:26:38] something i didn't know we were going to be able to do so quickly yeah we were lucky you know we

[00:26:43] were very lucky to be obviously responsible for quite a few of those big moments as you say that

[00:26:47] it's hurry one first dqn at alpha go alpha zero and then more recently alpha fold but i think the

[00:26:52] one externally that that really was significant was gpt3 from open AI not so much because they invented

[00:26:59] any new technology behind it but they were the first ones to really try and go for it at scale like

[00:27:05] landlady understanding in a sort of brute force way really from the ground up no sin tactical knowledge

[00:27:11] you know basically not using any of the normal ways one would do natural language understanding

[00:27:16] and what's surprising is and you know i saw this development gpt2 which was the earlier version of

[00:27:21] that and that was not very impressive right it was sort of it was doing exactly what i expected

[00:27:27] that kind of system to do which is just be a sort of poor memorization of its training data right

[00:27:32] and and basically not when you asked it a new question it wouldn't do a very good job of giving

[00:27:36] you back irrelevant answer right it would you could sort of see it was just memorizing things and

[00:27:41] then trying to pick like the the nearest word to it and stuff like that so it was it was not

[00:27:46] very convincing and and i thought for a long time that the two problems with doing language in this

[00:27:50] way would be it's not grounded in the real world in real experience so even in simulation

[00:27:57] that's still real grounding in sensory motor experience right you're getting sensory input

[00:28:01] and and and not just linguistic input and you then can form real concepts about things you know

[00:28:07] grounded concepts let's call them all abstractions about how the world works and real models about

[00:28:11] the world physics models and other things so i you know we used to debate this a lot in within

[00:28:16] de mind but also within the ai community about what would happen if you just read Wikipedia

[00:28:22] and nothing else what would you know you know this is a classic problem in traditional AI

[00:28:27] good old fashioned ai sometimes called where there were huge projects in the 80s and 90s the first

[00:28:31] time people tried to you know solve ai at places like MIT and they used to they know i don't know

[00:28:37] if you remember this there was this huge project called psych c-y-c and dug Leonard very famous AI

[00:28:43] pioneer and what it was was was literally hand inputting into a database i don't know how many

[00:28:49] PhDs were done on this but rules of the world logical rules of like how the world works and i

[00:28:55] think that there are a million rules typed in to this database and then at some point it would

[00:29:00] sort of you would you know they i think the dream behind it was you would ask a question and somehow

[00:29:04] it would then you know maybe once you had 10 million rules in there it would be able to tell you

[00:29:08] back answers you know common sense answers and it never really worked because it's very hard

[00:29:13] to it from from various reasons very hard to encapsulate all of our knowledge in terms of rules

[00:29:17] but one of the really big problems is it wasn't grounded it was just living in the world of symbols

[00:29:21] so when you asked it about a dog you know it didn't really know a dog's got four legs and

[00:29:27] box and chases cats and sort of all of this stuff that we intuitively understand because we've

[00:29:33] interacted with dogs and that system that doesn't really know what a dog is even though it had

[00:29:38] all these logic rules about it um but what with gpt3 is is it turned out that guests going bigger

[00:29:45] didn't just incrementally improve it which i don't think would have been very interesting but sort of

[00:29:49] cross the threshold somehow and suddenly it was doing some impressive things of not just regurgitating

[00:29:56] back exactly text that it's seen but actually sort of merging and averaging in a semi-smart way

[00:30:04] different things it learned about and of course now obviously we have our own very advanced models

[00:30:09] Google meta as well as open AI so that so we've all been you know pushing these systems to the

[00:30:14] maximum and it's very interesting that at this scale some of those original assumptions that one

[00:30:19] might have about intelligence and doing it in a brain like mano and so on may not hold yeah

[00:30:25] it is such an interesting time I mean you know you probably saw that paper that Google did a while

[00:30:29] ago maybe four months ago where they had there I think it's either palm or lambda explaining jokes

[00:30:36] that it had never seen before so they give it a joke that someone made up so it had never existed

[00:30:41] before as a joke and they give them a series of these jokes and ask the the algorithm to explain

[00:30:46] why the joke was funny and you know this hasn't been duplicated you know there's been accused

[00:30:51] of cherry picking and all these questions about whether it is but the answers that they supplied

[00:30:55] in this paper are very sophisticated and it's hard not to feel it's very important here I think

[00:31:01] for the listeners to understand this when we talk about the AI being capable of understanding a joke

[00:31:08] we are not implying that the AI is sentient or conscious or having an internal experience in

[00:31:14] any way but rather that it seems to be able to represent the concepts behind the joke and what makes

[00:31:20] it funny in a way that is intelligible and that can be condensed down or shared or translated into

[00:31:25] a different metaphor different kind of explanatory model and that is something I think most people in

[00:31:31] the field was not at all clear that that was going to be something that large language models

[00:31:35] were going to be capable of yeah no I totally agree and and also agree with the point on consciousness

[00:31:41] in which we can come back to later but you know there's a really interesting question but these systems

[00:31:45] are not nowhere near you know there's no in my view there's no even semblance or hint of sentience

[00:31:51] or consciousness yet right so that I think we can put that to one side for now but certainly even

[00:31:56] understanding I don't get the feeling these systems really understand in the sense we mean you know

[00:32:02] we usually mean it what it's saying you know but despite that what is really interesting is that

[00:32:08] it can still say intelligible things somewhat useful and including things like potentially

[00:32:13] explain jokes which has always been thought of as quite a high order intelligence thing to do

[00:32:18] you know understand irony or sarcasm or something like this you know sort of like a meta level of

[00:32:24] understanding right it's pretty high high level function so and of course there's still questions about

[00:32:28] you know how well was it generalized was it really in the training data somewhere because I think

[00:32:33] one reason we did not necessarily realize these systems could do would be able to do this type

[00:32:39] of thing once they got to the right scale is I think we're way beyond the scale now where

[00:32:44] we can do a human thought as humans we can do thought experiments usefully about it right so I

[00:32:49] used to sit there and something you know a few years ago and dream about oh what if I read all

[00:32:53] of Wikipedia as an iEVA system what would I know right this kind of you know you can kind of

[00:32:58] we've all kind of spent hours on Wikipedia following links through and just enjoying reading random

[00:33:03] articles and stuff like that is kind of information on there but I think I don't think any of us

[00:33:08] with our sort of limited minds can possibly comprehend what it would be like to read the entire

[00:33:14] internet right it's just what is that even me yeah what person would we want to yeah that's a

[00:33:21] good question and you know who knows what's on there and but what would it contain we've had what

[00:33:28] 30 years now of human beings billions of us putting things on this shed you know knowledge resource

[00:33:35] the internet and and think about the number of videos there being recorded now across all devices

[00:33:41] I mean it's just mind boggling if you actually think about that and perhaps we've you know it could be

[00:33:47] that we've actually recorded every corner of the world somehow almost everything that can be done

[00:33:51] I mean it's possible it's that that sounds like it must be inculcually big but but perhaps you

[00:33:56] know it perhaps it isn't as quite as big as one might imagine and so therefore if a large model

[00:34:01] sort of ingest all that information somehow in a useful way and of course these models are very

[00:34:05] data inefficient currently right especially compared something like the brain but that can be

[00:34:09] improved too is probably then you know what actual information is is out there in fact and it

[00:34:16] might turn out the explaining jokes is possible right in my experience with large language models

[00:34:22] which has predominantly been through GPT-3 the issue that I feel like is the hardest not to crack

[00:34:29] that is still very evident there is what sometimes called the tendency of the model to hallucinate

[00:34:34] so I once asked GPT-3 to write an essay about the Belgian chemist and political philosopher Antoine

[00:34:41] the Meshelae and it delivered this beautiful Wikipedia-like entry you know five paragraphs long filled

[00:34:47] with all these details quotes of his books his biography whatever I made up this guy he doesn't exist

[00:34:52] and and the software just doesn't seem to be able to say I don't know the answer to that and

[00:34:58] it will just riff if it doesn't have something to build on and I think that my question is you know

[00:35:03] is there a way to solve that problem because that's a major reliability problem going forward I think

[00:35:08] that's going to be I mean it's a hard problem but I can see how that would be solved you know I

[00:35:12] think the model needs an estimation of its own confidence probably in an answer and if that's

[00:35:17] below some threshold it should say I don't know and at the moment I think we're not really allowing

[00:35:22] the system to do that I mean very modern systems are doing that now where I don't know or you

[00:35:27] know what is that or ask a follow-up question you know well who is he would actually be the

[00:35:31] reasonable response well they remind me of currently is you know in neuroscience I studied people

[00:35:37] with hippocampal amnesia and things like that and they have a tendency to confabulate because

[00:35:41] they don't really have memory and so these systems also are deficient in memory and I think that

[00:35:48] having an estimate about your own answers and whether they're likely to be good or not and then

[00:35:53] if you're not confident you know once you just not answer one thing that has been interesting to

[00:35:58] me someone has written about this a little bit is just how heeded the public opinions are about

[00:36:07] artificial intelligence and language models it feels to me like people have almost like strongly

[00:36:13] felt kind of political feelings about these tools and you know we don't have to get into those

[00:36:19] particular arguments but I'm just curious if that surprises you looking at the broader discussion

[00:36:25] about these things is that something you anticipated coming when you first started thinking about

[00:36:29] you know starting deep-mind or has that surprised you in some way? So it doesn't surprise me in some

[00:36:34] sense I mean the exact manifestation of it obviously is you know one maybe couldn't have predicted but

[00:36:39] we we plan from even back in 2010 when we started deep-mind from the beginning we plan for success

[00:36:44] so we always had ethics and safety as key components of what we were doing and

[00:36:51] and what we thought about and the actual you know eventual impacts of these technologies

[00:36:56] and because we believed in what we were doing we believed AI would be one of the most important if not

[00:37:01] the most important invention humanity ever makes and could be massively you know of course broadly

[00:37:07] applicable and so I think it's natural for screws to need to happen and a lot of you know arguments

[00:37:13] partly because it's very nascent technologies and you know it's still being figured out and also

[00:37:18] you know there's a lot of potential both for good and for bad in these technologies like most

[00:37:23] you know powerful new technologies and so we have to steward this correctly and be very thoughtful

[00:37:28] about it in my view I think we should be using the scientific method to do that be thoughtful and

[00:37:32] hypothesis generate and try and get better understanding of our things rather than just maybe like

[00:37:37] you know the solar silicon valley trope of move fast and break things I think we should not do that

[00:37:42] with these kinds of technologies right because breaking things in the real world could be very very

[00:37:48] damaging if the technology is very powerful potentially right it's not like a you know a game app

[00:37:53] or a you know photo app or something right and I think language specifically has been a light

[00:37:59] any road because unlike maybe games or even science which is the two things we're kind of best known

[00:38:07] for a deep mind and my personal interests those are relatively niche domains in the sense of like

[00:38:13] there were people who obsessed with those things and I think they're hugely important and enjoyable

[00:38:16] and in science I think is probably you know my view the most important thing we can do with AI

[00:38:21] but they're kind of niche as far as the mainstream public are concerned whereas language

[00:38:27] you don't have to be an AI researcher to interact with one of these systems and go wow you know

[00:38:31] what's going on here what does this mean and of course it's already interacting with some of the

[00:38:37] difficulties we are seeing with social media in general and deep fakes and all these these worries

[00:38:43] that we have already in a possible without AI but AI may end up helping or be part of the solution

[00:38:50] if used correctly right but that's that's that's that's also up for debate so I think it's been caught

[00:38:55] up in the language space with all of the wider political and cultural dynamics that we see

[00:39:02] so you alluded to alpha fold earlier let and let's turn to that because that has really interesting

[00:39:07] implications for science and for health alpha fold was on the cover of science magazine tell us first

[00:39:12] what what alpha fold is and and where you see that going so alpha fold is our system to

[00:39:20] solve what what's been called the protein folding problem and if I explain a little bit about

[00:39:25] the problem first so proteins are essential to life your genome codes for proteins each gene

[00:39:31] codes for a protein more or less and proteins are like they're sometimes called the work courses

[00:39:37] of biology basically all biological functions in your body are governed by proteins and the protein

[00:39:44] folding problem is basically this problem of can you predict from the genetic sequence called the

[00:39:50] amino acid sequence can you predict the 3d shape that that protein will fold up into when it's in the

[00:39:56] body and the reason the 3d shape is important is that the shape of a protein it often is what

[00:40:03] governs its function so if you want to understand the function of the protein what it's doing and if

[00:40:07] what how it goes wrong in disease and what drugs to target and so on you sort of need to understand

[00:40:12] the 3d shape so for 50 years people are working on this problem so it was it was first articulated by

[00:40:18] a Nobel Prize winner called Christian and Fiensten and and is part of his Nobel acceptance speech

[00:40:24] in 1972 and he said this should be in theory possible to go from the one-dimensional sequence

[00:40:31] to the three-dimensional shape and and you should better predict that computationally

[00:40:36] and the normal way it's done is painstakingly with experimental work so using massive machines cryo

[00:40:44] e cryo e m and x-ray crystallography machines and and the rule of thumb is it takes one PhD student

[00:40:52] basically their entire PhD for five years to do one protein and so in the whole history of

[00:40:57] experimental biology there's only been a hundred and fifty thousand proteins that have been the

[00:41:02] structure has been identified and what's the total range of proteins and there's more than 100

[00:41:07] million known to science right so and and you know millions added every year because because our

[00:41:14] genetic sequencing is very fast now but this this this protein structure prediction is very slow

[00:41:20] experimentally so we use that initial data hundred fifty thousand to train alpha fold which is a

[00:41:26] you know a bespoke innovative deep learning system with some special case things in it that

[00:41:31] related to biology and physics that we put into the system and it's able to predict and take

[00:41:37] an amino acid sequence and give you back the 3D structure in a matter of seconds within on average

[00:41:44] the atomic accuracy so within one angstrom error so and that is the threshold which it then

[00:41:51] becomes useful for biologists and chemists so they need it to that was always the magic threshold

[00:41:56] that had to be reached so the chemist and biologists and life scientists could rely on it for

[00:42:01] downstream tasks like drug discovery and other things without necessarily having to do the painstaking

[00:42:06] experiments so it's like there's kind of a code book where you have these sequences of amino acids

[00:42:12] and a small subset of the entire field has been translated into a three-dimensional shape

[00:42:18] and so you've given that to alpha fold and it is able to detect some kind of underlying pattern in

[00:42:26] all of those translated codes that it can then apply to novel codes that it's been given

[00:42:33] in sort of other amino acids. That's right so the system is somehow sort of understood

[00:42:37] something about the way protein physics works and they how they fold together and it can almost

[00:42:42] do a translation between the one-dimensional sequence and then eventually the three-dimensional

[00:42:47] structure so it's a pretty amazing system and what we did with it the first things we did with

[00:42:53] it is actually fold almost every protein in the human body so 20,000 proteins only 17% of them

[00:43:01] were known to science the structure and overnight we more than doubled that to high accuracy structures

[00:43:08] overnight it's a 20,000 proteins and then now we've released over a million and over the next

[00:43:13] year we plan to release all hundred million proteins known to science and then continually update

[00:43:19] the database and we teamed up with the European Bioinformatics Institute at Cambridge who hosts

[00:43:26] a lot of the biggest databases in biology in the world and the real fantastic partners to openly

[00:43:32] release all these data all these predictions all these 3D predictions for the benefit of the

[00:43:36] scientific community and actually as a sort of gift to humanity and we allowed it for any use

[00:43:41] so drug discovery company you know farmer are using it already within less than a year we released

[00:43:46] this all last summer and you know it's been cited around 3000 times already which is you know

[00:43:51] enormous number for less than a year so we think that's pretty much every biologist in the world has

[00:43:57] has looked up their proteins on this database. For a non-scientist just ordinary person walking around

[00:44:03] the world where will this matter first in terms of kind of the downstream consequences of this being

[00:44:10] released in terms of their health say. What do you think the most immediate application of this?

[00:44:15] So I think the most immediate application is something we're following up on actually is drug

[00:44:19] discovery so when you try and just you know design a new molecule or new compound and you drug

[00:44:26] basically what you're trying to do is figure out where on the surface of the protein does that

[00:44:31] molecule need to bind to to you know fix the problem in hip-hip the problem or block it and so

[00:44:38] if you now know the 3D structure and the surface then you know you know much better where you

[00:44:45] should be targeting your drug or your molecule. It's just one part of the drug discovery process

[00:44:50] but it's an important part and you know should speed up all of those processors. The other thing is

[00:44:57] what I hope is that a lot of diseases are currently thought to be to do with proteins that

[00:45:03] misfold so sometimes fold in the wrong way instead of the normal healthy way and the speculation

[00:45:09] that Alzheimer's might be because of that for example with amyloid beta protein. So again a lot

[00:45:14] of these regions of proteins are actually unstructured until they interact with something and so

[00:45:20] alpha-fold turns out to be a very good predictor of those types of disordered regions so people are

[00:45:27] ready is that you know it's the best predictor of those types of disorder regions so not only can it

[00:45:31] give you back the 3D structure it can also tell you which bits are going to kind of be unfolded

[00:45:36] unless they interact with something and some of those regions are implicated in disease. So I think

[00:45:41] those are the two most obvious near term things. I love thinking about the long term story here with

[00:45:45] drug discovery which is you know 100 years ago the state of the art was Fleming leaving the

[00:45:50] Petri dish out on his desk and you know just a random mold for how things to fall throughout

[00:45:55] the window when he goes on vacation and now we've got alpha-fold who's hopefully it is accelerating

[00:46:00] the process a little bit a little more reliable. We have a question that we ask all of our guests which

[00:46:06] is also kind of a prediction question which is really I think there are few people in the world I

[00:46:12] would like to hear the answer more from than you which is in your field what is the unsolved

[00:46:18] problem that you are most fascinated in seeing the results of or seeing the mystery solved for if

[00:46:24] you could fast forward 10 years what would the problem you'd most like to see solved be.

[00:46:29] Well the one I spend most of my time thinking about and I still think is the most fascinating

[00:46:33] outstanding problem is the notion of abstract concepts or conceptual knowledge so you know there

[00:46:41] is some evidence that these large models today have some kind of compositionality capability but

[00:46:47] it's still quite rudimentary I feel and so you know I don't think yet that this is part of

[00:46:53] understanding I would say is actually be able to abstract things and then apply those abstractions

[00:46:57] in a new situation right seamlessly and it's called transfer learning or analytical reasoning

[00:47:03] in psychology and of course humans we do this effortlessly with our brains right sort of

[00:47:09] learning in something in one domain find the underlying structure and then apply it in a new domain

[00:47:14] and so far AI systems don't really do that in a satisfactory way I would say and I think if one

[00:47:20] was to crack that then we would bridge the chasm that's still there at the moment of how do we get

[00:47:26] these learning systems which can deal with messy vision and pixels on screens and other things

[00:47:31] and find structure in that back up to symbolic manipulation so things like mathematics

[00:47:37] and maybe do mathematical discovery and things like that and I think we're still quite far from

[00:47:42] that and no one quite knows how to bridge the that chasm we have our idea we have you know half a

[00:47:47] dozen at least prototype projects in the works are on this problem but so far I would say we don't

[00:47:55] know yet how to how to solve that problem it's a bit of a mystery what these representations these

[00:48:00] conceptual representations should even look like I could talk to you for for an entire day but

[00:48:06] one last question there's a famous moment in early computing history when Charles Babbage was

[00:48:12] creating the analytic engine in the 1830s and working with Ada Lovelace arguably the world's

[00:48:18] first programmer and she wrote this extraordinary passage as part of a footnote for she predicted

[00:48:25] that in the future computers would not just be useful for math but would one day be capable of

[00:48:31] composing music and doing other creative work and so I was curious where you felt we were now

[00:48:37] and where we will be in the coming years in terms of creativity this is the way I see it right now

[00:48:42] is that I would put creativity into three buckets if we're thinking of if we define creativity

[00:48:48] is coming up with something novel or new for a purpose then you know I think what AI systems are

[00:48:54] quite good at the moment of doing is interpolation and extrapolation I would say so interpolation is

[00:48:58] sort of averaging from examples right so you give it lots of images of cats you know can you

[00:49:04] generate me a new cat yes right some kind of we're you know kind of sophisticated averaging

[00:49:10] extrapolation is more like what AlphaGo did which is play 10 million games of Go look at human games

[00:49:17] come up with a new Go strategy or chess strategy that's never seen before right and move 37 in game

[00:49:23] two of the big match we played against the world champion was lauded as a move that no human

[00:49:28] would ever thought of even though we've been playing Go for 3000 years it was played on the wrong

[00:49:32] position although professionals laughed at it and now books are being written don't books are being

[00:49:36] written about you know that move right and it's sort of gone down and go history now but what's

[00:49:43] missing is I would say true invention and you can see that because our systems like Alpha zero

[00:49:49] and AlphaGo can invent new strategies and chess and Go but they can't invent Go

[00:49:57] right so that would then be the highest level of creativity can you invent a game as great as Go

[00:50:03] or as great as chess and that they can't do and it's just a little bit of mystery what is that out

[00:50:08] of the box thinking but I think it's related to this concepts abstractions I'm mentioned earlier

[00:50:13] and and I think if we solve that one could then have systems that do what we would regard as true

[00:50:21] creativity out of the box thinking because if you imagine what sort of instruction would you want

[00:50:25] to give a large model to invent Go what what you would say is something like can you invent me a game

[00:50:33] that I can learn in five minutes but could not be mastered in many lifetimes but only takes four

[00:50:40] hours to play so it fits in my day and is beautiful aesthetically right something like that but all

[00:50:46] of those words are super high concepts I mean what would our current AI system maybe I should type

[00:50:51] that into one of my like the models and see what he would do but I'm pretty sure it wouldn't come

[00:50:55] up with Go right and so so so that that's the kind of instruction I think we'd like to give our

[00:51:00] systems and you could imagine in science I think about well okay we've done alpha fold amazing

[00:51:07] big advanced in life sciences but what would it take for a system to come up with general relativity

[00:51:12] like Einstein did okay and then really advance our knowledge of the world and physics which is

[00:51:17] ultimately what I want to do with AI actually is understand the universe around us that's the whole

[00:51:22] reason I've worked on AI my entire life um and uh that question I think is going to require true

[00:51:29] creativity which you know we're not there yet well Devon Sassavis you mentioned earlier that you

[00:51:36] were adding some things to your to-do list I can't imagine what your to-do list looks like mine is

[00:51:40] like take up the groceries in the laundry ears is a little more ambitious so we don't want to take

[00:51:44] any more your time you should go off and check more of those boxes but thank you so much for this

[00:51:48] conversation it's been a real treat thank you so much for having me the TED interview is part of

[00:51:56] the TED audio collective the show is brought to you by TED and transmitter media

[00:52:01] Sammy Kase is our story editor fact checking by Miri yeshutassen Farah Degrange is our project manager

[00:52:09] Greta Cohn is our executive producer special thanks to Michelle Quint and Anna Fielin

[00:52:15] I'm your host Steven Johnson for more information on my other projects including my latest book

[00:52:20] Extra Life you can follow me on twitter at steven b johnson or sign up for my substack newsletter

[00:52:27] adjacent possible