Demis Hassabis: DeepMind - Lex Fridman 知识图谱

Lex Fridman (01:23.080)

that's tasked with interviewing

Lex Fridman (01:25.080)

that it is, in fact, AI?

Lex Fridman (01:27.440)

Maybe we're in a kind of meta Turing test.

Demis Hassabis (01:29.680)

Probably it would be a good idea not to tell you,

Lex Fridman (01:32.440)

so it doesn't change your behavior, right?

Demis Hassabis (01:33.920)

This is a kind of link.

Lex Fridman (01:35.080)

Heisenberg uncertainty principle situation.

Demis Hassabis (01:37.100)

If I told you, you'd behave differently.

Lex Fridman (01:39.080)

Maybe that's what's happening with us, of course.

Demis Hassabis (01:40.980)

This is a benchmark from the future

Lex Fridman (01:42.800)

where they replay 2022 as a year

Demis Hassabis (01:46.560)

before AIs were good enough yet,

Lex Fridman (01:49.440)

and now we want to see, is it gonna pass?

Demis Hassabis (01:52.080)

Exactly.

Lex Fridman (01:52.920)

If I was such a program,

Lex Fridman (01:56.000)

would you be able to tell, do you think?

Lex Fridman (01:57.960)

So to the Turing test question,

Demis Hassabis (01:59.960)

you've talked about the benchmark for solving intelligence.

Lex Fridman (02:05.840)

What would be the impressive thing?

Demis Hassabis (02:07.320)

You've talked about winning a Nobel Prize

Lex Fridman (02:09.120)

and AIS system winning a Nobel Prize,

Lex Fridman (02:11.440)

but I still return to the Turing test as a compelling test,

Lex Fridman (02:14.400)

the spirit of the Turing test as a compelling test.

Demis Hassabis (02:17.280)

Yeah, the Turing test, of course,

Lex Fridman (02:18.520)

it's been unbelievably influential,

Lex Fridman (02:20.200)

and Turing's one of my all time heroes,

Lex Fridman (02:22.120)

but I think if you look back at the 1950 paper,

Demis Hassabis (02:24.920)

his original paper and read the original,

Lex Fridman (02:27.000)

you'll see, I don't think he meant it

Demis Hassabis (02:28.840)

to be a rigorous formal test.

Lex Fridman (02:30.880)

I think it was more like a thought experiment,

Demis Hassabis (02:32.880)

almost a bit of philosophy he was writing

Lex Fridman (02:34.640)

if you look at the style of the paper,

Lex Fridman (02:36.440)

and you can see he didn't specify it very rigorously.

Lex Fridman (02:38.640)

So for example, he didn't specify the knowledge

Demis Hassabis (02:41.840)

that the expert or judge would have.

Lex Fridman (02:45.440)

How much time would they have to investigate this?

Lex Fridman (02:48.320)

So these are important parameters

Lex Fridman (02:49.500)

if you were gonna make it a true sort of formal test.

Lex Fridman (02:54.320)

And by some measures, people claim the Turing test passed

Lex Fridman (02:58.400)

several, a decade ago, I remember someone claiming that

Demis Hassabis (03:00.920)

with a kind of very bog standard, normal logic model,

Lex Fridman (03:06.000)

because they pretended it was a kid.

Lex Fridman (03:08.440)

So the judges thought that the machine was a child.

Lex Fridman (03:13.280)

So that would be very different

Demis Hassabis (03:15.360)

from an expert AI person interrogating a machine

Lex Fridman (03:18.680)

and knowing how it was built and so on.

Lex Fridman (03:20.720)

So I think we should probably move away from that

Lex Fridman (03:24.600)

as a formal test and move more towards a general test

Demis Hassabis (03:28.800)

where we test the AI capabilities on a range of tasks

Lex Fridman (03:32.040)

and see if it reaches human level or above performance

Demis Hassabis (03:35.360)

on maybe thousands, perhaps even millions of tasks

Lex Fridman (03:38.400)

eventually and cover the entire sort of cognitive space.

Lex Fridman (03:41.940)

So I think for its time,

Lex Fridman (03:44.140)

it was an amazing thought experiment.

Lex Fridman (03:45.520)

And also 1950s, obviously there's barely the dawn

Lex Fridman (03:48.240)

of the computer age.

Lex Fridman (03:49.440)

So of course he only thought about text

Lex Fridman (03:51.480)

and now we have a lot more different inputs.

Lex Fridman (03:54.580)

So yeah, maybe the better thing to test

Lex Fridman (03:57.080)

is the generalizability, so across multiple tasks.

Lex Fridman (03:59.660)

But I think it's also possible as systems like Gato show

Lex Fridman (04:04.540)

that eventually that might map right back to language.

Lex Fridman (04:08.320)

So you might be able to demonstrate your ability

Lex Fridman (04:10.800)

to generalize across tasks by then communicating

Demis Hassabis (04:14.760)

your ability to generalize across tasks,

Lex Fridman (04:17.080)

which is kind of what we do through conversation anyway

Demis Hassabis (04:19.200)

when we jump around.

Lex Fridman (04:20.800)

Ultimately what's in there in that conversation

Demis Hassabis (04:23.720)

is not just you moving around knowledge,

Lex Fridman (04:27.000)

it's you moving around like these entirely different

Demis Hassabis (04:30.360)

modalities of understanding that ultimately map

Lex Fridman (04:34.920)

to your ability to operate successfully

Demis Hassabis (04:38.920)

in all of these domains, which you can think of as tasks.

Lex Fridman (04:42.600)

Yeah, I think certainly we as humans use language

Demis Hassabis (04:45.600)

as our main generalization communication tool.

Lex Fridman (04:48.440)

So I think we end up thinking in language

Lex Fridman (04:51.280)

and expressing our solutions in language.

Lex Fridman (04:54.440)

So it's going to be a very powerful mode in which

Demis Hassabis (04:58.320)

to explain the system, to explain what it's doing.

Lex Fridman (05:03.080)

But I don't think it's the only modality that matters.

Lex Fridman (05:07.600)

So I think there's going to be a lot of different ways

Lex Fridman (05:10.920)

to express capabilities other than just language.

Demis Hassabis (05:15.600)

Yeah, visual, robotics, body language,

Lex Fridman (05:21.120)

yeah, actions, the interactive aspect of all that.

Demis Hassabis (05:23.760)

That's all part of it.

Lex Fridman (05:24.760)

But what's interesting with Gato is that

Demis Hassabis (05:27.600)

it's sort of pushing prediction to the maximum

Lex Fridman (05:30.200)

in terms of like mapping arbitrary sequences

Demis Hassabis (05:33.400)

to other sequences and sort of just predicting

Lex Fridman (05:35.480)

what's going to happen next.

Lex Fridman (05:36.400)

So prediction seems to be fundamental to intelligence.

Lex Fridman (05:41.040)

And what you're predicting doesn't so much matter.

Demis Hassabis (05:44.160)

Yeah, it seems like you can generalize that quite well.

Lex Fridman (05:46.840)

So obviously language models predict the next word,

Demis Hassabis (05:49.640)

Gato predicts potentially any action or any token.

Lex Fridman (05:53.880)

And it's just the beginning really.

Demis Hassabis (05:55.320)

It's our most general agent one could call it so far,

Lex Fridman (05:58.120)

but that itself can be scaled up massively more

Demis Hassabis (06:01.240)

than we've done so far.

Lex Fridman (06:02.120)

And obviously we're in the middle of doing that.

Lex Fridman (06:04.280)

But the big part of solving AGI is creating benchmarks

Lex Fridman (06:08.240)

that help us get closer and closer,

Demis Hassabis (06:11.080)

sort of creating benchmarks that test the generalizability.

Lex Fridman (06:14.920)

And it's just still interesting that this fella,

Demis Hassabis (06:17.440)

Alan Turing, was one of the first

Lex Fridman (06:20.520)

and probably still one of the only people

Demis Hassabis (06:22.600)

that was trying, maybe philosophically,

Lex Fridman (06:25.040)

but was trying to formulate a benchmark

Demis Hassabis (06:26.840)

that could be followed.

Lex Fridman (06:27.880)

It is, even though it's fuzzy,

Demis Hassabis (06:30.960)

it's still sufficiently rigorous

Lex Fridman (06:32.520)

to where you can run that test.

Lex Fridman (06:33.960)

And I still think something like the Turing test

Lex Fridman (06:36.640)

will, at the end of the day,

Demis Hassabis (06:38.720)

be the thing that truly impresses other humans

Lex Fridman (06:42.560)

so that you can have a close friend who's an AI system.

Lex Fridman (06:46.400)

And for that friend to be a good friend,

Lex Fridman (06:48.320)

they're going to have to be able to play StarCraft

Lex Fridman (06:53.160)

and they're gonna have to do all of these tasks,

Lex Fridman (06:56.600)

get you a beer, so the robotics tasks,

Demis Hassabis (07:00.440)

play games with you, use language,

Lex Fridman (07:03.160)

humor, all of those kinds of things.

Lex Fridman (07:04.800)

But that ultimately can boil down to language.

Lex Fridman (07:08.000)

It feels like, not in terms of the AI community,

Lex Fridman (07:11.200)

but in terms of the actual impact

Lex Fridman (07:13.120)

of general intelligence on the world,

Demis Hassabis (07:14.760)

it feels like language will be the place

Lex Fridman (07:16.640)

where it truly shines.

Demis Hassabis (07:18.480)

I think so, because it's such an important

Lex Fridman (07:20.640)

kind of input output for us.

Demis Hassabis (07:22.480)

I think you're right.

Lex Fridman (07:23.320)

I think the Turing test,

Demis Hassabis (07:24.680)

the kind of the philosophy behind it,

Lex Fridman (07:27.440)

which is the idea of can a machine mimic

Lex Fridman (07:31.120)

the behaviors of a human?

Lex Fridman (07:34.960)

And I would say wider than just language and text,

Demis Hassabis (07:38.680)

then in terms of actions and everything else,

Lex Fridman (07:41.320)

creativity, all these things,

Demis Hassabis (07:42.680)

then if it can sort of match or exceed

Lex Fridman (07:45.320)

human cognitive capabilities,

Demis Hassabis (07:47.320)

then I think we have a true intelligence.

Lex Fridman (07:49.880)

So from that perspective, you're right.

Demis Hassabis (07:51.640)

I think he did formulate the right kind of setup.

Lex Fridman (07:55.960)

I just, I think there'll be a kind of humor

Demis Hassabis (07:59.280)

in the AI systems of the future looking back

Lex Fridman (08:01.440)

to this conversation and thinking about the Turing test

Lex Fridman (08:06.040)

and also thinking about by that time,

Lex Fridman (08:09.360)

they would know which year they were finally able

Demis Hassabis (08:11.880)

to sort of cross the threshold of human level intelligence

Lex Fridman (08:15.560)

and think how funny it is that we humans

Demis Hassabis (08:18.800)

were still confused about this whole problem

Lex Fridman (08:22.400)

that has been long ago been solved.

Demis Hassabis (08:24.720)

Well, hello to the AI agents of the future.

Lex Fridman (08:27.240)

Anyway, so going back to your journey,

Lex Fridman (08:30.320)

when did you fall in love with programming first?

Lex Fridman (08:33.800)

Well, I was pretty young age actually.

Demis Hassabis (08:35.960)

So, I started off, actually games was my first love.

Lex Fridman (08:40.840)

So starting to play chess when I was around four years old

Lex Fridman (08:43.680)

and then it was actually with winnings

Lex Fridman (08:46.160)

from a chess competition that I managed

Demis Hassabis (08:48.400)

to buy my first chess computer

Lex Fridman (08:49.800)

when I was about eight years old.

Demis Hassabis (08:50.840)

It was a ZX Spectrum, which was hugely popular

Lex Fridman (08:53.160)

in the UK at the time.

Lex Fridman (08:54.720)

And it was amazing machine because I think it trained

Lex Fridman (08:58.520)

a whole generation of programmers in the UK

Demis Hassabis (09:00.440)

because it was so accessible.

Lex Fridman (09:02.520)

You know, you literally switched it on

Lex Fridman (09:03.800)

and there was the basic prompt

Lex Fridman (09:05.120)

and you could just get going.

Lex Fridman (09:06.680)

And my parents didn't really know anything about computers.

Lex Fridman (09:09.920)

So, but because it was my money from a chess competition,

Demis Hassabis (09:12.600)

I could say I wanted to buy it.

Lex Fridman (09:15.760)

And then, you know, I just went to bookstores,

Demis Hassabis (09:17.960)

got books on programming and started typing in,

Lex Fridman (09:22.040)

you know, the programming code.

Lex Fridman (09:23.520)

And then of course, once you start doing that,

Lex Fridman (09:26.480)

you start adjusting it and then making your own games.

Lex Fridman (09:29.120)

And that's when I fell in love with computers

Lex Fridman (09:30.840)

and realized that they were a very magical device.

Demis Hassabis (09:34.840)

In a way, I kind of, I wouldn't have been able

Lex Fridman (09:36.440)

to explain this at the time,

Lex Fridman (09:37.440)

but I felt that they were sort of almost

Lex Fridman (09:38.960)

a magical extension of your mind.

Demis Hassabis (09:40.920)

I always had this feeling and I've always loved this

Lex Fridman (09:43.080)

about computers that you can set them off doing something,

Demis Hassabis (09:46.160)

some task for you, you can go to sleep,

Lex Fridman (09:48.520)

come back the next day and it's solved.

Demis Hassabis (09:51.240)

You know, that feels magical to me.

Lex Fridman (09:53.080)

So, I mean, all machines do that to some extent.

Demis Hassabis (09:55.280)

They all enhance our natural capabilities.

Lex Fridman (09:57.640)

Obviously cars make us, allow us to move faster

Demis Hassabis (10:00.080)

than we can run, but this was a machine to extend the mind.

Lex Fridman (10:04.520)

And then of course, AI is the ultimate expression

Demis Hassabis (10:08.480)

of what a machine may be able to do or learn.

Lex Fridman (10:11.360)

So very naturally for me, that thought extended

Demis Hassabis (10:14.440)

into AI quite quickly.

Lex Fridman (10:16.040)

Do you remember the programming language

Lex Fridman (10:18.560)

that was first started and was it special to the machine?

Lex Fridman (10:22.080)

No, I think it was just basic on the ZX Spectrum.

Demis Hassabis (10:25.880)

I don't know what specific form it was.

Lex Fridman (10:27.480)

And then later on I got a Commodore Amiga,

Demis Hassabis (10:29.600)

which was a fantastic machine.

Lex Fridman (10:32.760)

Now you're just showing off.

Lex Fridman (10:33.800)

So yeah, well, lots of my friends had Atari STs

Lex Fridman (10:36.520)

and I managed to get Amigas, it was a bit more powerful

Lex Fridman (10:38.880)

and that was incredible and used to do programming

Lex Fridman (10:42.800)

in assembler and also Amos basic,

Demis Hassabis (10:46.240)

this specific form of basic, it was incredible actually.

Lex Fridman (10:49.280)

So I learned all my coding skills.

Lex Fridman (10:51.000)

And when did you fall in love with AI?

Lex Fridman (10:53.040)

So when did you first start to gain an understanding

Demis Hassabis (10:56.880)

that you can not just write programs

Lex Fridman (10:58.880)

that do some mathematical operations for you

Demis Hassabis (11:01.640)

while you sleep, but something that's akin

Lex Fridman (11:05.400)

to bringing an entity to life,

Demis Hassabis (11:08.800)

sort of a thing that can figure out something

Lex Fridman (11:11.840)

more complicated than a simple mathematical operation.

Demis Hassabis (11:15.920)

Yeah, so there was a few stages for me

Lex Fridman (11:17.600)

all while I was very young.

Lex Fridman (11:18.920)

So first of all, as I was trying to improve

Lex Fridman (11:21.680)

at playing chess, I was captaining

Demis Hassabis (11:23.160)

various England junior chess teams.

Lex Fridman (11:24.680)

And at the time when I was about maybe 10, 11 years old,

Demis Hassabis (11:27.480)

I was gonna become a professional chess player.

Lex Fridman (11:29.360)

That was my first thought.

Lex Fridman (11:32.000)

So that dream was there to try to get

Lex Fridman (11:34.680)

to the highest levels of chess.

Demis Hassabis (11:35.520)

Yeah, so when I was about 12 years old,

Lex Fridman (11:39.160)

I got to master standard and I was second highest rated

Demis Hassabis (11:41.320)

player in the world to Judith Polgar,

Lex Fridman (11:42.680)

who obviously ended up being an amazing chess player

Lex Fridman (11:45.760)

and a world women's champion.

Lex Fridman (11:48.560)

And when I was trying to improve at chess,

Demis Hassabis (11:50.760)

where what you do is you obviously, first of all,

Lex Fridman (11:52.760)

you're trying to improve your own thinking processes.

Lex Fridman (11:55.120)

So that leads you to thinking about thinking,

Lex Fridman (11:58.080)

how is your brain coming up with these ideas?

Lex Fridman (12:00.400)

Why is it making mistakes?

Lex Fridman (12:01.880)

How can you improve that thought process?

Lex Fridman (12:04.520)

But the second thing is that you,

Lex Fridman (12:06.360)

it was just the beginning, this was like in the early 80s,

Demis Hassabis (12:09.640)

mid 80s of chess computers.

Lex Fridman (12:11.240)

If you remember, they were physical balls

Demis Hassabis (12:12.760)

like the one we have in front of us.

Lex Fridman (12:14.000)

And you press down the squares.

Lex Fridman (12:17.000)

And I think Kasparov had a branded version of it

Lex Fridman (12:19.600)

that I got.

Lex Fridman (12:21.000)

And you used to, they're not as strong as they are today,

Lex Fridman (12:24.640)

but they were pretty strong and you used to practice

Demis Hassabis (12:27.320)

against them to try and improve your openings

Lex Fridman (12:30.560)

and other things.

Lex Fridman (12:31.480)

And so I remember, I think I probably got my first one,

Lex Fridman (12:33.440)

I was around 11 or 12.

Lex Fridman (12:34.920)

And I remember thinking, this is amazing,

Lex Fridman (12:37.760)

how has someone programmed this chess board to play chess?

Lex Fridman (12:42.920)

And it was very formative book I bought,

Lex Fridman (12:45.600)

which was called The Chess Computer Handbook

Demis Hassabis (12:47.760)

by David Levy.

Lex Fridman (12:49.000)

This thing came out in 1984 or something.

Lex Fridman (12:50.680)

So I must've got it when I was about 11, 12.

Lex Fridman (12:52.360)

And it explained fully how these chess programs were made.

Lex Fridman (12:56.120)

And I remember my first AI program

Lex Fridman (12:57.680)

being programming my Amiga.

Demis Hassabis (13:00.440)

It couldn't, it wasn't powerful enough to play chess.

Lex Fridman (13:02.920)

I couldn't write a whole chess program,

Lex Fridman (13:04.200)

but I wrote a program for it to play Othello or reverse it,

Lex Fridman (13:07.480)

sometimes called I think in the US.

Lex Fridman (13:09.320)

And so a slightly simpler game than chess,

Lex Fridman (13:11.760)

but I used all of the principles that chess programs had,

Demis Hassabis (13:14.360)

alpha, beta, search, all of that.

Lex Fridman (13:16.000)

And that was my first AI program.

Demis Hassabis (13:17.440)

I remember that very well, I was around 12 years old.

Lex Fridman (13:19.440)

So that brought me into AI.

Lex Fridman (13:21.680)

And then the second part was later on,

Lex Fridman (13:24.120)

when I was around 16, 17,

Lex Fridman (13:25.560)

and I was writing games professionally, designing games,

Lex Fridman (13:28.840)

writing a game called Theme Park,

Demis Hassabis (13:30.640)

which had AI as a core gameplay component

Lex Fridman (13:34.040)

as part of the simulation.

Lex Fridman (13:35.680)

And it sold millions of copies around the world.

Lex Fridman (13:38.480)

And people loved the way that the AI,

Demis Hassabis (13:41.000)

even though it was relatively simple

Lex Fridman (13:42.280)

by today's AI standards,

Demis Hassabis (13:44.440)

was reacting to the way you as the player played it.

Lex Fridman (13:47.720)

So it was called a sandbox game.

Lex Fridman (13:49.240)

So it was one of the first types of games like that,

Lex Fridman (13:51.320)

along with SimCity.

Lex Fridman (13:52.680)

And it meant that every game you played was unique.

Lex Fridman (13:55.720)

Is there something you could say just on a small tangent

Demis Hassabis (13:58.840)

about really impressive AI

Lex Fridman (14:02.160)

from a game design, human enjoyment perspective,

Demis Hassabis (14:06.560)

really impressive AI that you've seen in games

Lex Fridman (14:09.680)

and maybe what does it take to create an AI system?

Lex Fridman (14:12.480)

And how hard of a problem is that?

Lex Fridman (14:14.240)

So a million questions just as a brief tangent.

Demis Hassabis (14:18.360)

Well, look, I think games have been significant in my life

Lex Fridman (14:23.000)

for three reasons.

Lex Fridman (14:23.840)

So first of all, I was playing them

Lex Fridman (14:26.080)

and training myself on games when I was a kid.

Demis Hassabis (14:28.800)

Then I went through a phase of designing games

Lex Fridman (14:31.480)

and writing AI for games.

Lex Fridman (14:33.000)

So all the games I professionally wrote

Lex Fridman (14:35.960)

had AI as a core component.

Lex Fridman (14:37.680)

And that was mostly in the 90s.

Lex Fridman (14:40.040)

And the reason I was doing that in games industry

Demis Hassabis (14:42.960)

was at the time the games industry,

Lex Fridman (14:45.080)

I think was the cutting edge of technology.

Lex Fridman (14:47.160)

So whether it was graphics with people like John Carmack

Lex Fridman (14:49.800)

and Quake and those kinds of things or AI,

Demis Hassabis (14:53.040)

I think actually all the action was going on in games.

Lex Fridman (14:56.160)

And we're still reaping the benefits of that

Demis Hassabis (14:58.440)

even with things like GPUs, which I find ironic

Lex Fridman (15:01.480)

was obviously invented for graphics, computer graphics,

Lex Fridman (15:03.680)

but then turns out to be amazingly useful for AI.

Lex Fridman (15:06.320)

It just turns out everything's a matrix multiplication

Demis Hassabis (15:08.760)

it appears in the whole world.

Lex Fridman (15:11.080)

So I think games at the time had the most cutting edge AI.

Lex Fridman (15:15.800)

And a lot of the games, I was involved in writing.

Lex Fridman (15:19.800)

So there was a game called Black and White,

Demis Hassabis (15:21.280)

which was one game I was involved with

Lex Fridman (15:22.760)

in the early stages of,

Demis Hassabis (15:24.000)

which I still think is the most impressive example

Lex Fridman (15:28.400)

of reinforcement learning in a computer game.

Lex Fridman (15:30.560)

So in that game, you trained a little pet animal and...

Lex Fridman (15:34.600)

It's a brilliant game.

Lex Fridman (15:35.440)

And it sort of learned from how you were treating it.

Lex Fridman (15:37.640)

So if you treated it badly, then it became mean.

Lex Fridman (15:40.680)

And then it would be mean to your villagers

Lex Fridman (15:42.920)

and your population, the sort of the little tribe

Demis Hassabis (15:45.760)

that you were running.

Lex Fridman (15:47.240)

But if you were kind to it, then it would be kind.

Lex Fridman (15:49.400)

And people were fascinated by how that works.

Lex Fridman (15:51.080)

And so was I to be honest with the way it kind of developed.

Demis Hassabis (15:54.160)

And...

Lex Fridman (15:55.120)

Especially the mapping to good and evil.

Demis Hassabis (15:57.240)

Yeah.

Lex Fridman (15:58.080)

Made you realize, made me realize that you can sort of

Demis Hassabis (16:01.640)

in the choices you make can define where you end up.

Lex Fridman (16:07.440)

And that means all of us are capable of the good, evil.

Demis Hassabis (16:12.640)

It all matters in the different choices

Lex Fridman (16:15.240)

along the trajectory to those places that you make.

Demis Hassabis (16:18.240)

It's fascinating.

Lex Fridman (16:19.080)

I mean, games can do that philosophically to you.

Lex Fridman (16:21.360)

And it's rare.

Lex Fridman (16:22.200)

It seems rare.

Demis Hassabis (16:23.040)

Yeah.

Lex Fridman (16:23.880)

Well, games are, I think, a unique medium

Demis Hassabis (16:24.720)

because you as the player,

Lex Fridman (16:26.600)

you're not just passively consuming the entertainment,

Lex Fridman (16:30.080)

right?

Lex Fridman (16:30.920)

You're actually actively involved as an agent.

Lex Fridman (16:34.280)

So I think that's what makes it in some ways

Lex Fridman (16:36.160)

can be more visceral than other mediums

Demis Hassabis (16:38.400)

like films and books.

Lex Fridman (16:40.000)

So the second, so that was designing AI in games.

Lex Fridman (16:42.640)

And then the third use we've used of AI

Lex Fridman (16:46.440)

is in DeepMind from the beginning,

Demis Hassabis (16:48.440)

which is using games as a testing ground

Lex Fridman (16:50.920)

for proving out AI algorithms and developing AI algorithms.

Lex Fridman (16:55.000)

And that was a sort of a core component

Lex Fridman (16:58.480)

of our vision at the start of DeepMind

Demis Hassabis (17:00.320)

was that we would use games very heavily

Lex Fridman (17:03.200)

as our main testing ground, certainly to begin with,

Demis Hassabis (17:06.360)

because it's super efficient to use games.

Lex Fridman (17:08.560)

And also, it's very easy to have metrics

Demis Hassabis (17:11.480)

to see how well your systems are improving

Lex Fridman (17:14.080)

and what direction your ideas are going in

Lex Fridman (17:15.840)

and whether you're making incremental improvements.

Lex Fridman (17:18.400)

And because those games are often rooted

Demis Hassabis (17:20.400)

in something that humans did for a long time beforehand,

Lex Fridman (17:23.360)

there's already a strong set of rules.

Demis Hassabis (17:26.520)

Like it's already a damn good benchmark.

Lex Fridman (17:28.240)

Yes, it's really good for so many reasons

Demis Hassabis (17:30.200)

because you've got clear measures

Lex Fridman (17:32.800)

of how good humans can be at these things.

Lex Fridman (17:35.520)

And in some cases like Go,

Lex Fridman (17:36.840)

we've been playing it for thousands of years

Lex Fridman (17:39.720)

and often they have scores or at least win conditions.

Lex Fridman (17:43.280)

So it's very easy for reward learning systems

Demis Hassabis (17:45.640)

to get a reward.

Lex Fridman (17:46.480)

It's very easy to specify what that reward is.

Lex Fridman (17:49.320)

And also at the end, it's easy to test externally

Lex Fridman (17:54.320)

at how strong is your system by of course,

Demis Hassabis (17:56.920)

playing against the world's strongest players at those games.

Lex Fridman (18:00.240)

So it's so good for so many reasons

Lex Fridman (18:02.680)

and it's also very efficient to run potentially millions

Lex Fridman (18:05.520)

of simulations in parallel on the cloud.

Lex Fridman (18:08.280)

So I think there's a huge reason why we were so successful

Lex Fridman (18:12.800)

back in starting out 2010,

Lex Fridman (18:14.760)

how come we were able to progress so quickly

Lex Fridman (18:16.680)

because we've utilized games.

Lex Fridman (18:18.880)

And at the beginning of DeepMind,

Lex Fridman (18:21.320)

we also hired some amazing game engineers

Demis Hassabis (18:24.600)

who I knew from my previous lives in the games industry.

Lex Fridman (18:28.000)

And that helped to bootstrap us very quickly.

Lex Fridman (18:30.920)

And plus it's somehow super compelling

Lex Fridman (18:33.880)

almost at a philosophical level of man versus machine

Demis Hassabis (18:38.080)

over a chess board or a Go board.

Lex Fridman (18:41.200)

And especially given that the entire history of AI

Demis Hassabis (18:43.600)

is defined by people saying it's gonna be impossible

Lex Fridman (18:45.960)

to make a machine that beats a human being in chess.

Lex Fridman (18:50.960)

And then once that happened,

Lex Fridman (18:53.200)

people were certain when I was coming up in AI

Demis Hassabis (18:55.880)

that Go is not a game that can be solved

Lex Fridman (18:58.760)

because of the combinatorial complexity is just too,

Demis Hassabis (19:02.000)

it's no matter how much Moore's law you have,

Lex Fridman (19:06.640)

compute is just never going to be able

Demis Hassabis (19:08.560)

to crack the game of Go.

Lex Fridman (19:10.200)

And so then there's something compelling about facing,

Demis Hassabis (19:14.920)

sort of taking on the impossibility of that task

Lex Fridman (19:18.160)

from the AI researcher perspective,

Demis Hassabis (19:22.480)

engineer perspective, and then as a human being,

Lex Fridman (19:24.520)

just observing this whole thing.

Demis Hassabis (19:27.040)

Your beliefs about what you thought was impossible

Lex Fridman (19:32.520)

being broken apart,

Demis Hassabis (19:35.920)

it's humbling to realize we're not as smart as we thought.

Lex Fridman (19:40.480)

It's humbling to realize that the things we think

Demis Hassabis (19:43.160)

are impossible now perhaps will be done in the future.

Lex Fridman (19:47.000)

There's something really powerful about a game,

Demis Hassabis (19:50.800)

AI system beating human being in a game

Lex Fridman (19:52.880)

that drives that message home

Demis Hassabis (19:55.680)

for like millions, billions of people,

Lex Fridman (19:58.000)

especially in the case of Go.

Demis Hassabis (19:59.320)

Sure.

Lex Fridman (1:00:00.200)

amazing as it is, is just the beginning.

Lex Fridman (1:00:03.000)

And I hope it's evidence of what could be done

Lex Fridman (1:00:07.160)

with computational methods.

Lex Fridman (1:00:08.800)

So AlphaFold solved this huge problem

Lex Fridman (1:00:12.200)

of the structure of proteins, but biology is dynamic.

Lex Fridman (1:00:15.240)

So really what I imagine from here,

Lex Fridman (1:00:16.880)

and we're working on all these things now,

Demis Hassabis (1:00:18.640)

is protein, protein interaction, protein ligand binding,

Lex Fridman (1:00:23.160)

so reacting with molecules,

Demis Hassabis (1:00:25.400)

then you wanna build up to pathways,

Lex Fridman (1:00:27.640)

and then eventually a virtual cell.

Demis Hassabis (1:00:30.000)

That's my dream, maybe in the next 10 years.

Lex Fridman (1:00:32.680)

And I've been talking actually

Demis Hassabis (1:00:33.600)

to a lot of biologists, friends of mine,

Lex Fridman (1:00:35.000)

Paul Nurse, who runs the Crick Institute,

Demis Hassabis (1:00:36.760)

amazing biologists, Nobel Prize winning biologists.

Lex Fridman (1:00:39.080)

We've been discussing for 20 years now, virtual cells.

Lex Fridman (1:00:42.080)

Could you build a virtual simulation of a cell?

Lex Fridman (1:00:44.720)

And if you could, that would be incredible

Demis Hassabis (1:00:46.240)

for biology and disease discovery,

Lex Fridman (1:00:48.120)

because you could do loads of experiments

Demis Hassabis (1:00:49.520)

on the virtual cell, and then only at the last stage,

Lex Fridman (1:00:52.400)

validate it in the wet lab.

Lex Fridman (1:00:53.920)

So you could, in terms of the search space

Lex Fridman (1:00:56.400)

of discovering new drugs, it takes 10 years roughly

Demis Hassabis (1:00:59.200)

to go from identifying a target,

Lex Fridman (1:01:03.360)

to having a drug candidate.

Demis Hassabis (1:01:06.480)

Maybe that could be shortened by an order of magnitude,

Lex Fridman (1:01:09.760)

if you could do most of that work in silico.

Lex Fridman (1:01:13.120)

So in order to get to a virtual cell,

Lex Fridman (1:01:15.720)

we have to build up understanding

Demis Hassabis (1:01:18.320)

of different parts of biology and the interactions.

Lex Fridman (1:01:20.760)

And so every few years we talk about this,

Demis Hassabis (1:01:24.560)

I talked about this with Paul.

Lex Fridman (1:01:25.600)

And then finally, last year after AlphaFold,

Demis Hassabis (1:01:27.840)

I said, now's the time we can finally go for it.

Lex Fridman (1:01:30.600)

And AlphaFold is the first proof point

Demis Hassabis (1:01:32.360)

that this might be possible.

Lex Fridman (1:01:33.800)

And he's very excited, and we have some collaborations

Demis Hassabis (1:01:35.920)

with his lab, they're just across the road actually

Lex Fridman (1:01:38.480)

from us, it's wonderful being here in King's Cross

Demis Hassabis (1:01:40.960)

with the Crick Institute across the road.

Lex Fridman (1:01:42.880)

And I think the next steps,

Demis Hassabis (1:01:45.960)

I think there's gonna be some amazing advances

Lex Fridman (1:01:48.040)

in biology built on top of things like AlphaFold.

Demis Hassabis (1:01:50.960)

We're already seeing that with the community doing that

Lex Fridman (1:01:53.160)

after we've open sourced it and released it.

Lex Fridman (1:01:56.000)

And I often say that I think if you think of mathematics

Lex Fridman (1:02:02.360)

is the perfect description language for physics,

Demis Hassabis (1:02:05.080)

I think AI might be end up being

Lex Fridman (1:02:06.920)

the perfect description language for biology

Demis Hassabis (1:02:09.280)

because biology is so messy, it's so emergent,

Lex Fridman (1:02:13.040)

so dynamic and complex.

Demis Hassabis (1:02:15.320)

I think I find it very hard to believe

Lex Fridman (1:02:16.920)

we'll ever get to something as elegant

Lex Fridman (1:02:18.600)

as Newton's laws of motions to describe a cell, right?

Lex Fridman (1:02:21.760)

It's just too complicated.

Lex Fridman (1:02:23.600)

So I think AI is the right tool for that.

Lex Fridman (1:02:26.160)

So you have to start at the basic building blocks

Lex Fridman (1:02:29.480)

and use AI to run the simulation

Lex Fridman (1:02:31.680)

for all those building blocks.

Lex Fridman (1:02:32.880)

So have a very strong way to do prediction

Lex Fridman (1:02:36.040)

of what given these building blocks,

Lex Fridman (1:02:37.800)

what kind of biology, how the function

Lex Fridman (1:02:40.880)

and the evolution of that biological system.

Demis Hassabis (1:02:43.640)

It's almost like a cellular automata,

Lex Fridman (1:02:45.280)

you have to run it, you can't analyze it from a high level.

Demis Hassabis (1:02:47.880)

You have to take the basic ingredients,

Lex Fridman (1:02:49.840)

figure out the rules and let it run.

Lex Fridman (1:02:51.960)

But in this case, the rules are very difficult

Lex Fridman (1:02:53.960)

to figure out, you have to learn them.

Demis Hassabis (1:02:56.200)

That's exactly it.

Lex Fridman (1:02:57.040)

So the biology is too complicated to figure out the rules.

Demis Hassabis (1:03:00.800)

It's too emergent, too dynamic,

Lex Fridman (1:03:03.600)

say compared to a physics system,

Lex Fridman (1:03:05.080)

like the motion of a planet, right?

Lex Fridman (1:03:07.040)

And so you have to learn the rules

Lex Fridman (1:03:09.200)

and that's exactly the type of systems that we're building.

Lex Fridman (1:03:11.920)

So you mentioned you've open sourced AlphaFold

Lex Fridman (1:03:14.800)

and even the data involved.

Lex Fridman (1:03:16.640)

To me personally, also really happy

Lex Fridman (1:03:20.040)

and a big thank you for open sourcing Mojoko,

Lex Fridman (1:03:23.520)

the physics simulation engine that's often used

Demis Hassabis (1:03:27.080)

for robotics research and so on.

Lex Fridman (1:03:29.080)

So I think that's a pretty gangster move.

Lex Fridman (1:03:31.120)

So what's the, I mean, very few companies

Lex Fridman (1:03:37.200)

or people do that kind of thing.

Lex Fridman (1:03:39.080)

What's the philosophy behind that?

Lex Fridman (1:03:41.240)

You know, it's a case by case basis.

Lex Fridman (1:03:42.920)

And in both of those cases,

Lex Fridman (1:03:44.040)

we felt that was the maximum benefit to humanity to do that.

Lex Fridman (1:03:47.360)

And the scientific community, in one case,

Lex Fridman (1:03:50.040)

the robotics physics community with Mojoko, so.

Demis Hassabis (1:03:53.360)

We purchased it.

Lex Fridman (1:03:54.200)

We purchased it for, yes,

Demis Hassabis (1:03:55.840)

we purchased it for the express principle to open source it.

Lex Fridman (1:03:58.520)

So, you know, I hope people appreciate that.

Demis Hassabis (1:04:02.440)

It's great to hear that you do.

Lex Fridman (1:04:04.040)

And then the second thing was,

Lex Fridman (1:04:05.800)

and mostly we did it because the person building it

Lex Fridman (1:04:08.040)

was not able to cope with supporting it anymore

Demis Hassabis (1:04:11.920)

because it got too big for him.

Lex Fridman (1:04:13.600)

He's an amazing professor who built it in the first place.

Lex Fridman (1:04:16.720)

So we helped him out with that.

Lex Fridman (1:04:18.240)

And then with AlphaFold is even bigger, I would say.

Lex Fridman (1:04:20.520)

And I think in that case,

Lex Fridman (1:04:21.960)

we decided that there were so many downstream applications

Demis Hassabis (1:04:25.520)

of AlphaFold that we couldn't possibly even imagine

Lex Fridman (1:04:29.400)

what they all were.

Lex Fridman (1:04:30.480)

So the best way to accelerate drug discovery

Lex Fridman (1:04:34.360)

and also fundamental research would be to give all

Demis Hassabis (1:04:38.680)

that data away and the system itself.

Lex Fridman (1:04:43.240)

You know, it's been so gratifying to see

Lex Fridman (1:04:45.280)

what people have done that within just one year,

Lex Fridman (1:04:47.040)

which is a short amount of time in science.

Lex Fridman (1:04:49.240)

And it's been used by over 500,000 researchers have used it.

Lex Fridman (1:04:54.160)

We think that's almost every biologist in the world.

Demis Hassabis (1:04:56.560)

I think there's roughly 500,000 biologists in the world,

Lex Fridman (1:04:58.840)

professional biologists,

Demis Hassabis (1:05:00.000)

have used it to look at their proteins of interest.

Lex Fridman (1:05:04.480)

We've seen amazing fundamental research done.

Lex Fridman (1:05:06.520)

So a couple of weeks ago, front cover,

Lex Fridman (1:05:09.040)

there was a whole special issue of science,

Demis Hassabis (1:05:10.840)

including the front cover,

Lex Fridman (1:05:12.040)

which had the nuclear pore complex on it,

Demis Hassabis (1:05:14.000)

which is one of the biggest proteins in the body.

Lex Fridman (1:05:15.800)

The nuclear pore complex is a protein that governs

Demis Hassabis (1:05:18.960)

all the nutrients going in and out of your cell nucleus.

Lex Fridman (1:05:21.680)

So they're like little gateways that open and close

Demis Hassabis (1:05:24.760)

to let things go in and out of your cell nucleus.

Lex Fridman (1:05:27.320)

So they're really important, but they're huge

Demis Hassabis (1:05:29.400)

because they're massive donut ring shaped things.

Lex Fridman (1:05:31.680)

And they've been looking to try and figure out

Demis Hassabis (1:05:33.440)

that structure for decades.

Lex Fridman (1:05:34.960)

And they have lots of experimental data,

Lex Fridman (1:05:37.160)

but it's too low resolution, there's bits missing.

Lex Fridman (1:05:39.600)

And they were able to, like a giant Lego jigsaw puzzle,

Demis Hassabis (1:05:43.080)

use alpha fold predictions plus experimental data

Lex Fridman (1:05:46.200)

and combined those two independent sources of information,

Demis Hassabis (1:05:49.760)

actually four different groups around the world

Lex Fridman (1:05:51.240)

were able to put it together more or less simultaneously

Demis Hassabis (1:05:54.600)

using alpha fold predictions.

Lex Fridman (1:05:56.280)

So that's been amazing to see.

Lex Fridman (1:05:57.720)

And pretty much every pharma company,

Lex Fridman (1:05:59.400)

every drug company executive I've spoken to

Demis Hassabis (1:06:01.440)

has said that their teams are using alpha fold

Lex Fridman (1:06:03.760)

to accelerate whatever drugs they're trying to discover.

Lex Fridman (1:06:08.040)

So I think the knock on effect has been enormous

Lex Fridman (1:06:11.440)

in terms of the impact that alpha fold has made.

Lex Fridman (1:06:15.240)

And it's probably bringing in, it's creating biologists,

Lex Fridman (1:06:17.840)

it's bringing more people into the field,

Demis Hassabis (1:06:20.800)

both on the excitement and both on the technical skills

Lex Fridman (1:06:23.320)

involved in, it's almost like a gateway drug to biology.

Demis Hassabis (1:06:28.760)

Yes, it is.

Lex Fridman (1:06:29.600)

And to get more computational people involved too, hopefully.

Lex Fridman (1:06:32.640)

And I think for us, the next stage, as I said,

Lex Fridman (1:06:35.920)

in future we have to have other considerations too.

Demis Hassabis (1:06:37.960)

We're building on top of alpha fold

Lex Fridman (1:06:39.640)

and these other ideas I discussed with you

Demis Hassabis (1:06:41.200)

about protein interactions and genomics and other things.

Lex Fridman (1:06:44.800)

And not everything will be open source.

Demis Hassabis (1:06:46.200)

Some of it we'll do commercially

Lex Fridman (1:06:48.000)

because that will be the best way

Demis Hassabis (1:06:49.000)

to actually get the most resources and impact behind it.

Lex Fridman (1:06:51.720)

In other ways, some other projects

Demis Hassabis (1:06:53.480)

we'll do nonprofit style.

Lex Fridman (1:06:55.280)

And also we have to consider for future things as well,

Demis Hassabis (1:06:58.520)

safety and ethics as well.

Lex Fridman (1:06:59.720)

Like synthetic biology, there is dual use.

Lex Fridman (1:07:03.600)

And we have to think about that as well.

Lex Fridman (1:07:05.080)

With alpha fold, we consulted with 30 different bioethicists

Lex Fridman (1:07:08.600)

and other people expert in this field

Lex Fridman (1:07:10.240)

to make sure it was safe before we released it.

Lex Fridman (1:07:13.280)

So there'll be other considerations in future.

Lex Fridman (1:07:15.280)

But for right now, I think alpha fold

Demis Hassabis (1:07:17.120)

is a kind of a gift from us to the scientific community.

Lex Fridman (1:07:20.840)

So I'm pretty sure that something like alpha fold

Demis Hassabis (1:07:25.600)

will be part of Nobel prizes in the future.

Lex Fridman (1:07:29.080)

But us humans, of course,

Demis Hassabis (1:07:30.840)

are horrible with credit assignment.

Lex Fridman (1:07:32.480)

So we'll of course give it to the humans.

Lex Fridman (1:07:35.560)

Do you think there will be a day

Lex Fridman (1:07:37.440)

when AI system can't be denied

Lex Fridman (1:07:42.520)

that it earned that Nobel prize?

Lex Fridman (1:07:45.120)

Do you think we will see that in 21st century?

Lex Fridman (1:07:47.400)

It depends what type of AIs we end up building, right?

Lex Fridman (1:07:50.200)

Whether they're goal seeking agents

Demis Hassabis (1:07:53.600)

who specifies the goals, who comes up with the hypotheses,

Lex Fridman (1:07:57.800)

who determines which problems to tackle, right?

Lex Fridman (1:08:00.320)

So I think...

Lex Fridman (1:08:01.160)

And tweets about it, announcement of the results.

Demis Hassabis (1:08:02.440)

Yes, and tweets about results exactly as part of it.

Lex Fridman (1:08:05.440)

So I think right now, of course,

Demis Hassabis (1:08:07.760)

it's amazing human ingenuity that's behind these systems.

Lex Fridman (1:08:12.200)

And then the system, in my opinion, is just a tool.

Demis Hassabis (1:08:15.120)

Be a bit like saying with Galileo and his telescope,

Lex Fridman (1:08:18.400)

the ingenuity that the credit should go to the telescope.

Demis Hassabis (1:08:21.160)

I mean, it's clearly Galileo building the tool

Lex Fridman (1:08:23.560)

which he then uses.

Lex Fridman (1:08:25.160)

So I still see that in the same way today,

Lex Fridman (1:08:27.320)

even though these tools learn for themselves.

Demis Hassabis (1:08:30.440)

There, I think of things like alpha fold

Lex Fridman (1:08:32.960)

and the things we're building as the ultimate tools

Demis Hassabis (1:08:35.840)

for science and for acquiring new knowledge

Lex Fridman (1:08:38.560)

to help us as scientists acquire new knowledge.

Demis Hassabis (1:08:41.160)

I think one day there will come a point

Lex Fridman (1:08:43.200)

where an AI system may solve

Demis Hassabis (1:08:46.360)

or come up with something like general relativity

Lex Fridman (1:08:48.800)

of its own bat, not just by averaging everything

Demis Hassabis (1:08:52.040)

on the internet or averaging everything on PubMed,

Lex Fridman (1:08:55.240)

although that would be interesting to see

Lex Fridman (1:08:56.320)

what that would come up with.

Lex Fridman (1:08:58.520)

So that to me is a bit like our earlier debate

Demis Hassabis (1:09:00.400)

about creativity, you know, inventing go

Lex Fridman (1:09:03.240)

rather than just coming up with a good go move.

Lex Fridman (1:09:06.280)

And so I think solving, I think to, you know,

Lex Fridman (1:09:10.400)

if we wanted to give it the credit

Demis Hassabis (1:09:11.800)

of like a Nobel type of thing,

Lex Fridman (1:09:13.520)

then it would need to invent go

Lex Fridman (1:09:15.800)

and sort of invent that new conjecture out of the blue

Lex Fridman (1:09:19.280)

rather than being specified by the human scientists

Demis Hassabis (1:09:22.720)

or the human creators.

Lex Fridman (1:09:23.560)

So I think right now it's definitely just a tool.

Demis Hassabis (1:09:26.280)

Although it is interesting how far you get

Lex Fridman (1:09:27.880)

by averaging everything on the internet, like you said,

Demis Hassabis (1:09:29.960)

because, you know, a lot of people do see science

Lex Fridman (1:09:33.160)

as you're always standing on the shoulders of giants.

Lex Fridman (1:09:35.640)

And the question is how much are you really reaching

Lex Fridman (1:09:40.040)

up above the shoulders of giants?

Demis Hassabis (1:09:42.000)

Maybe it's just simulating different kinds

Lex Fridman (1:09:44.700)

of results of the past with ultimately this new perspective

Demis Hassabis (1:09:49.360)

that gives you this breakthrough idea.

Lex Fridman (1:09:51.120)

But that idea may not be novel in the way

Demis Hassabis (1:09:54.860)

that it can't be already discovered on the internet.

Lex Fridman (1:09:56.740)

Maybe the Nobel prizes of the next 100 years

Demis Hassabis (1:10:00.080)

are already all there on the internet to be discovered.

Lex Fridman (1:10:03.040)

They could be, they could be.

Demis Hassabis (1:10:04.560)

I mean, I think this is one of the big mysteries,

Lex Fridman (1:10:08.560)

I think is that I, first of all,

Demis Hassabis (1:10:11.720)

I believe a lot of the big new breakthroughs

Lex Fridman (1:10:13.760)

that are gonna come in the next few decades

Lex Fridman (1:10:15.280)

and even in the last decade are gonna come

Lex Fridman (1:10:17.400)

at the intersection between different subject areas

Demis Hassabis (1:10:20.200)

where there'll be some new connection that's found

Lex Fridman (1:10:23.480)

between what seemingly were disparate areas.

Lex Fridman (1:10:26.180)

And one can even think of DeepMind, as I said earlier,

Lex Fridman (1:10:28.840)

as a sort of interdisciplinary between neuroscience ideas

Lex Fridman (1:10:31.720)

and AI engineering ideas originally.

Lex Fridman (1:10:35.040)

And so I think there's that.

Lex Fridman (1:10:37.960)

And then one of the things we can't imagine today is,

Lex Fridman (1:10:40.380)

and one of the reasons I think people,

Demis Hassabis (1:10:41.720)

we were so surprised by how well large models worked

Lex Fridman (1:10:44.440)

is that actually it's very hard for our human minds,

Demis Hassabis (1:10:47.900)

our limited human minds to understand

Lex Fridman (1:10:49.440)

what it would be like to read the whole internet, right?

Demis Hassabis (1:10:52.020)

I think we can do a thought experiment

Lex Fridman (1:10:53.520)

and I used to do this of like,

Lex Fridman (1:10:54.680)

well, what if I read the whole of Wikipedia?

Lex Fridman (1:10:57.600)

What would I know?

Lex Fridman (1:10:58.440)

And I think our minds can just about comprehend

Lex Fridman (1:11:00.480)

maybe what that would be like,

Lex Fridman (1:11:01.920)

but the whole internet is beyond comprehension.

Lex Fridman (1:11:04.440)

So I think we just don't understand what it would be like

Lex Fridman (1:11:07.420)

to be able to hold all of that in mind potentially, right?

Lex Fridman (1:11:10.320)

And then active at once,

Lex Fridman (1:11:12.920)

and then maybe what are the connections

Lex Fridman (1:11:14.520)

that are available there?

Lex Fridman (1:11:15.780)

So I think no doubt there are huge things

Lex Fridman (1:11:17.520)

to be discovered just like that.

Lex Fridman (1:11:19.280)

But I do think there is this other type of creativity

Lex Fridman (1:11:22.280)

of true spark of new knowledge, new idea,

Demis Hassabis (1:11:25.400)

never thought before about,

Lex Fridman (1:11:26.680)

can't be averaged from things that are known,

Demis Hassabis (1:11:29.320)

that really, of course, everything come,

Lex Fridman (1:11:32.000)

nobody creates in a vacuum,

Lex Fridman (1:11:33.680)

so there must be clues somewhere,

Lex Fridman (1:11:35.420)

but just a unique way of putting those things together.

Demis Hassabis (1:11:38.280)

I think some of the greatest scientists in history

Lex Fridman (1:11:40.480)

have displayed that I would say,

Demis Hassabis (1:11:42.240)

although it's very hard to know going back to their time,

Lex Fridman (1:11:45.120)

what was exactly known when they came up with those things.

Demis Hassabis (1:11:48.120)

Although you're making me really think because just a thought

Lex Fridman (1:11:52.440)

experiment of deeply knowing a hundred Wikipedia pages.

Demis Hassabis (1:11:57.360)

I don't think I can,

Lex Fridman (1:11:59.200)

I've been really impressed by Wikipedia for technical topics.

Lex Fridman (1:12:03.400)

So if you know a hundred pages or a thousand pages,

Lex Fridman (1:12:07.040)

I don't think we can truly comprehend

Lex Fridman (1:12:10.120)

what kind of intelligence that is.

Lex Fridman (1:12:13.400)

That's a pretty powerful intelligence.

Demis Hassabis (1:12:14.760)

If you know how to use that

Lex Fridman (1:12:16.120)

and integrate that information correctly,

Demis Hassabis (1:12:18.320)

I think you can go really far.

Lex Fridman (1:12:20.000)

You can probably construct thought experiments

Demis Hassabis (1:12:22.080)

based on that, like simulate different ideas.

Lex Fridman (1:12:25.840)

So if this is true, let me run this thought experiment

Demis Hassabis (1:12:28.840)

that maybe this is true.

Lex Fridman (1:12:30.160)

It's not really invention.

Demis Hassabis (1:12:31.360)

It's like just taking literally the knowledge

Lex Fridman (1:12:34.640)

and using it to construct the very basic simulation

Demis Hassabis (1:12:37.240)

of the world.

Lex Fridman (1:12:38.080)

I mean, some argue it's romantic in part,

Lex Fridman (1:12:40.080)

but Einstein would do the same kind of things

Lex Fridman (1:12:42.400)

with a thought experiment.

Demis Hassabis (1:12:43.720)

Yeah, one could imagine doing that systematically

Lex Fridman (1:12:46.320)

across millions of Wikipedia pages,

Demis Hassabis (1:12:48.440)

plus PubMed, all these things.

Lex Fridman (1:12:50.400)

I think there are many, many things to be discovered

Demis Hassabis (1:12:53.680)

like that that are hugely useful.

Lex Fridman (1:12:55.280)

You could imagine,

Lex Fridman (1:12:56.200)

and I want us to do some of these things in material science

Lex Fridman (1:12:58.520)

like room temperature superconductors

Demis Hassabis (1:13:00.000)

is something on my list one day.

Lex Fridman (1:13:01.560)

I'd like to have an AI system to help build

Demis Hassabis (1:13:05.000)

better optimized batteries,

Lex Fridman (1:13:06.640)

all of these sort of mechanical things.

Demis Hassabis (1:13:09.000)

I think a systematic sort of search

Lex Fridman (1:13:11.600)

could be guided by a model,

Demis Hassabis (1:13:14.360)

could be extremely powerful.

Lex Fridman (1:13:17.120)

So speaking of which,

Demis Hassabis (1:13:18.160)

you have a paper on nuclear fusion,

Lex Fridman (1:13:21.320)

magnetic control of tachymic plasmas

Demis Hassabis (1:13:23.120)

through deep reinforcement learning.

Lex Fridman (1:13:24.720)

So you're seeking to solve nuclear fusion with deep RL.

Lex Fridman (1:13:29.800)

So it's doing control of high temperature plasmas.

Lex Fridman (1:13:31.840)

Can you explain this work

Lex Fridman (1:13:33.520)

and can AI eventually solve nuclear fusion?

Lex Fridman (1:13:37.240)

It's been very fun last year or two and very productive

Demis Hassabis (1:13:40.200)

because we've been taking off a lot of my dream projects,

Lex Fridman (1:13:43.360)

if you like, of things that I've collected

Demis Hassabis (1:13:44.960)

over the years of areas of science

Lex Fridman (1:13:46.960)

that I would like to,

Demis Hassabis (1:13:48.200)

I think could be very transformative if we helped accelerate

Lex Fridman (1:13:51.200)

and really interesting problems,

Demis Hassabis (1:13:53.600)

scientific challenges in of themselves.

Lex Fridman (1:13:55.760)

So this is energy.

Lex Fridman (1:13:57.040)

So energy, yes, exactly.

Lex Fridman (1:13:58.520)

So energy and climate.

Lex Fridman (1:13:59.960)

So we talked about disease and biology

Lex Fridman (1:14:01.760)

as being one of the biggest places I think AI can help with.

Demis Hassabis (1:14:04.520)

I think energy and climate is another one.

Lex Fridman (1:14:07.120)

So maybe they would be my top two.

Lex Fridman (1:14:09.240)

And fusion is one area I think AI can help with.

Lex Fridman (1:14:12.520)

Now, fusion has many challenges,

Demis Hassabis (1:14:15.360)

mostly physics and material science

Lex Fridman (1:14:17.240)

and engineering challenges as well

Demis Hassabis (1:14:18.600)

to build these massive fusion reactors

Lex Fridman (1:14:20.520)

and contain the plasma.

Lex Fridman (1:14:21.920)

And what we try to do,

Lex Fridman (1:14:22.760)

and whenever we go into a new field to apply our systems,

Demis Hassabis (1:14:26.280)

is we look for, we talk to domain experts.

Lex Fridman (1:14:29.240)

We try and find the best people in the world

Demis Hassabis (1:14:30.680)

to collaborate with.

Lex Fridman (1:14:33.000)

In this case, in fusion,

Demis Hassabis (1:14:34.120)

we collaborated with EPFL in Switzerland,

Lex Fridman (1:14:36.400)

the Swiss Technical Institute, who are amazing.

Demis Hassabis (1:14:38.280)

They have a test reactor.

Lex Fridman (1:14:39.640)

They were willing to let us use,

Demis Hassabis (1:14:41.360)

which I double checked with the team

Lex Fridman (1:14:43.400)

we were gonna use carefully and safely.

Demis Hassabis (1:14:46.120)

I was impressed they managed to persuade them

Lex Fridman (1:14:47.760)

to let us use it.

Lex Fridman (1:14:49.160)

And it's an amazing test reactor they have there.

Lex Fridman (1:14:53.440)

And they try all sorts of pretty crazy experiments on it.

Lex Fridman (1:14:57.000)

And what we tend to look at is,

Lex Fridman (1:14:59.720)

if we go into a new domain like fusion,

Lex Fridman (1:15:01.760)

what are all the bottleneck problems?

Lex Fridman (1:15:04.160)

Like thinking from first principles,

Lex Fridman (1:15:05.960)

what are all the bottleneck problems

Lex Fridman (1:15:07.000)

that are still stopping fusion working today?

Lex Fridman (1:15:09.280)

And then we look at, we get a fusion expert to tell us,

Lex Fridman (1:15:12.080)

and then we look at those bottlenecks

Lex Fridman (1:15:13.760)

and we look at the ones,

Lex Fridman (1:15:14.600)

which ones are amenable to our AI methods today, right?

Lex Fridman (1:15:18.920)

And would be interesting from a research perspective,

Lex Fridman (1:15:22.200)

from our point of view, from an AI point of view,

Lex Fridman (1:15:24.400)

and that would address one of their bottlenecks.

Lex Fridman (1:15:26.760)

And in this case, plasma control was perfect.

Demis Hassabis (1:15:29.720)

So, the plasma, it's a million degrees Celsius,

Lex Fridman (1:15:32.480)

something like that, it's hotter than the sun.

Lex Fridman (1:15:34.640)

And there's obviously no material that can contain it.

Lex Fridman (1:15:37.640)

So, they have to be containing these magnetic,

Demis Hassabis (1:15:39.440)

very powerful and superconducting magnetic fields.

Lex Fridman (1:15:42.520)

But the problem is plasma,

Demis Hassabis (1:15:43.960)

it's pretty unstable as you imagine,

Lex Fridman (1:15:45.360)

you're kind of holding a mini sun, mini star in a reactor.

Demis Hassabis (1:15:49.320)

So, you kind of want to predict ahead of time,

Lex Fridman (1:15:52.520)

what the plasma is gonna do.

Demis Hassabis (1:15:54.040)

So, you can move the magnetic field

Lex Fridman (1:15:56.240)

within a few milliseconds,

Demis Hassabis (1:15:58.440)

to basically contain what it's gonna do next.

Lex Fridman (1:16:00.960)

So, it seems like a perfect problem if you think of it

Demis Hassabis (1:16:03.160)

for like a reinforcement learning prediction problem.

Lex Fridman (1:16:06.280)

So, you got controller, you're gonna move the magnetic field.

Lex Fridman (1:16:09.720)

And until we came along, they were doing it

Lex Fridman (1:16:12.560)

with traditional operational research type of controllers,

Demis Hassabis (1:16:16.720)

which are kind of handcrafted.

Lex Fridman (1:16:18.320)

And the problem is, of course,

Demis Hassabis (1:16:19.160)

they can't react in the moment

Lex Fridman (1:16:20.480)

to something the plasma is doing,

Demis Hassabis (1:16:21.640)

they have to be hard coded.

Lex Fridman (1:16:23.040)

And again, knowing that that's normally our go to solution

Demis Hassabis (1:16:26.040)

is we would like to learn that instead.

Lex Fridman (1:16:27.960)

And they also had a simulator of these plasma.

Demis Hassabis (1:16:30.320)

So, there were lots of criteria

Lex Fridman (1:16:31.480)

that matched what we like to use.

Lex Fridman (1:16:34.760)

So, can AI eventually solve nuclear fusion?

Lex Fridman (1:16:38.440)

Well, so with this problem,

Lex Fridman (1:16:39.760)

and we published it in a nature paper last year,

Lex Fridman (1:16:42.040)

we held the fusion, we held the plasma in a specific shapes.

Demis Hassabis (1:16:46.160)

So, actually, it's almost like carving the plasma

Lex Fridman (1:16:48.360)

into different shapes and hold it there

Demis Hassabis (1:16:51.000)

for a record amount of time.

Lex Fridman (1:16:52.880)

So, that's one of the problems of fusion sort of solved.

Demis Hassabis (1:16:57.600)

So, have a controller that's able to,

Lex Fridman (1:16:59.840)

no matter the shape.

Demis Hassabis (1:17:01.480)

Contain it. Contain it.

Lex Fridman (1:17:02.360)

Yeah, contain it and hold it in structure.

Lex Fridman (1:17:04.160)

And there's different shapes that are better

Lex Fridman (1:17:05.760)

for the energy productions called droplets and so on.

Demis Hassabis (1:17:10.080)

So, that was huge.

Lex Fridman (1:17:11.880)

And now we're looking,

Demis Hassabis (1:17:12.720)

we're talking to lots of fusion startups

Lex Fridman (1:17:14.400)

to see what's the next problem we can tackle

Demis Hassabis (1:17:17.400)

in the fusion area.

Lex Fridman (1:17:19.360)

So, another fascinating place in a paper titled,

Demis Hassabis (1:17:23.080)

Pushing the Frontiers of Density Functionals

Lex Fridman (1:17:25.120)

by Solving the Fractional Electron Problem.

Demis Hassabis (1:17:27.520)

So, you're taking on modeling and simulating

Lex Fridman (1:17:30.880)

the quantum mechanical behavior of electrons.

Demis Hassabis (1:17:33.320)

Yes.

Lex Fridman (1:17:36.040)

Can you explain this work and can AI model

Lex Fridman (1:17:39.240)

and simulate arbitrary quantum mechanical systems

Lex Fridman (1:17:41.560)

in the future?

Demis Hassabis (1:17:42.400)

Yeah, so this is another problem I've had my eye on

Lex Fridman (1:17:44.240)

for a decade or more,

Demis Hassabis (1:17:47.160)

which is sort of simulating the properties of electrons.

Lex Fridman (1:17:51.200)

If you can do that, you can basically describe

Lex Fridman (1:17:54.280)

how elements and materials and substances work.

Lex Fridman (1:17:58.040)

So, it's kind of like fundamental

Demis Hassabis (1:18:00.040)

if you want to advance material science.

Lex Fridman (1:18:02.840)

And we have Schrodinger's equation

Lex Fridman (1:18:05.240)

and then we have approximations

Lex Fridman (1:18:06.480)

to that density functional theory.

Demis Hassabis (1:18:08.400)

These things are famous.

Lex Fridman (1:18:10.560)

And people try and write approximations

Demis Hassabis (1:18:13.200)

to these functionals and kind of come up

Lex Fridman (1:18:17.040)

with descriptions of the electron clouds,

Demis Hassabis (1:18:19.880)

where they're going to go,

Lex Fridman (1:18:20.720)

how they're going to interact

Demis Hassabis (1:18:22.120)

when you put two elements together.

Lex Fridman (1:18:24.240)

And what we try to do is learn a simulation,

Demis Hassabis (1:18:27.680)

learn a functional that will describe more chemistry,

Lex Fridman (1:18:30.560)

types of chemistry.

Demis Hassabis (1:18:31.760)

So, until now, you can run expensive simulations,

Lex Fridman (1:18:35.560)

but then you can only simulate very small molecules,

Demis Hassabis (1:18:38.760)

very simple molecules.

Lex Fridman (1:18:40.160)

We would like to simulate large materials.

Lex Fridman (1:18:43.080)

And so, today there's no way of doing that.

Lex Fridman (1:18:45.760)

And we're building up towards building functionals

Demis Hassabis (1:18:48.560)

that approximate Schrodinger's equation

Lex Fridman (1:18:51.240)

and then allow you to describe what the electrons are doing.

Lex Fridman (1:18:55.600)

And all material sort of science

Lex Fridman (1:18:57.480)

and material properties are governed by the electrons

Lex Fridman (1:18:59.920)

and how they interact.

Lex Fridman (1:19:01.360)

So, have a good summarization of the simulation

Demis Hassabis (1:19:05.840)

through the functional,

Lex Fridman (1:19:08.720)

but one that is still close

Demis Hassabis (1:19:11.360)

to what the actual simulation would come out with.

Lex Fridman (1:19:13.200)

So, how difficult is that task?

Lex Fridman (1:19:16.720)

What's involved in that task?

Lex Fridman (1:19:17.760)

Is it running those complicated simulations

Lex Fridman (1:19:20.720)

and learning the task of mapping

Lex Fridman (1:19:23.280)

from the initial conditions

Lex Fridman (1:19:24.560)

and the parameters of the simulation,

Lex Fridman (1:19:26.400)

learning what the functional would be?

Demis Hassabis (1:19:27.720)

Yeah.

Lex Fridman (1:19:28.560)

So, it's pretty tricky.

Lex Fridman (1:19:29.440)

And we've done it with,

Lex Fridman (1:19:31.320)

the nice thing is we can run a lot of the simulations,

Demis Hassabis (1:19:35.440)

the molecular dynamic simulations on our compute clusters.

Lex Fridman (1:19:39.080)

And so, that generates a lot of data.

Demis Hassabis (1:19:40.840)

So, in this case, the data is generated.

Lex Fridman (1:19:42.800)

So, we like those sort of systems and that's why we use games.

Demis Hassabis (1:19:45.880)

It's simulated, generated data.

Lex Fridman (1:19:48.480)

And we can kind of create as much of it as we want, really.

Lex Fridman (1:19:51.160)

And just let's leave some,

Lex Fridman (1:19:53.280)

if any computers are free in the cloud,

Lex Fridman (1:19:55.280)

we just run, we run some of these calculations, right?

Lex Fridman (1:19:57.680)

Compute cluster calculation.

Demis Hassabis (1:19:59.360)

I like how the free compute time

Lex Fridman (1:20:01.080)

is used up on quantum mechanics.

Demis Hassabis (1:20:02.200)

Yeah, quantum mechanics, exactly.

Lex Fridman (1:20:03.560)

Simulations and protein simulations and other things.

Lex Fridman (1:20:06.280)

And so, when you're not searching on YouTube

Lex Fridman (1:20:09.880)

for free video, cat videos,

Demis Hassabis (1:20:11.360)

we're using those computers usefully in quantum chemistry.

Lex Fridman (1:20:13.960)

It's the idea.

Demis Hassabis (1:20:14.800)

Finally.

Lex Fridman (1:20:15.640)

And putting them to good use.

Lex Fridman (1:20:17.000)

And then, yeah, and then all of that computational data

Lex Fridman (1:20:19.760)

that's generated,

Demis Hassabis (1:20:20.840)

we can then try and learn the functionals from that,

Lex Fridman (1:20:23.480)

which of course are way more efficient

Demis Hassabis (1:20:25.640)

once we learn the functional

Lex Fridman (1:20:27.080)

than running those simulations would be.

Lex Fridman (1:20:30.560)

Do you think one day AI may allow us

Lex Fridman (1:20:33.120)

to do something like basically crack open physics?

Lex Fridman (1:20:36.360)

So, do something like travel faster than the speed of light?

Lex Fridman (1:20:39.520)

My ultimate aim is always being with AI

Demis Hassabis (1:20:41.600)

is the reason I am personally working on AI

Lex Fridman (1:20:45.560)

for my whole life, it was to build a tool

Demis Hassabis (1:20:48.200)

to help us understand the universe.

Lex Fridman (1:20:50.360)

So, I wanted to, and that means physics, really,

Lex Fridman (1:20:53.800)

and the nature of reality.

Lex Fridman (1:20:54.920)

So, I don't think we have systems

Demis Hassabis (1:20:58.000)

that are capable of doing that yet,

Lex Fridman (1:20:59.400)

but when we get towards AGI,

Demis Hassabis (1:21:01.000)

I think that's one of the first things

Lex Fridman (1:21:02.920)

I think we should apply AGI to.

Demis Hassabis (1:21:05.320)

I would like to test the limits of physics

Lex Fridman (1:21:07.160)

and our knowledge of physics.

Demis Hassabis (1:21:08.600)

There's so many things we don't know.

Lex Fridman (1:21:10.080)

This is one thing I find fascinating about science.

Lex Fridman (1:21:12.320)

And as a huge proponent of the scientific method

Lex Fridman (1:21:15.080)

as being one of the greatest ideas humanity has ever had

Lex Fridman (1:21:17.880)

and allowed us to progress with our knowledge,

Lex Fridman (1:21:20.160)

but I think as a true scientist,

Demis Hassabis (1:21:22.000)

I think what you find is the more you find out,

Lex Fridman (1:21:25.200)

the more you realize we don't know.

Lex Fridman (1:21:27.040)

And I always think that it's surprising

Lex Fridman (1:21:29.880)

that more people aren't troubled.

Demis Hassabis (1:21:31.880)

Every night I think about all these things

Lex Fridman (1:21:34.000)

we interact with all the time,

Demis Hassabis (1:21:35.240)

that we have no idea how they work.

Lex Fridman (1:21:36.880)

Time, consciousness, gravity, life, we can't,

Demis Hassabis (1:21:41.440)

I mean, these are all the fundamental things of nature.

Lex Fridman (1:21:43.840)

I think the way we don't really know what they are.

Demis Hassabis (1:21:47.320)

To live life, we pin certain assumptions on them

Lex Fridman (1:21:51.480)

and kind of treat our assumptions as if they're a fact.

Demis Hassabis (1:21:55.240)

That allows us to sort of box them off somehow.

Lex Fridman (1:21:57.560)

Yeah, box them off somehow.

Lex Fridman (1:21:59.000)

But the reality is when you think of time,

Lex Fridman (1:22:02.320)

you should remind yourself,

Demis Hassabis (1:22:03.560)

you should take it off the shelf

Lex Fridman (1:22:06.760)

and realize like, no, we have a bunch of assumptions.

Demis Hassabis (1:22:09.040)

There's still a lot of, there's even now a lot of debate.

Lex Fridman (1:22:11.520)

There's a lot of uncertainty about exactly what is time.

Lex Fridman (1:22:15.520)

Is there an error of time?

Lex Fridman (1:22:17.480)

You know, there's a lot of fundamental questions

Demis Hassabis (1:22:19.480)

that you can't just make assumptions about.

Lex Fridman (1:22:21.160)

And maybe AI allows you to not put anything on the shelf.

Demis Hassabis (1:22:27.680)

Yeah.

Lex Fridman (1:22:28.520)

Not make any hard assumptions

Lex Fridman (1:22:30.200)

and really open it up and see what's.

Lex Fridman (1:22:32.080)

Exactly, I think we should be truly open minded about that.

Lex Fridman (1:22:34.640)

And exactly that, not be dogmatic to a particular theory.

Lex Fridman (1:22:39.040)

It'll also allow us to build better tools,

Demis Hassabis (1:22:41.960)

experimental tools eventually,

Lex Fridman (1:22:44.400)

that can then test certain theories

Demis Hassabis (1:22:46.280)

that may not be testable today.

Lex Fridman (1:22:48.080)

Things about like what we spoke about at the beginning

Demis Hassabis (1:22:51.240)

about the computational nature of the universe.

Lex Fridman (1:22:53.520)

How one might, if that was true,

Lex Fridman (1:22:55.320)

how one might go about testing that, right?

Lex Fridman (1:22:57.360)

And how much, you know, there are people

Demis Hassabis (1:22:59.840)

who've conjectured people like Scott Aaronson and others

Lex Fridman (1:23:02.520)

about, you know, how much information

Demis Hassabis (1:23:04.720)

can a specific plank unit of space and time

Lex Fridman (1:23:08.040)

contain, right?

Lex Fridman (1:23:09.200)

So one might be able to think about testing those ideas

Lex Fridman (1:23:11.960)

if you had AI helping you build

Demis Hassabis (1:23:15.360)

some new exquisite experimental tools.

Lex Fridman (1:23:19.400)

This is what I imagine that, you know,

Demis Hassabis (1:23:20.960)

many decades from now we'll be able to do.

Lex Fridman (1:23:23.120)

And what kind of questions can be answered

Lex Fridman (1:23:25.840)

through running a simulation of them?

Lex Fridman (1:23:28.840)

So there's a bunch of physics simulations

Demis Hassabis (1:23:30.760)

you can imagine that could be run

Lex Fridman (1:23:32.600)

in some kind of efficient way,

Demis Hassabis (1:23:35.760)

much like you're doing in the quantum simulation work.

Lex Fridman (1:23:40.320)

And perhaps even the origin of life.

Lex Fridman (1:23:42.160)

So figuring out how going even back

Lex Fridman (1:23:45.160)

before the work of AlphaFold begins

Demis Hassabis (1:23:47.640)

of how this whole thing emerges from a rock.

Lex Fridman (1:23:52.640)

Yes.

Demis Hassabis (1:23:53.480)

From a static thing.

Lex Fridman (1:23:54.320)

What do you think AI will allow us to,

Lex Fridman (1:23:57.040)

is that something you have your eye on?

Lex Fridman (1:23:58.920)

It's trying to understand the origin of life.

Demis Hassabis (1:24:01.560)

First of all, yourself, what do you think,

Lex Fridman (1:24:06.320)

how the heck did life originate on Earth?

Demis Hassabis (1:24:08.760)

Yeah, well, maybe I'll come to that in a second,

Lex Fridman (1:24:11.120)

but I think the ultimate use of AI

Demis Hassabis (1:24:13.800)

is to kind of use it to accelerate science to the maximum.

Lex Fridman (1:24:18.120)

So I think of it a little bit

Demis Hassabis (1:24:21.040)

like the tree of all knowledge.

Lex Fridman (1:24:22.600)

If you imagine that's all the knowledge there is

Demis Hassabis (1:24:24.160)

in the universe to attain.

Lex Fridman (1:24:25.840)

And we sort of barely scratched the surface of that so far.

Lex Fridman (1:24:29.320)

And even though we've done pretty well

Lex Fridman (1:24:31.960)

since the enlightenment, right, as humanity.

Lex Fridman (1:24:34.320)

And I think AI will turbocharge all of that,

Lex Fridman (1:24:36.840)

like we've seen with AlphaFold.

Lex Fridman (1:24:38.600)

And I want to explore as much of that tree of knowledge

Lex Fridman (1:24:41.400)

as is possible to do.

Lex Fridman (1:24:42.920)

And I think that involves AI helping us

Lex Fridman (1:24:46.400)

with understanding or finding patterns,

Lex Fridman (1:24:49.680)

but also potentially designing and building new tools,

Lex Fridman (1:24:52.200)

experimental tools.

Lex Fridman (1:24:53.600)

So I think that's all,

Lex Fridman (1:24:56.040)

and also running simulations and learning simulations,

Demis Hassabis (1:24:58.920)

all of that we're sort of doing at a baby steps level here.

Lex Fridman (1:25:05.000)

But I can imagine that in the decades to come

Demis Hassabis (1:25:08.560)

as what's the full flourishing of that line of thinking.

Lex Fridman (1:25:12.920)

It's gonna be truly incredible, I would say.

Demis Hassabis (1:25:15.160)

If I visualized this tree of knowledge,

Lex Fridman (1:25:17.320)

something tells me that that tree of knowledge for humans

Demis Hassabis (1:25:20.840)

is much smaller in the set of all possible trees

Lex Fridman (1:25:24.440)

of knowledge, it's actually quite small

Demis Hassabis (1:25:26.600)

given our cognitive limitations,

Lex Fridman (1:25:31.480)

limited cognitive capabilities,

Demis Hassabis (1:25:33.680)

that even with the tools we build,

Lex Fridman (1:25:35.720)

we still won't be able to understand a lot of things.

Lex Fridman (1:25:38.120)

And that's perhaps what nonhuman systems

Lex Fridman (1:25:41.160)

might be able to reach farther, not just as tools,

Lex Fridman (1:25:44.920)

but in themselves understanding something

Lex Fridman (1:25:47.200)

that they can bring back.

Demis Hassabis (1:25:48.480)

Yeah, it could well be.

Lex Fridman (1:25:50.200)

So, I mean, there's so many things

Demis Hassabis (1:25:51.800)

that are sort of encapsulated in what you just said there.

Lex Fridman (1:25:55.000)

I think first of all, there's two different things.

Lex Fridman (1:25:58.320)

There's like, what do we understand today?

Lex Fridman (1:26:00.560)

What could the human mind understand?

Lex Fridman (1:26:02.680)

And what is the totality of what is there to be understood?

Lex Fridman (1:26:06.400)

And so there's three concentric,

Demis Hassabis (1:26:08.640)

you can think of them as three larger and larger trees

Lex Fridman (1:26:10.720)

or exploring more branches of that tree.

Lex Fridman (1:26:12.880)

And I think with AI, we're gonna explore that whole lot.

Lex Fridman (1:26:15.960)

Now, the question is, if you think about

Lex Fridman (1:26:19.120)

what is the totality of what could be understood,

Lex Fridman (1:26:22.680)

there may be some fundamental physics reasons

Lex Fridman (1:26:24.800)

why certain things can't be understood,

Lex Fridman (1:26:26.280)

like what's outside a simulation or outside the universe.

Demis Hassabis (1:26:29.000)

Maybe it's not understandable from within the universe.

Lex Fridman (1:26:32.320)

So there may be some hard constraints like that.

Demis Hassabis (1:26:34.840)

It could be smaller constraints,

Lex Fridman (1:26:36.000)

like we think of space time as fundamental.

Demis Hassabis (1:26:40.520)

Our human brains are really used to this idea

Lex Fridman (1:26:42.880)

of a three dimensional world with time, maybe.

Lex Fridman (1:26:46.040)

But our tools could go beyond that.

Lex Fridman (1:26:47.760)

They wouldn't have that limitation necessarily.

Demis Hassabis (1:26:49.760)

They could think in 11 dimensions, 12 dimensions,

Lex Fridman (1:26:51.760)

whatever is needed.

Lex Fridman (1:26:52.920)

But we could still maybe understand that

Lex Fridman (1:26:55.640)

in several different ways.

Demis Hassabis (1:26:56.720)

The example I always give is,

Lex Fridman (1:26:59.040)

when I play Garry Kasparov for speed chess,

Demis Hassabis (1:27:01.400)

or we've talked about chess and these kinds of things,

Lex Fridman (1:27:04.400)

you know, if you're reasonably good at chess,

Demis Hassabis (1:27:07.520)

you can't come up with the move Garry comes up with

Lex Fridman (1:27:11.200)

in his move, but he can explain it to you.

Lex Fridman (1:27:13.320)

And you can understand.

Lex Fridman (1:27:14.160)

And you can understand post hoc the reasoning.

Lex Fridman (1:27:16.720)

So I think there's an even further level of like,

Lex Fridman (1:27:19.400)

well, maybe you couldn't have invented that thing,

Lex Fridman (1:27:21.640)

but going back to using language again,

Lex Fridman (1:27:24.320)

perhaps you can understand and appreciate that.

Demis Hassabis (1:27:27.040)

Same way that you can appreciate, you know,

Lex Fridman (1:27:28.920)

Vivaldi or Mozart or something without,

Demis Hassabis (1:27:31.120)

you can appreciate the beauty of that

Lex Fridman (1:27:32.680)

without being able to construct it yourself, right?

Demis Hassabis (1:27:35.800)

Invent the music yourself.

Lex Fridman (1:27:37.400)

So I think we see this in all forms of life.

Lex Fridman (1:27:39.280)

So it will be that times, you know, a million,

Lex Fridman (1:27:42.440)

but you can imagine also one sign of intelligence

Lex Fridman (1:27:45.800)

is the ability to explain things clearly and simply, right?

Lex Fridman (1:27:49.320)

You know, people like Richard Feynman,

Lex Fridman (1:27:50.400)

another one of my old time heroes used to say that, right?

Lex Fridman (1:27:52.400)

If you can't, you know, if you can explain it

Demis Hassabis (1:27:54.480)

something simply, then that's the best sign,

Lex Fridman (1:27:57.360)

a complex topic simply,

Demis Hassabis (1:27:58.640)

then that's one of the best signs of you understanding it.

Lex Fridman (1:28:00.680)

Yeah.

Demis Hassabis (1:28:01.520)

I can see myself talking trash in the AI system in that way.

Lex Fridman (1:28:04.600)

Yes.

Demis Hassabis (1:28:05.680)

It gets frustrated how dumb I am

Lex Fridman (1:28:07.800)

and trying to explain something to me.

Demis Hassabis (1:28:09.880)

I was like, well, that means you're not intelligent

Lex Fridman (1:28:11.600)

because if you were intelligent,

Demis Hassabis (1:28:12.720)

you'd be able to explain it simply.

Lex Fridman (1:28:14.440)

Yeah, of course, you know, there's also the other option.

Demis Hassabis (1:28:16.720)

Of course, we could enhance ourselves and with our devices,

Lex Fridman (1:28:19.560)

we are already sort of symbiotic with our compute devices,

Demis Hassabis (1:28:23.120)

right, with our phones and other things.

Lex Fridman (1:28:24.600)

And, you know, there's stuff like Neuralink and Xceptra

Demis Hassabis (1:28:27.120)

that could advance that further.

Lex Fridman (1:28:30.000)

So I think there's lots of really amazing possibilities

Demis Hassabis (1:28:33.880)

that I could foresee from here.

Lex Fridman (1:28:35.360)

Well, let me ask you some wild questions.

Lex Fridman (1:28:37.040)

So out there looking for friends,

Lex Fridman (1:28:39.920)

do you think there's a lot of alien civilizations out there?

Lex Fridman (1:28:43.120)

So I guess this also goes back

Lex Fridman (1:28:44.960)

to your origin of life question too,

Demis Hassabis (1:28:46.640)

because I think that that's key.

Lex Fridman (1:28:48.240)

My personal opinion, looking at all this,

Demis Hassabis (1:28:51.360)

and, you know, it's one of my hobbies, physics, I guess.

Lex Fridman (1:28:53.680)

So, you know, it's something I think about a lot

Lex Fridman (1:28:56.880)

and talk to a lot of experts on and read a lot of books on.

Lex Fridman (1:29:00.760)

And I think my feeling currently is that we are alone.

Demis Hassabis (1:29:05.280)

I think that's the most likely scenario

Lex Fridman (1:29:07.160)

given what evidence we have.

Demis Hassabis (1:29:08.800)

So, and the reasoning is I think that, you know,

Lex Fridman (1:29:13.160)

we've tried since things like SETI program

Lex Fridman (1:29:16.120)

and I guess since the dawning of the space age,

Lex Fridman (1:29:19.840)

we've, you know, had telescopes,

Demis Hassabis (1:29:21.240)

open radio telescopes and other things.

Lex Fridman (1:29:23.280)

And if you think about and try to detect signals,

Demis Hassabis (1:29:27.280)

now, if you think about the evolution of humans on earth,

Lex Fridman (1:29:30.120)

we could have easily been a million years ahead

Demis Hassabis (1:29:33.960)

of our time now or million years behind,

Lex Fridman (1:29:36.280)

right, easily with just some slightly different quirk

Demis Hassabis (1:29:39.440)

thing happening hundreds of thousands of years ago.

Lex Fridman (1:29:42.120)

You know, things could have been slightly different

Demis Hassabis (1:29:43.640)

if the meteor would hit the dinosaurs a million years earlier,

Lex Fridman (1:29:46.200)

maybe things would have evolved.

Demis Hassabis (1:29:48.080)

We'd be a million years ahead of where we are now.

Lex Fridman (1:29:50.920)

So what that means is if you imagine where humanity will be

Demis Hassabis (1:29:54.080)

in a few hundred years, let alone a million years,

Lex Fridman (1:29:56.720)

especially if we hopefully, you know,

Demis Hassabis (1:29:59.840)

solve things like climate change and other things,

Lex Fridman (1:30:02.200)

and we continue to flourish and we build things like AI

Lex Fridman (1:30:05.640)

and we do space traveling and all of the stuff

Lex Fridman (1:30:07.920)

that humans have dreamed of forever, right?

Lex Fridman (1:30:10.800)

And sci fi is talked about forever.

Lex Fridman (1:30:14.360)

We will be spreading across the stars, right?

Lex Fridman (1:30:16.760)

And von Neumann famously calculated, you know,

Lex Fridman (1:30:19.240)

it would only take about a million years

Demis Hassabis (1:30:20.800)

if you sent out von Neumann probes to the nearest,

Lex Fridman (1:30:23.240)

you know, the nearest other solar systems.

Lex Fridman (1:30:26.200)

And then all they did was build two more versions

Lex Fridman (1:30:29.040)

of themselves and sent those two out

Demis Hassabis (1:30:30.440)

to the next nearest systems.

Lex Fridman (1:30:32.240)

You know, within a million years,

Demis Hassabis (1:30:33.480)

I think you would have one of these probes

Lex Fridman (1:30:35.040)

in every system in the galaxy.

Lex Fridman (1:30:36.920)

So it's not actually in cosmological time.

Lex Fridman (1:30:40.040)

That's actually a very short amount of time.

Demis Hassabis (1:30:42.080)

So, and you know, people like Dyson have thought

Lex Fridman (1:30:44.600)

about constructing Dyson spheres around stars

Demis Hassabis (1:30:47.280)

to collect all the energy coming out of the star.

Lex Fridman (1:30:49.800)

You know, there would be constructions like that

Demis Hassabis (1:30:51.800)

would be visible across space,

Lex Fridman (1:30:54.120)

probably even across a galaxy.

Demis Hassabis (1:30:56.000)

So, and then, you know, if you think about

Lex Fridman (1:30:57.920)

all of our radio, television emissions

Demis Hassabis (1:31:00.760)

that have gone out since the, you know, 30s and 40s,

Lex Fridman (1:31:05.120)

imagine a million years of that.

Lex Fridman (1:31:06.720)

And now hundreds of civilizations doing that.

Lex Fridman (1:31:10.000)

When we opened our ears at the point

Demis Hassabis (1:31:12.200)

we got technologically sophisticated enough

Lex Fridman (1:31:14.840)

in the space age,

Demis Hassabis (1:31:15.880)

we should have heard a cacophony of voices.

Lex Fridman (1:31:19.120)

We should have joined that cacophony of voices.

Lex Fridman (1:31:20.920)

And what we did, we opened our ears and we heard nothing.

Lex Fridman (1:31:24.480)

And many people who argue that there are aliens

Demis Hassabis (1:31:27.120)

would say, well, we haven't really done

Lex Fridman (1:31:28.800)

exhaustive search yet.

Lex Fridman (1:31:29.920)

And maybe we're looking in the wrong bands

Lex Fridman (1:31:31.880)

and we've got the wrong devices

Lex Fridman (1:31:33.760)

and we wouldn't notice what an alien form was like

Lex Fridman (1:31:36.080)

because it'd be so different to what we're used to.

Demis Hassabis (1:31:38.280)

But, you know, I don't really buy that,

Lex Fridman (1:31:40.640)

that it shouldn't be as difficult as that.

Demis Hassabis (1:31:42.640)

Like, I think we've searched enough.

Lex Fridman (1:31:44.320)

There should be everywhere.

Demis Hassabis (1:31:45.680)

If it was, yeah, it should be everywhere.

Lex Fridman (1:31:47.280)

We should see Dyson spheres being put up,

Demis Hassabis (1:31:49.240)

sun's blinking in and out.

Lex Fridman (1:31:50.600)

You know, there should be a lot of evidence

Demis Hassabis (1:31:52.000)

for those things.

Lex Fridman (1:31:52.920)

And then there are other people who argue,

Demis Hassabis (1:31:54.160)

well, the sort of safari view of like,

Lex Fridman (1:31:56.000)

well, we're a primitive species still

Demis Hassabis (1:31:57.840)

because we're not space faring yet.

Lex Fridman (1:31:59.400)

And we're, you know, there's some kind of global,

Demis Hassabis (1:32:01.400)

like universal rule not to interfere,

Lex Fridman (1:32:03.360)

you know, Star Trek rule.

Lex Fridman (1:32:04.600)

But like, look, we can't even coordinate humans

Lex Fridman (1:32:07.360)

to deal with climate change and we're one species.

Lex Fridman (1:32:10.040)

What is the chance that of all of these different

Lex Fridman (1:32:12.400)

human civilization, you know, alien civilizations,

Demis Hassabis (1:32:14.800)

they would have the same priorities

Lex Fridman (1:32:16.760)

and agree across these kinds of matters.

Lex Fridman (1:32:20.200)

And even if that was true

Lex Fridman (1:32:21.840)

and we were in some sort of safari for our own good,

Demis Hassabis (1:32:25.040)

to me, that's not much different

Lex Fridman (1:32:26.360)

from the simulation hypothesis

Lex Fridman (1:32:27.640)

because what does it mean, the simulation hypothesis?

Lex Fridman (1:32:29.880)

I think in its most fundamental level,

Lex Fridman (1:32:31.360)

it means what we're seeing is not quite reality, right?

Lex Fridman (1:32:34.960)

It's something, there's something more deeper underlying it,

Demis Hassabis (1:32:37.760)

maybe computational.

Lex Fridman (1:32:39.120)

Now, if we were in a sort of safari park

Lex Fridman (1:32:42.600)

and everything we were seeing was a hologram

Lex Fridman (1:32:44.440)

and it was projected by the aliens or whatever,

Demis Hassabis (1:32:46.520)

that to me is not much different

Lex Fridman (1:32:47.840)

than thinking we're inside of another universe

Lex Fridman (1:32:50.280)

because we still can't see true reality, right?

Lex Fridman (1:32:53.160)

I mean, there's other explanations.

Demis Hassabis (1:32:55.120)

It could be that the way they're communicating

Lex Fridman (1:32:58.000)

is just fundamentally different,

Demis Hassabis (1:32:59.280)

that we're too dumb to understand the much better methods

Lex Fridman (1:33:02.440)

of communication they have.

Demis Hassabis (1:33:03.840)

It could be, I mean, it's silly to say,

Lex Fridman (1:33:06.600)

but our own thoughts could be the methods

Demis Hassabis (1:33:09.960)

by which they're communicating.

Lex Fridman (1:33:11.200)

Like the place from which our ideas,

Demis Hassabis (1:33:13.240)

writers talk about this, like the muse.

Lex Fridman (1:33:15.160)

Yeah.

Demis Hassabis (1:33:17.120)

I mean, it sounds like very kind of wild,

Lex Fridman (1:33:20.880)

but it could be thoughts.

Demis Hassabis (1:33:22.160)

It could be some interactions with our mind

Lex Fridman (1:33:24.600)

that we think are originating from us

Demis Hassabis (1:33:27.840)

is actually something that is coming

Lex Fridman (1:33:31.440)

from other life forms elsewhere.

Demis Hassabis (1:33:33.040)

Consciousness itself might be that.

Lex Fridman (1:33:34.880)

It could be, but I don't see any sensible argument

Demis Hassabis (1:33:37.360)

to the why would all of the alien species

Lex Fridman (1:33:40.560)

behave in this way?

Demis Hassabis (1:33:41.600)

Yeah, some of them will be more primitive.

Lex Fridman (1:33:43.200)

They will be close to our level.

Demis Hassabis (1:33:44.920)

There should be a whole sort of normal distribution

Lex Fridman (1:33:47.760)

of these things, right?

Demis Hassabis (1:33:48.680)

Some would be aggressive.

Lex Fridman (1:33:49.640)

Some would be curious.

Demis Hassabis (1:33:52.120)

Others would be very historical and philosophical

Lex Fridman (1:33:55.560)

because maybe they're a million years older than us,

Lex Fridman (1:33:58.080)

but it's not, it shouldn't be like,

Lex Fridman (1:34:00.160)

I mean, one alien civilization might be like that,

Demis Hassabis (1:34:03.000)

communicating thoughts and others,

Lex Fridman (1:34:04.200)

but I don't see why potentially the hundreds there should be

Lex Fridman (1:34:07.720)

would be uniform in this way, right?

Lex Fridman (1:34:10.040)

It could be a violent dictatorship that the people,

Demis Hassabis (1:34:13.040)

the alien civilizations that become successful

Lex Fridman (1:34:20.560)

gain the ability to be destructive,

Demis Hassabis (1:34:23.080)

an order of magnitude more destructive,

Lex Fridman (1:34:26.000)

but of course the sad thought,

Demis Hassabis (1:34:29.880)

well, either humans are very special.

Lex Fridman (1:34:32.640)

We took a lot of leaps that arrived

Demis Hassabis (1:34:35.480)

at what it means to be human.

Lex Fridman (1:34:38.600)

There's a question there, which was the hardest,

Demis Hassabis (1:34:41.160)

which was the most special,

Lex Fridman (1:34:42.720)

but also if others have reached this level

Lex Fridman (1:34:45.200)

and maybe many others have reached this level,

Lex Fridman (1:34:47.680)

the great filter that prevented them from going farther

Demis Hassabis (1:34:52.680)

to becoming a multi planetary species

Lex Fridman (1:34:54.800)

or reaching out into the stars.

Lex Fridman (1:34:57.520)

And those are really important questions for us,

Lex Fridman (1:34:59.960)

whether there's other alien civilizations out there or not,

Demis Hassabis (1:35:04.720)

this is very useful for us to think about.

Lex Fridman (1:35:06.960)

If we destroy ourselves, how will we do it?

Lex Fridman (1:35:10.240)

And how easy is it to do?

Lex Fridman (1:35:11.960)

Yeah, well, these are big questions

Lex Fridman (1:35:14.160)

and I've thought about these a lot,

Lex Fridman (1:35:15.320)

but the interesting thing is that if we're alone,

Demis Hassabis (1:35:19.600)

that's somewhat comforting from the great filter perspective

Lex Fridman (1:35:22.000)

because it probably means the great filters were passed us.

Lex Fridman (1:35:25.240)

And I'm pretty sure they are.

Lex Fridman (1:35:26.240)

So going back to your origin of life question,

Demis Hassabis (1:35:29.040)

there are some incredible things

Lex Fridman (1:35:30.560)

that no one knows how happened,

Demis Hassabis (1:35:31.760)

like obviously the first life form from chemical soup,

Lex Fridman (1:35:35.240)

that seems pretty hard,

Lex Fridman (1:35:36.800)

but I would guess the multicellular,

Lex Fridman (1:35:38.720)

I wouldn't be that surprised if we saw single cell

Demis Hassabis (1:35:42.120)

sort of life forms elsewhere, bacteria type things,

Lex Fridman (1:35:45.440)

but multicellular life seems incredibly hard,

Demis Hassabis (1:35:48.000)

that step of capturing mitochondria

Lex Fridman (1:35:50.200)

and then sort of using that as part of yourself,

Demis Hassabis (1:35:53.120)

you know, when you've just eaten it.

Lex Fridman (1:35:53.960)

Would you say that's the biggest, the most,

Demis Hassabis (1:35:57.560)

like if you had to choose one sort of,

Lex Fridman (1:36:01.400)

Hitchhiker's Galaxy, one sentence summary of like,

Demis Hassabis (1:36:04.400)

oh, those clever creatures did this,

Lex Fridman (1:36:07.280)

that would be the multicellular.

Demis Hassabis (1:36:08.280)

I think that was probably the one that's the biggest.

Lex Fridman (1:36:10.600)

I mean, there's a great book

Demis Hassabis (1:36:11.440)

called The 10 Great Inventions of Evolution by Nick Lane,

Lex Fridman (1:36:14.760)

and he speculates on 10 of these, you know,

Lex Fridman (1:36:17.440)

what could be great filters.

Lex Fridman (1:36:19.800)

I think that's one.

Demis Hassabis (1:36:21.000)

I think the advent of intelligence

Lex Fridman (1:36:23.880)

and conscious intelligence and in order, you know,

Demis Hassabis (1:36:26.360)

to us to be able to do science and things like that

Lex Fridman (1:36:28.600)

is huge as well.

Demis Hassabis (1:36:29.880)

I mean, it's only evolved once as far as, you know,

Lex Fridman (1:36:32.840)

in Earth history.

Lex Fridman (1:36:34.880)

So that would be a later candidate,

Lex Fridman (1:36:37.160)

but there's certainly for the early candidates,

Demis Hassabis (1:36:39.160)

I think multicellular life forms is huge.

Lex Fridman (1:36:41.440)

By the way, what it's interesting to ask you,

Demis Hassabis (1:36:43.560)

if you can hypothesize about

Lex Fridman (1:36:45.760)

what is the origin of intelligence?

Lex Fridman (1:36:48.000)

Is it that we started cooking meat over fire?

Lex Fridman (1:36:53.640)

Is it that we somehow figured out

Lex Fridman (1:36:55.520)

that we could be very powerful when we started collaborating?

Lex Fridman (1:36:58.120)

So cooperation between our ancestors

Lex Fridman (1:37:03.560)

so that we can overthrow the alpha male.

Lex Fridman (1:37:07.040)

What is it, Richard?

Demis Hassabis (1:37:07.880)

I talked to Richard Ranham,

Lex Fridman (1:37:08.920)

who thinks we're all just beta males

Demis Hassabis (1:37:10.760)

who figured out how to collaborate to defeat the one,

Lex Fridman (1:37:13.840)

the dictator, the authoritarian alpha male

Demis Hassabis (1:37:16.360)

that controlled the tribe.

Lex Fridman (1:37:18.360)

Is there other explanation?

Demis Hassabis (1:37:20.120)

Was there 2001 Space Odyssey type of monolith

Lex Fridman (1:37:24.080)

that came down to Earth?

Demis Hassabis (1:37:25.280)

Well, I think all of those things

Lex Fridman (1:37:27.480)

you suggested are good candidates,

Lex Fridman (1:37:28.640)

fire and cooking, right?

Lex Fridman (1:37:30.680)

So that's clearly important for energy efficiency,

Demis Hassabis (1:37:35.520)

cooking our meat and then being able to be more efficient

Lex Fridman (1:37:39.600)

about eating it and consuming the energy.

Demis Hassabis (1:37:42.840)

I think that's huge and then utilizing fire and tools

Lex Fridman (1:37:45.720)

I think you're right about the tribal cooperation aspects

Lex Fridman (1:37:48.640)

and probably language is part of that

Lex Fridman (1:37:51.040)

because probably that's what allowed us

Demis Hassabis (1:37:52.400)

to outcompete Neanderthals

Lex Fridman (1:37:53.680)

and perhaps less cooperative species.

Lex Fridman (1:37:56.160)

So that may be the case.

Lex Fridman (1:37:58.760)

Tool making, spears, axes, I think that let us,

Demis Hassabis (1:38:02.400)

I mean, I think it's pretty clear now

Lex Fridman (1:38:03.920)

that humans were responsible

Demis Hassabis (1:38:05.080)

for a lot of the extinctions of megafauna,

Lex Fridman (1:38:07.840)

especially in the Americas when humans arrived.

Lex Fridman (1:38:10.840)

So you can imagine once you discover tool usage

Lex Fridman (1:38:14.440)

how powerful that would have been

Lex Fridman (1:38:15.800)

and how scary for animals.

Lex Fridman (1:38:17.520)

So I think all of those could have been explanations for it.

Demis Hassabis (1:38:20.720)

The interesting thing is that it's a bit

Lex Fridman (1:38:22.760)

like general intelligence too,

Demis Hassabis (1:38:24.080)

is it's very costly to begin with to have a brain

Lex Fridman (1:38:28.040)

and especially a general purpose brain

Demis Hassabis (1:38:29.520)

rather than a special purpose one

Lex Fridman (1:38:30.920)

because the amount of energy our brains use,

Demis Hassabis (1:38:32.320)

I think it's like 20% of the body's energy

Lex Fridman (1:38:34.400)

and it's massive and even your thinking chest,

Demis Hassabis (1:38:36.680)

one of the funny things that we used to say

Lex Fridman (1:38:39.000)

is it's as much as a racing driver uses

Demis Hassabis (1:38:41.560)

for a whole Formula One race,

Lex Fridman (1:38:43.560)

just playing a game of serious high level chess,

Demis Hassabis (1:38:46.360)

which you wouldn't think just sitting there

Lex Fridman (1:38:49.280)

because the brain's using so much energy.

Lex Fridman (1:38:52.040)

So in order for an animal, an organism to justify that,

Lex Fridman (1:38:54.760)

there has to be a huge payoff.

Lex Fridman (1:38:57.840)

And the problem with half a brain

Lex Fridman (1:39:00.280)

or half intelligence, say an IQs of like a monkey brain,

Demis Hassabis (1:39:06.720)

it's not clear you can justify that evolutionary

Lex Fridman (1:39:10.240)

until you get to the human level brain.

Lex Fridman (1:39:12.440)

And so, but how do you do that jump?

Lex Fridman (1:39:14.720)

It's very difficult,

Demis Hassabis (1:39:15.560)

which is why I think it has only been done once

Lex Fridman (1:39:17.120)

from the sort of specialized brains that you see in animals

Demis Hassabis (1:39:19.800)

to this sort of general purpose,

Lex Fridman (1:39:22.480)

chewing powerful brains that humans have

Lex Fridman (1:39:26.200)

and which allows us to invent the modern world.

Lex Fridman (1:39:29.800)

And it takes a lot to cross that barrier.

Lex Fridman (1:39:33.600)

And I think we've seen the same with AI systems,

Lex Fridman (1:39:35.600)

which is that maybe until very recently,

Demis Hassabis (1:39:38.160)

it's always been easier to craft a specific solution

Lex Fridman (1:39:40.880)

to a problem like chess than it has been

Demis Hassabis (1:39:43.040)

to build a general learning system

Lex Fridman (1:39:44.480)

that could potentially do many things.

Demis Hassabis (1:39:46.280)

Cause initially that system will be way worse

Lex Fridman (1:39:49.480)

than less efficient than the specialized system.

Lex Fridman (1:39:52.120)

So one of the interesting quirks of the human mind

Lex Fridman (1:39:55.880)

of this evolved system is that it appears to be conscious.

Demis Hassabis (1:40:01.320)

This thing that we don't quite understand,

Lex Fridman (1:40:02.920)

but it seems very special is ability

Demis Hassabis (1:40:07.360)

to have a subjective experience

Lex Fridman (1:40:08.760)

that it feels like something to eat a cookie,

Demis Hassabis (1:40:12.280)

the deliciousness of it or see a color

Lex Fridman (1:40:14.320)

and that kind of stuff.

Lex Fridman (1:40:15.560)

Do you think in order to solve intelligence,

Lex Fridman (1:40:17.960)

we also need to solve consciousness along the way?

Lex Fridman (1:40:20.680)

Do you think AGI systems need to have consciousness

Lex Fridman (1:40:23.920)

in order to be truly intelligent?

Demis Hassabis (1:40:28.000)

Yeah, we thought about this a lot actually.

Lex Fridman (1:40:29.640)

And I think that my guess is that consciousness

Lex Fridman (1:40:33.440)

and intelligence are double dissociable.

Lex Fridman (1:40:35.800)

So you can have one without the other both ways.

Lex Fridman (1:40:38.360)

And I think you can see that with consciousness

Lex Fridman (1:40:40.920)

in that I think some animals and pets,

Demis Hassabis (1:40:44.160)

if you have a pet dog or something like that,

Lex Fridman (1:40:46.240)

you can see some of the higher animals and dolphins,

Demis Hassabis (1:40:48.560)

things like that have self awareness

Lex Fridman (1:40:51.680)

and are very sociable, seem to dream.

Demis Hassabis (1:40:57.360)

A lot of the traits one would regard

Lex Fridman (1:40:59.000)

as being kind of conscious and self aware,

Lex Fridman (1:41:02.800)

but yet they're not that smart, right?

Lex Fridman (1:41:05.080)

So they're not that intelligent

Demis Hassabis (1:41:06.320)

by say IQ standards or something like that.

Lex Fridman (1:41:08.920)

Yeah, it's also possible that our understanding

Demis Hassabis (1:41:11.080)

of intelligence is flawed, like putting an IQ to it.

Lex Fridman (1:41:14.920)

Maybe the thing that a dog can do

Demis Hassabis (1:41:17.360)

is actually gone very far along the path of intelligence

Lex Fridman (1:41:20.640)

and we humans are just able to play chess

Lex Fridman (1:41:23.240)

and maybe write poems.

Lex Fridman (1:41:24.840)

Right, but if we go back to the idea of AGI

Lex Fridman (1:41:27.040)

and general intelligence, dogs are very specialized, right?

Lex Fridman (1:41:29.480)

Most animals are pretty specialized.

Demis Hassabis (1:41:30.920)

They can be amazing at what they do,

Lex Fridman (1:41:32.360)

but they're like kind of elite sports people or something,

Demis Hassabis (1:41:35.600)

right, so they do one thing extremely well

Lex Fridman (1:41:38.040)

because their entire brain is optimized.

Demis Hassabis (1:41:40.080)

They have somehow convinced the entirety

Lex Fridman (1:41:41.880)

of the human population to feed them and service them.

Lex Fridman (1:41:44.520)

So in some way they're controlling.

Lex Fridman (1:41:46.400)

Yes, exactly.

Lex Fridman (1:41:47.240)

Well, we co evolved to some crazy degree, right?

Lex Fridman (1:41:50.120)

Including the way the dogs even wag their tails

Lex Fridman (1:41:53.800)

and twitch their noses, right?

Lex Fridman (1:41:55.160)

We find inextricably cute.

Lex Fridman (1:41:58.640)

But I think you can also see intelligence on the other side.

Lex Fridman (1:42:01.840)

So systems like artificial systems

Demis Hassabis (1:42:03.800)

that are amazingly smart at certain things

Lex Fridman (1:42:07.240)

like maybe playing go and chess and other things,

Lex Fridman (1:42:09.800)

but they don't feel at all in any shape or form conscious

Lex Fridman (1:42:13.440)

in the way that you do to me or I do to you.

Lex Fridman (1:42:17.240)

And I think actually building AI

Lex Fridman (1:42:21.440)

is these intelligent constructs

Demis Hassabis (1:42:24.200)

is one of the best ways to explore

Lex Fridman (1:42:25.920)

the mystery of consciousness, to break it down

Demis Hassabis (1:42:28.000)

because we're gonna have devices

Lex Fridman (1:42:31.200)

that are pretty smart at certain things

Demis Hassabis (1:42:34.440)

or capable at certain things,

Lex Fridman (1:42:36.200)

but potentially won't have any semblance

Demis Hassabis (1:42:39.160)

of self awareness or other things.

Lex Fridman (1:42:40.800)

And in fact, I would advocate if there's a choice,

Demis Hassabis (1:42:43.880)

building systems in the first place,

Lex Fridman (1:42:45.680)

AI systems that are not conscious to begin with

Demis Hassabis (1:42:48.640)

are just tools until we understand them better

Lex Fridman (1:42:52.440)

and the capabilities better.

Lex Fridman (1:42:53.960)

So on that topic, just not as the CEO of DeepMind,

Lex Fridman (1:42:58.320)

just as a human being, let me ask you

Demis Hassabis (1:43:00.880)

about this one particular anecdotal evidence

Lex Fridman (1:43:03.480)

of the Google engineer who made a comment

Demis Hassabis (1:43:07.080)

or believed that there's some aspect of a language model,

Lex Fridman (1:43:11.800)

the Lambda language model that exhibited sentience.

Lex Fridman (1:43:15.960)

So you said you believe there might be a responsibility

Lex Fridman (1:43:18.440)

to build systems that are not sentient.

Lex Fridman (1:43:21.120)

And this experience of a particular engineer,

Lex Fridman (1:43:23.560)

I think I'd love to get your general opinion

Demis Hassabis (1:43:25.880)

on this kind of thing, but I think it will happen

Lex Fridman (1:43:28.000)

more and more and more, which not when engineers,

Lex Fridman (1:43:31.480)

but when people out there that don't have

Lex Fridman (1:43:33.120)

an engineering background start interacting

Demis Hassabis (1:43:34.760)

with increasingly intelligent systems,

Lex Fridman (1:43:37.120)

we anthropomorphize them.

Demis Hassabis (1:43:38.960)

They start to have deep, impactful interactions with us

Lex Fridman (1:43:44.680)

in a way that we miss them when they're gone.

Lex Fridman (1:43:47.920)

And we sure as heck feel like they're living entities,

Lex Fridman (1:43:51.960)

self aware entities, and maybe even

Demis Hassabis (1:43:54.200)

we project sentience onto them.

Lex Fridman (1:43:55.960)

So what's your thought about this particular system?

Lex Fridman (1:44:01.320)

Have you ever met a language model that's sentient?

Lex Fridman (1:44:04.600)

No, no.

Lex Fridman (1:44:06.320)

What do you make of the case of when you kind of feel

Lex Fridman (1:44:10.200)

that there's some elements of sentience to the system?

Demis Hassabis (1:44:12.920)

Yeah, so this is an interesting question

Lex Fridman (1:44:15.040)

and obviously a very fundamental one.

Lex Fridman (1:44:17.760)

So the first thing to say is I think that none

Lex Fridman (1:44:20.280)

of the systems we have today, I would say,

Demis Hassabis (1:44:22.200)

even have one iota of semblance

Lex Fridman (1:44:25.080)

of consciousness or sentience.

Demis Hassabis (1:44:26.320)

That's my personal feeling interacting with them every day.

Lex Fridman (1:44:29.720)

So I think this way premature to be discussing

Lex Fridman (1:44:32.400)

what that engineer talked about.

Lex Fridman (1:44:34.160)

I think at the moment it's more of a projection

Demis Hassabis (1:44:36.480)

of the way our own minds work,

Lex Fridman (1:44:37.840)

which is to see sort of purpose and direction

Demis Hassabis (1:44:43.120)

in almost anything that we, you know,

Lex Fridman (1:44:44.600)

our brains are trained to interpret agency,

Demis Hassabis (1:44:48.200)

basically in things, even inanimate things sometimes.

Lex Fridman (1:44:52.280)

And of course with a language system,

Demis Hassabis (1:44:54.880)

because language is so fundamental to intelligence,

Lex Fridman (1:44:57.080)

that's going to be easy for us to anthropomorphize that.

Demis Hassabis (1:45:00.440)

I mean, back in the day, even the first, you know,

Lex Fridman (1:45:03.840)

the dumbest sort of template chatbots ever,

Demis Hassabis (1:45:05.800)

Eliza and the ilk of the original chatbots

Lex Fridman (1:45:09.200)

back in the sixties fooled some people

Lex Fridman (1:45:11.160)

under certain circumstances, right?

Lex Fridman (1:45:12.600)

It pretended to be a psychologist.

Lex Fridman (1:45:14.040)

So just basically rabbit back to you

Lex Fridman (1:45:16.080)

the same question you asked it back to you.

Lex Fridman (1:45:19.240)

And some people believe that.

Lex Fridman (1:45:21.320)

So I don't think we can, this is why I think

Demis Hassabis (1:45:23.280)

the Turing test is a little bit flawed as a formal test

Lex Fridman (1:45:25.440)

because it depends on the sophistication of the judge,

Demis Hassabis (1:45:29.240)

whether or not they are qualified to make that distinction.

Lex Fridman (1:45:33.280)

So I think we should talk to, you know,

Demis Hassabis (1:45:36.800)

the top philosophers about this,

Lex Fridman (1:45:38.320)

people like Daniel Dennett and David Chalmers and others

Demis Hassabis (1:45:41.160)

who've obviously thought deeply about consciousness.

Lex Fridman (1:45:43.680)

Of course, consciousness itself hasn't been well,

Demis Hassabis (1:45:46.040)

there's no agreed definition.

Lex Fridman (1:45:47.760)

If I was to, you know, speculate about that, you know,

Demis Hassabis (1:45:52.160)

I kind of, the working definition I like is

Lex Fridman (1:45:55.120)

it's the way information feels when it gets processed.

Demis Hassabis (1:45:58.080)

I think maybe Max Tegmark came up with that.

Lex Fridman (1:46:00.160)

I like that idea.

Demis Hassabis (1:46:01.040)

I don't know if it helps us get towards

Lex Fridman (1:46:02.280)

any more operational thing,

Lex Fridman (1:46:03.920)

but I think it's a nice way of viewing it.

Lex Fridman (1:46:07.800)

I think we can obviously see from neuroscience

Demis Hassabis (1:46:10.000)

certain prerequisites that are required,

Lex Fridman (1:46:11.720)

like self awareness, I think is necessary,

Lex Fridman (1:46:14.440)

but not sufficient component.

Lex Fridman (1:46:16.080)

This idea of a self and other

Lex Fridman (1:46:18.160)

and set of coherent preferences

Lex Fridman (1:46:20.520)

that are coherent over time.

Demis Hassabis (1:46:22.480)

You know, these things are maybe memory.

Lex Fridman (1:46:24.800)

These things are probably needed

Demis Hassabis (1:46:26.200)

for a sentient or conscious being.

Lex Fridman (1:46:29.320)

But the reason, the difficult thing,

Demis Hassabis (1:46:31.160)

I think for us when we get,

Lex Fridman (1:46:32.240)

and I think this is a really interesting

Demis Hassabis (1:46:33.400)

philosophical debate is when we get closer to AGI

Lex Fridman (1:46:37.280)

and, you know, and much more powerful systems

Demis Hassabis (1:46:40.680)

than we have today,

Lex Fridman (1:46:42.240)

how are we going to make this judgment?

Lex Fridman (1:46:44.440)

And one way, which is the Turing test

Lex Fridman (1:46:46.960)

is sort of a behavioral judgment,

Demis Hassabis (1:46:48.640)

is the system exhibiting all the behaviors

Lex Fridman (1:46:52.080)

that a human sentient or a sentient being would exhibit?

Lex Fridman (1:46:56.880)

Is it answering the right questions?

Lex Fridman (1:46:58.160)

Is it saying the right things?

Lex Fridman (1:46:59.160)

Is it indistinguishable from a human?

Lex Fridman (1:47:01.960)

And so on.

Lex Fridman (1:47:03.360)

But I think there's a second thing

Lex Fridman (1:47:05.760)

that makes us as humans regard each other as sentient,

Lex Fridman (1:47:09.040)

right?

Lex Fridman (1:47:09.880)

Why do we think this?

Lex Fridman (1:47:10.920)

And I debated this with Daniel Dennett.

Lex Fridman (1:47:12.720)

And I think there's a second reason

Demis Hassabis (1:47:13.880)

that's often overlooked,

Lex Fridman (1:47:15.600)

which is that we're running on the same substrate, right?

Lex Fridman (1:47:18.280)

So if we're exhibiting the same behavior,

Lex Fridman (1:47:21.120)

more or less as humans,

Lex Fridman (1:47:22.680)

and we're running on the same, you know,

Lex Fridman (1:47:24.400)

carbon based biological substrate,

Demis Hassabis (1:47:26.200)

the squishy, you know, few pounds of flesh in our skulls,

Lex Fridman (1:47:29.560)

then the most parsimonious, I think, explanation

Lex Fridman (1:47:32.800)

is that you're feeling the same thing as I'm feeling, right?

Lex Fridman (1:47:35.520)

But we will never have that second part,

Lex Fridman (1:47:37.840)

the substrate equivalence with a machine, right?

Lex Fridman (1:47:41.200)

So we will have to only judge based on the behavior.

Lex Fridman (1:47:43.880)

And I think the substrate equivalence

Lex Fridman (1:47:45.920)

is a critical part of why we make assumptions

Demis Hassabis (1:47:48.200)

that we're conscious.

Lex Fridman (1:47:49.080)

And in fact, even with animals, high level animals,

Lex Fridman (1:47:51.680)

why we think they might be,

Lex Fridman (1:47:52.680)

because they're exhibiting some of the behaviors

Demis Hassabis (1:47:54.160)

we would expect from a sentient animal.

Lex Fridman (1:47:55.880)

And we know they're made of the same things,

Demis Hassabis (1:47:57.600)

biological neurons.

Lex Fridman (1:47:58.640)

So we're gonna have to come up with explanations

Demis Hassabis (1:48:02.880)

or models of the gap between substrate differences,

Lex Fridman (1:48:06.320)

between machines and humans

Demis Hassabis (1:48:08.040)

to get anywhere beyond the behavioral.

Lex Fridman (1:48:10.840)

But to me, sort of the practical question

Demis Hassabis (1:48:12.920)

is very interesting and very important.

Lex Fridman (1:48:16.040)

When you have millions, perhaps billions of people

Demis Hassabis (1:48:18.640)

believing that you have a sentient AI,

Lex Fridman (1:48:20.800)

believing what that Google engineer believed,

Demis Hassabis (1:48:24.040)

which I just see as an obvious, very near term future thing,

Lex Fridman (1:48:28.760)

certainly on the path to AGI,

Lex Fridman (1:48:31.160)

how does that change the world?

Lex Fridman (1:48:33.160)

What's the responsibility of the AI system

Lex Fridman (1:48:35.240)

to help those millions of people?

Lex Fridman (1:48:38.160)

And also what's the ethical thing?

Demis Hassabis (1:48:39.760)

Because you can make a lot of people happy

Lex Fridman (1:48:44.760)

by creating a meaningful, deep experience

Demis Hassabis (1:48:48.040)

with a system that's faking it before it makes it.

Lex Fridman (1:48:52.800)

And I don't, are we the right,

Lex Fridman (1:48:56.120)

who is to say what's the right thing to do?

Lex Fridman (1:48:59.720)

Should AI always be tools?

Lex Fridman (1:49:01.920)

Why are we constraining AI to always be tools

Lex Fridman (1:49:05.880)

as opposed to friends?

Demis Hassabis (1:49:07.680)

Yeah, I think, well, I mean, these are fantastic questions

Lex Fridman (1:49:11.800)

and also critical ones.

Lex Fridman (1:49:13.840)

And we've been thinking about this

Lex Fridman (1:49:16.240)

since the start of DeepMind and before that,

Demis Hassabis (1:49:18.080)

because we plan for success

Lex Fridman (1:49:19.560)

and however remote that looked like back in 2010.

Lex Fridman (1:49:24.640)

And we've always had sort of these ethical considerations

Lex Fridman (1:49:26.960)

as fundamental at DeepMind.

Lex Fridman (1:49:29.400)

And my current thinking on the language models

Lex Fridman (1:49:32.000)

and large models is they're not ready,

Demis Hassabis (1:49:33.920)

we don't understand them well enough yet.

Lex Fridman (1:49:36.480)

And in terms of analysis tools and guard rails,

Lex Fridman (1:49:40.240)

what they can and can't do and so on,

Lex Fridman (1:49:42.080)

to deploy them at scale, because I think,

Demis Hassabis (1:49:45.440)

there are big, still ethical questions

Lex Fridman (1:49:46.840)

like should an AI system always announce

Lex Fridman (1:49:48.640)

that it is an AI system to begin with?

Lex Fridman (1:49:50.600)

Probably yes.

Lex Fridman (1:49:52.800)

What do you do about answering those philosophical questions

Lex Fridman (1:49:55.520)

about the feelings people may have about AI systems,

Lex Fridman (1:49:58.800)

perhaps incorrectly attributed?

Lex Fridman (1:50:00.760)

So I think there's a whole bunch of research

Demis Hassabis (1:50:02.840)

that needs to be done first to responsibly,

Lex Fridman (1:50:06.040)

before you can responsibly deploy these systems at scale.

Demis Hassabis (1:50:09.120)

That will be at least be my current position.

Lex Fridman (1:50:12.080)

Over time, I'm very confident we'll have those tools

Demis Hassabis (1:50:15.080)

like interpretability questions and analysis questions.

Lex Fridman (1:50:20.680)

And then with the ethical quandary,

Demis Hassabis (1:50:23.200)

I think there it's important to look beyond just science.

Lex Fridman (1:50:28.520)

That's why I think philosophy, social sciences,

Demis Hassabis (1:50:31.440)

even theology, other things like that come into it,

Lex Fridman (1:50:34.440)

where arts and humanities,

Lex Fridman (1:50:37.120)

what does it mean to be human and the spirit of being human

Lex Fridman (1:50:40.320)

and to enhance that and the human condition, right?

Lex Fridman (1:50:43.680)

And allow us to experience things

Lex Fridman (1:50:45.080)

we could never experience before

Lex Fridman (1:50:46.400)

and improve the overall human condition

Lex Fridman (1:50:49.080)

and humanity overall, get radical abundance,

Demis Hassabis (1:50:51.640)

solve many scientific problems, solve disease.

Lex Fridman (1:50:54.120)

So this is the era I think, this is the amazing era

Demis Hassabis (1:50:56.560)

I think we're heading into if we do it right.

Lex Fridman (1:50:59.480)

But we've got to be careful.

Demis Hassabis (1:51:00.800)

We've already seen with things like social media,

Lex Fridman (1:51:02.680)

how dual use technologies can be misused by,

Demis Hassabis (1:51:05.920)

firstly, by bad actors or naive actors or crazy actors,

Lex Fridman (1:51:12.040)

right, so there's that set of just the common

Demis Hassabis (1:51:14.120)

or garden misuse of existing dual use technology.

Lex Fridman (1:51:18.000)

And then of course, there's an additional thing

Demis Hassabis (1:51:20.960)

that has to be overcome with AI

Lex Fridman (1:51:21.960)

that eventually it may have its own agency.

Lex Fridman (1:51:24.480)

So it could be good or bad in and of itself.

Lex Fridman (1:51:28.720)

So I think these questions have to be approached

Demis Hassabis (1:51:31.480)

very carefully using the scientific method, I would say,

Lex Fridman (1:51:35.360)

in terms of hypothesis generation, careful control testing,

Demis Hassabis (1:51:38.680)

not live A, B testing out in the world,

Lex Fridman (1:51:40.680)

because with powerful technologies like AI,

Demis Hassabis (1:51:44.400)

if something goes wrong, it may cause a lot of harm

Lex Fridman (1:51:47.640)

before you can fix it.

Demis Hassabis (1:51:49.120)

It's not like an imaging app or game app

Lex Fridman (1:51:52.000)

where if something goes wrong, it's relatively easy to fix

Lex Fridman (1:51:56.160)

and the harm is relatively small.

Lex Fridman (1:51:57.960)

So I think it comes with the usual cliche of,

Demis Hassabis (1:52:02.720)

like with a lot of power comes a lot of responsibility.

Lex Fridman (1:52:05.240)

And I think that's the case here with things like AI,

Demis Hassabis (1:52:07.800)

given the enormous opportunity in front of us.

Lex Fridman (1:52:11.040)

And I think we need a lot of voices

Lex Fridman (1:52:14.040)

and as many inputs into things like the design

Lex Fridman (1:52:17.160)

of the systems and the values they should have

Lex Fridman (1:52:19.880)

and what goals should they be put to.

Lex Fridman (1:52:22.400)

I think as wide a group of voices as possible

Demis Hassabis (1:52:24.560)

beyond just the technologists is needed to input into that

Lex Fridman (1:52:27.720)

and to have a say in that,

Demis Hassabis (1:52:29.080)

especially when it comes to deployment of these systems,

Lex Fridman (1:52:31.840)

which is when the rubber really hits the road,

Demis Hassabis (1:52:33.440)

it really affects the general person in the street

Lex Fridman (1:52:35.440)

rather than fundamental research.

Lex Fridman (1:52:37.400)

And that's why I say, I think as a first step,

Lex Fridman (1:52:40.240)

it would be better if we have the choice

Demis Hassabis (1:52:42.360)

to build these systems as tools to give,

Lex Fridman (1:52:45.120)

and I'm not saying that they should never go beyond tools

Demis Hassabis (1:52:47.960)

because of course the potential is there

Lex Fridman (1:52:50.360)

for it to go way beyond just tools.

Lex Fridman (1:52:52.960)

But I think that would be a good first step

Lex Fridman (1:52:55.800)

in order for us to allow us to carefully experiment

Lex Fridman (1:52:58.880)

and understand what these things can do.

Lex Fridman (1:53:01.000)

So the leap between tool, the sentient entity being

Demis Hassabis (1:53:05.800)

is one we should take very careful of.

Lex Fridman (1:53:08.280)

Let me ask a dark personal question.

Lex Fridman (1:53:11.120)

So you're one of the most brilliant people

Lex Fridman (1:53:13.480)

in the AI community, you're also one of the most kind

Lex Fridman (1:53:16.800)

and if I may say sort of loved people in the community.

Lex Fridman (1:53:20.880)

That said, creation of a super intelligent AI system

Demis Hassabis (1:53:25.880)

would be one of the most powerful things in the world,

Lex Fridman (1:53:32.720)

tools or otherwise.

Lex Fridman (1:53:34.840)

And again, as the old saying goes, power corrupts

Lex Fridman (1:53:38.400)

and absolute power corrupts absolutely.

Demis Hassabis (1:53:41.640)

You are likely to be one of the people,

Lex Fridman (1:53:47.280)

I would say probably the most likely person

Demis Hassabis (1:53:50.320)

to be in the control of such a system.

Lex Fridman (1:53:53.280)

Do you think about the corrupting nature of power

Demis Hassabis (1:53:57.120)

when you talk about these kinds of systems

Lex Fridman (1:53:59.560)

that as all dictators and people have caused atrocities

Demis Hassabis (1:54:04.920)

in the past, always think they're doing good,

Lex Fridman (1:54:07.760)

but they don't do good because the power

Demis Hassabis (1:54:10.400)

has polluted their mind about what is good

Lex Fridman (1:54:12.560)

and what is evil.

Lex Fridman (1:54:13.720)

Do you think about this stuff

Lex Fridman (1:54:14.840)

or are we just focused on language model?

Demis Hassabis (1:54:16.440)

No, I think about them all the time

Lex Fridman (1:54:18.760)

and I think what are the defenses against that?

Demis Hassabis (1:54:22.360)

I think one thing is to remain very grounded

Lex Fridman (1:54:24.840)

and sort of humble, no matter what you do or achieve.

Lex Fridman (1:54:28.800)

And I try to do that, my best friends

Lex Fridman (1:54:31.160)

are still my set of friends

Demis Hassabis (1:54:32.200)

from my undergraduate Cambridge days,

Lex Fridman (1:54:34.680)

my family's and friends are very important.

Demis Hassabis (1:54:39.280)

I've always, I think trying to be a multidisciplinary person,

Lex Fridman (1:54:42.360)

it helps to keep you humble

Demis Hassabis (1:54:43.760)

because no matter how good you are at one topic,

Lex Fridman (1:54:45.880)

someone will be better than you at that.

Lex Fridman (1:54:47.560)

And always relearning a new topic again from scratch

Lex Fridman (1:54:50.920)

is a new field is very humbling, right?

Lex Fridman (1:54:53.320)

So for me, that's been biology over the last five years,

Lex Fridman (1:54:56.400)

huge area topic and I just love doing that,

Lex Fridman (1:55:00.200)

but it helps to keep you grounded

Lex Fridman (1:55:01.600)

like it keeps you open minded.

Lex Fridman (1:55:04.320)

And then the other important thing

Lex Fridman (1:55:06.360)

is to have a really good, amazing set of people around you

Demis Hassabis (1:55:10.040)

at your company or your organization

Lex Fridman (1:55:11.840)

who are also very ethical and grounded themselves

Lex Fridman (1:55:14.920)

and help to keep you that way.

Lex Fridman (1:55:16.840)

And then ultimately just to answer your question,

Demis Hassabis (1:55:18.880)

I hope we're gonna be a big part of birthing AI

Lex Fridman (1:55:22.000)

and that being the greatest benefit to humanity

Demis Hassabis (1:55:24.440)

of any tool or technology ever,

Lex Fridman (1:55:26.800)

and getting us into a world of radical abundance

Lex Fridman (1:55:29.560)

and curing diseases and solving many of the big challenges

Lex Fridman (1:55:33.960)

we have in front of us.

Lex Fridman (1:55:34.840)

And then ultimately help the ultimate flourishing

Lex Fridman (1:55:37.560)

of humanity to travel the stars

Lex Fridman (1:55:39.240)

and find those aliens if they are there.

Lex Fridman (1:55:41.160)

And if they're not there, find out why they're not there,

Lex Fridman (1:55:43.480)

what is going on here in the universe.

Lex Fridman (1:55:46.520)

This is all to come.

Lex Fridman (1:55:47.360)

And that's what I've always dreamed about.

Lex Fridman (1:55:50.720)

But I think AI is too big an idea.

Demis Hassabis (1:55:53.000)

It's not going to be,

Lex Fridman (1:55:54.760)

there'll be a certain set of pioneers who get there first.

Demis Hassabis (1:55:57.000)

I hope we're in the vanguard

Lex Fridman (1:55:58.600)

so we can influence how that goes.

Lex Fridman (1:56:00.400)

And I think it matters who builds,

Lex Fridman (1:56:02.480)

which cultures they come from and what values they have,

Demis Hassabis (1:56:06.480)

the builders of AI systems.

Lex Fridman (1:56:07.840)

Cause I think even though the AI system

Demis Hassabis (1:56:09.280)

is gonna learn for itself most of its knowledge,

Lex Fridman (1:56:11.560)

there'll be a residue in the system of the culture

Lex Fridman (1:56:14.760)

and the values of the creators of that system.

Lex Fridman (1:56:17.680)

And there's interesting questions

Demis Hassabis (1:56:18.720)

to discuss about that geopolitically.

Lex Fridman (1:56:21.600)

Different cultures,

Demis Hassabis (1:56:22.440)

we're in a more fragmented world than ever, unfortunately.

Lex Fridman (1:56:24.920)

I think in terms of global cooperation,

Demis Hassabis (1:56:27.480)

we see that in things like climate

Lex Fridman (1:56:29.240)

where we can't seem to get our act together globally

Demis Hassabis (1:56:32.000)

to cooperate on these pressing matters.

Lex Fridman (1:56:34.080)

I hope that will change over time.

Demis Hassabis (1:56:35.600)

Perhaps if we get to an era of radical abundance,

Lex Fridman (1:56:38.640)

we don't have to be so competitive anymore.

Demis Hassabis (1:56:40.440)

Maybe we can be more cooperative

Lex Fridman (1:56:42.680)

if resources aren't so scarce.

Demis Hassabis (1:56:44.360)

It's true that in terms of power corrupting

Lex Fridman (1:56:48.240)

and leading to destructive things,

Demis Hassabis (1:56:50.040)

it seems that some of the atrocities of the past happen

Lex Fridman (1:56:53.160)

when there's a significant constraint on resources.

Demis Hassabis (1:56:56.680)

I think that's the first thing.

Lex Fridman (1:56:57.560)

I don't think that's enough.

Demis Hassabis (1:56:58.400)

I think scarcity is one thing that's led to competition,

Lex Fridman (1:57:01.560)

sort of zero sum game thinking.

Demis Hassabis (1:57:03.960)

I would like us to all be in a positive sum world.

Lex Fridman (1:57:06.080)

And I think for that, you have to remove scarcity.

Demis Hassabis (1:57:08.480)

I don't think that's enough, unfortunately,

Lex Fridman (1:57:09.840)

to get world peace

Demis Hassabis (1:57:10.800)

because there's also other corrupting things

Lex Fridman (1:57:12.800)

like wanting power over people and this kind of stuff,

Demis Hassabis (1:57:15.520)

which is not necessarily satisfied by just abundance.

Lex Fridman (1:57:19.040)

But I think it will help.

Lex Fridman (1:57:22.400)

But I think ultimately, AI is not gonna be run

Lex Fridman (1:57:24.920)

by any one person or one organization.

Demis Hassabis (1:57:26.800)

I think it should belong to the world, belong to humanity.

Lex Fridman (1:57:29.600)

And I think there'll be many ways this will happen.

Lex Fridman (1:57:33.120)

And ultimately, everybody should have a say in that.

Lex Fridman (1:57:36.840)

Do you have advice for young people in high school,

Demis Hassabis (1:57:42.000)

in college, maybe if they're interested in AI

Lex Fridman (1:57:45.800)

or interested in having a big impact on the world,

Lex Fridman (1:57:50.640)

what they should do to have a career they can be proud of

Lex Fridman (1:57:53.200)

or to have a life they can be proud of?

Demis Hassabis (1:57:55.000)

I love giving talks to the next generation.

Lex Fridman (1:57:57.400)

What I say to them is actually two things.

Demis Hassabis (1:57:59.120)

I think the most important things to learn about

Lex Fridman (1:58:02.160)

and to find out about when you're young

Demis Hassabis (1:58:04.520)

is what are your true passions is first of all,

Lex Fridman (1:58:07.080)

as two things.

Demis Hassabis (1:58:07.920)

One is find your true passions.

Lex Fridman (1:58:09.720)

And I think you can do that by,

Demis Hassabis (1:58:11.720)

the way to do that is to explore as many things as possible

Lex Fridman (1:58:14.600)

when you're young and you have the time

Lex Fridman (1:58:16.520)

and you can take those risks.

Lex Fridman (1:58:19.160)

I would also encourage people to look at

Demis Hassabis (1:58:21.080)

finding the connections between things in a unique way.

Lex Fridman (1:58:24.600)

I think that's a really great way to find a passion.

Demis Hassabis (1:58:27.280)

Second thing I would say, advise is know yourself.

Lex Fridman (1:58:30.600)

So spend a lot of time understanding

Lex Fridman (1:58:33.920)

how you work best.

Lex Fridman (1:58:35.600)

Like what are the optimal times to work?

Lex Fridman (1:58:37.680)

What are the optimal ways that you study?

Lex Fridman (1:58:39.880)

What are your, how do you deal with pressure?

Demis Hassabis (1:58:42.240)

Sort of test yourself in various scenarios

Lex Fridman (1:58:44.560)

and try and improve your weaknesses,

Lex Fridman (1:58:47.240)

but also find out what your unique skills and strengths are

Lex Fridman (1:58:50.720)

and then hone those.

Lex Fridman (1:58:52.160)

So then that's what will be your super value

Lex Fridman (1:58:54.520)

in the world later on.

Lex Fridman (1:58:55.880)

And if you can then combine those two things

Lex Fridman (1:58:57.840)

and find passions that you're genuinely excited about

Demis Hassabis (1:59:01.200)

that intersect with what your unique strong skills are,

Lex Fridman (1:59:05.360)

then you're onto something incredible

Lex Fridman (1:59:07.880)

and I think you can make a huge difference in the world.

Lex Fridman (1:59:10.920)

So let me ask about know yourself.

Demis Hassabis (1:59:12.760)

This is fun.

Lex Fridman (1:59:13.600)

This is fun.

Demis Hassabis (1:59:14.440)

Quick questions about day in the life, the perfect day,

Lex Fridman (1:59:18.120)

the perfect productive day in the life of Demis's Hub.

Demis Hassabis (1:59:21.200)

Maybe these days you're, there's a lot involved.

Lex Fridman (1:59:26.200)

So maybe a slightly younger Demis's Hub

Demis Hassabis (1:59:29.000)

where you could focus on a single project maybe.

Lex Fridman (1:59:33.120)

How early do you wake up?

Lex Fridman (1:59:34.440)

Are you a night owl?

Lex Fridman (1:59:35.600)

Do you wake up early in the morning?

Lex Fridman (1:59:36.760)

What are some interesting habits?

Lex Fridman (1:59:39.160)

How many dozens of cups of coffees do you drink a day?

Lex Fridman (1:59:42.400)

What's the computer that you use?

Lex Fridman (1:59:46.320)

What's the setup?

Lex Fridman (1:59:47.160)

How many screens?

Lex Fridman (1:59:47.980)

What kind of keyboard?

Demis Hassabis (1:59:49.120)

Are we talking Emacs Vim

Lex Fridman (1:59:51.400)

or are we talking something more modern?

Lex Fridman (1:59:53.320)

So there's a bunch of those questions.

Lex Fridman (1:59:54.480)

So maybe day in the life, what's the perfect day involved?

Demis Hassabis (1:59:58.960)

Well, these days it's quite different

Lex Fridman (20:00.520)

Well, look, I think it's,

Demis Hassabis (20:01.640)

I mean, it has been a fascinating journey

Lex Fridman (20:03.720)

and especially as I think about it from,

Demis Hassabis (20:06.880)

I can understand it from both sides,

Lex Fridman (20:08.760)

both as the AI, creators of the AI,

Lex Fridman (20:13.080)

but also as a games player originally.

Lex Fridman (20:15.640)

So, it was a really interesting,

Demis Hassabis (20:17.960)

I mean, it was a fantastic, but also somewhat

Lex Fridman (20:21.160)

bittersweet moment, the AlphaGo match for me,

Demis Hassabis (20:24.680)

seeing that and being obviously heavily involved in that.

Lex Fridman (20:29.440)

But as you say, chess has been the,

Demis Hassabis (20:32.480)

I mean, Kasparov, I think rightly called it

Lex Fridman (20:34.360)

the Drosophila of intelligence, right?

Demis Hassabis (20:37.280)

So, it's sort of, I love that phrase

Lex Fridman (20:39.520)

and I think he's right because chess has been

Demis Hassabis (20:42.960)

hand in hand with AI from the beginning

Lex Fridman (20:45.360)

of the whole field, right?

Demis Hassabis (20:47.440)

So, I think every AI practitioner,

Lex Fridman (20:49.640)

starting with Turing and Claude Shannon and all those,

Demis Hassabis (20:52.480)

the sort of forefathers of the field,

Lex Fridman (20:56.320)

tried their hand at writing a chess program.

Demis Hassabis (20:58.840)

I've got original edition of Claude Shannon's

Lex Fridman (21:01.160)

first chess program, I think it was 1949,

Demis Hassabis (21:04.000)

the original sort of paper.

Lex Fridman (21:06.760)

And they all did that and Turing famously wrote

Demis Hassabis (21:09.960)

a chess program, but all the computers around them

Lex Fridman (21:12.480)

were obviously too slow to run it.

Lex Fridman (21:13.760)

So, he had to run, he had to be the computer, right?

Lex Fridman (21:16.040)

So, he literally, I think spent two or three days

Demis Hassabis (21:18.920)

running his own program by hand with pencil and paper

Lex Fridman (21:21.360)

and playing a friend of his with his chess program.

Demis Hassabis (21:24.960)

So, of course, Deep Blue was a huge moment,

Lex Fridman (21:28.560)

beating Kasparov, but actually when that happened,

Demis Hassabis (21:31.880)

I remember that very vividly, of course,

Lex Fridman (21:34.120)

because it was chess and computers and AI,

Demis Hassabis (21:36.640)

all the things I loved and I was at college at the time.

Lex Fridman (21:39.240)

But I remember coming away from that,

Demis Hassabis (21:40.800)

being more impressed by Kasparov's mind

Lex Fridman (21:43.080)

than I was by Deep Blue.

Demis Hassabis (21:44.480)

Because here was Kasparov with his human mind,

Lex Fridman (21:47.680)

not only could he play chess more or less

Demis Hassabis (21:49.400)

to the same level as this brute of a calculation machine,

Lex Fridman (21:53.160)

but of course, Kasparov can do everything else

Demis Hassabis (21:55.160)

humans can do, ride a bike, talk many languages,

Lex Fridman (21:57.480)

do politics, all the rest of the amazing things

Demis Hassabis (21:59.400)

that Kasparov does.

Lex Fridman (22:00.880)

And so, with the same brain.

Lex Fridman (22:03.160)

And yet Deep Blue, brilliant as it was at chess,

Lex Fridman (22:07.040)

it'd been hand coded for chess and actually had distilled

Demis Hassabis (22:12.040)

the knowledge of chess grandmasters into a cool program,

Lex Fridman (22:16.400)

but it couldn't do anything else.

Demis Hassabis (22:18.000)

It couldn't even play a strictly simpler game

Lex Fridman (22:20.080)

like tic tac toe.

Demis Hassabis (22:21.280)

So, something to me was missing from intelligence

Lex Fridman (22:25.880)

from that system that we would regard as intelligence.

Lex Fridman (22:28.480)

And I think it was this idea of generality

Lex Fridman (22:30.880)

and also learning.

Demis Hassabis (22:33.000)

So, and that's obviously what we tried to do with AlphaGo.

Lex Fridman (22:36.120)

Yeah, with AlphaGo and AlphaZero, MuZero,

Lex Fridman (22:38.600)

and then God and all the things that we'll get into

Lex Fridman (22:42.040)

some parts of, there's just a fascinating trajectory here.

Lex Fridman (22:45.640)

But let's just stick on chess briefly.

Lex Fridman (22:48.520)

On the human side of chess, you've proposed that

Demis Hassabis (22:51.960)

from a game design perspective,

Lex Fridman (22:53.400)

the thing that makes chess compelling as a game

Demis Hassabis (22:57.760)

is that there's a creative tension between a bishop

Lex Fridman (23:01.000)

and the knight.

Lex Fridman (23:02.960)

Can you explain this?

Lex Fridman (23:04.040)

First of all, it's really interesting to think about

Lex Fridman (23:06.440)

what makes a game compelling,

Lex Fridman (23:08.640)

makes it stick across centuries.

Demis Hassabis (23:12.000)

Yeah, I was sort of thinking about this,

Lex Fridman (23:13.480)

and actually a lot of even amazing chess players

Demis Hassabis (23:15.440)

don't think about it necessarily

Lex Fridman (23:16.840)

from a game's designer point of view.

Demis Hassabis (23:18.280)

So, it's with my game design hat on

Lex Fridman (23:20.240)

that I was thinking about this, why is chess so compelling?

Lex Fridman (23:23.080)

And I think a critical reason is the dynamicness

Lex Fridman (23:27.560)

of the different kind of chess positions you can have,

Demis Hassabis (23:30.000)

whether they're closed or open and other things,

Lex Fridman (23:32.200)

comes from the bishop and the knight.

Demis Hassabis (23:33.520)

So, if you think about how different

Lex Fridman (23:36.480)

the capabilities of the bishop and knight are

Demis Hassabis (23:39.240)

in terms of the way they move,

Lex Fridman (23:40.880)

and then somehow chess has evolved

Demis Hassabis (23:43.080)

to balance those two capabilities more or less equally.

Lex Fridman (23:46.080)

So, they're both roughly worth three points each.

Demis Hassabis (23:48.720)

So, you think that dynamics is always there

Lex Fridman (23:50.560)

and then the rest of the rules

Demis Hassabis (23:51.640)

are kind of trying to stabilize the game.

Lex Fridman (23:53.760)

Well, maybe, I mean, it's sort of,

Demis Hassabis (23:55.080)

I don't know, it's chicken and egg situation,

Lex Fridman (23:56.560)

probably both came together.

Lex Fridman (23:57.680)

But the fact that it's got to this beautiful equilibrium

Lex Fridman (24:00.520)

where you can have the bishop and knight

Demis Hassabis (24:02.360)

that are so different in power,

Lex Fridman (24:04.400)

but so equal in value across the set

Lex Fridman (24:06.920)

of the universe of all positions, right?

Lex Fridman (24:09.480)

Somehow they've been balanced by humanity

Demis Hassabis (24:11.560)

over hundreds of years,

Lex Fridman (24:13.480)

I think gives the game the creative tension

Demis Hassabis (24:16.880)

that you can swap the bishop and knights

Lex Fridman (24:19.000)

for a bishop for a knight,

Lex Fridman (24:20.160)

and they're more or less worth the same,

Lex Fridman (24:22.080)

but now you aim for a different type of position.

Demis Hassabis (24:24.040)

If you have the knight, you want a closed position.

Lex Fridman (24:26.040)

If you have the bishop, you want an open position.

Demis Hassabis (24:28.160)

So, I think that creates

Lex Fridman (24:29.000)

a lot of the creative tension in chess.

Demis Hassabis (24:30.920)

So, some kind of controlled creative tension.

Lex Fridman (24:34.040)

From an AI perspective,

Lex Fridman (24:35.960)

do you think AI systems could eventually design games

Lex Fridman (24:38.840)

that are optimally compelling to humans?

Demis Hassabis (24:41.640)

Well, that's an interesting question.

Lex Fridman (24:42.920)

Sometimes I get asked about AI and creativity,

Lex Fridman (24:46.000)

and the way I answered that is relevant to that question,

Lex Fridman (24:48.880)

which is that I think there are different levels

Demis Hassabis (24:51.240)

of creativity, one could say.

Lex Fridman (24:52.920)

So, I think if we define creativity

Demis Hassabis (24:55.320)

as coming up with something original, right,

Lex Fridman (24:57.280)

that's useful for a purpose,

Demis Hassabis (24:59.320)

then I think the kind of lowest level of creativity

Lex Fridman (25:02.240)

is like an interpolation.

Demis Hassabis (25:03.720)

So, an averaging of all the examples you see.

Lex Fridman (25:06.280)

So, maybe a very basic AI system could say

Demis Hassabis (25:08.320)

you could have that.

Lex Fridman (25:09.160)

So, you show it millions of pictures of cats,

Lex Fridman (25:11.400)

and then you say, give me an average looking cat, right?

Lex Fridman (25:13.920)

Generate me an average looking cat.

Demis Hassabis (25:15.480)

I would call that interpolation.

Lex Fridman (25:17.200)

Then there's extrapolation,

Demis Hassabis (25:18.720)

which something like AlphaGo showed.

Lex Fridman (25:20.440)

So, AlphaGo played millions of games of Go against itself,

Lex Fridman (25:24.320)

and then it came up with brilliant new ideas

Lex Fridman (25:26.600)

like Move 37 in game two, brilliant motif strategies in Go

Demis Hassabis (25:30.760)

that no humans had ever thought of,

Lex Fridman (25:32.840)

even though we've played it for thousands of years

Lex Fridman (25:34.800)

and professionally for hundreds of years.

Lex Fridman (25:36.600)

So, that I call that extrapolation,

Lex Fridman (25:38.840)

but then there's still a level above that,

Lex Fridman (25:41.080)

which is, you could call out of the box thinking

Lex Fridman (25:44.000)

or true innovation, which is, could you invent Go, right?

Lex Fridman (25:47.520)

Could you invent chess and not just come up

Demis Hassabis (25:49.200)

with a brilliant chess move or brilliant Go move,

Lex Fridman (25:51.320)

but can you actually invent chess

Lex Fridman (25:53.680)

or something as good as chess or Go?

Lex Fridman (25:55.880)

And I think one day AI could, but then what's missing

Demis Hassabis (26:00.080)

is how would you even specify that task

Lex Fridman (26:02.240)

to a program right now?

Lex Fridman (26:04.440)

And the way I would do it if I was telling a human to do it

Lex Fridman (26:07.560)

or a human games designer to do it is I would say,

Demis Hassabis (26:10.800)

something like Go, I would say, come up with a game

Lex Fridman (26:14.120)

that only takes five minutes to learn,

Demis Hassabis (26:16.080)

which Go does because it's got simple rules,

Lex Fridman (26:17.880)

but many lifetimes to master, right?

Demis Hassabis (26:20.280)

Or impossible to master in one lifetime

Lex Fridman (26:22.080)

because it's so deep and so complex.

Lex Fridman (26:23.920)

And then it's aesthetically beautiful.

Lex Fridman (26:26.520)

And also it can be completed in three or four hours

Demis Hassabis (26:30.080)

of gameplay time, which is useful for us in a human day.

Lex Fridman (26:35.280)

And so you might specify these sort of high level concepts

Demis Hassabis (26:38.560)

like that, and then with that

Lex Fridman (26:40.640)

and then maybe a few other things,

Demis Hassabis (26:42.800)

one could imagine that Go satisfies those constraints.

Lex Fridman (26:47.560)

But the problem is that we're not able

Demis Hassabis (26:49.600)

to specify abstract notions like that,

Lex Fridman (26:53.040)

high level abstract notions like that yet to our AI systems.

Lex Fridman (26:57.040)

And I think there's still something missing there

Lex Fridman (26:58.840)

in terms of high level concepts or abstractions

Demis Hassabis (27:01.840)

that they truly understand

Lex Fridman (27:03.080)

and they're combinable and compositional.

Lex Fridman (27:06.560)

So for the moment, I think AI is capable

Lex Fridman (27:09.760)

of doing interpolation and extrapolation,

Lex Fridman (27:11.760)

but not true invention.

Lex Fridman (27:13.520)

So coming up with rule sets and optimizing

Demis Hassabis (27:18.040)

with complicated objectives around those rule sets,

Lex Fridman (27:20.640)

we can't currently do.

Lex Fridman (27:22.280)

But you could take a specific rule set

Lex Fridman (27:25.480)

and then run a kind of self play experiment

Demis Hassabis (27:28.360)

to see how long, just observe how an AI system

Lex Fridman (27:32.040)

from scratch learns, how long is that journey of learning?

Lex Fridman (27:35.960)

And maybe if it satisfies some of those other things

Lex Fridman (27:39.160)

you mentioned in terms of quickness to learn and so on,

Lex Fridman (27:41.680)

and you could see a long journey to master

Lex Fridman (27:44.200)

for even an AI system, then you could say

Demis Hassabis (27:46.920)

that this is a promising game.

Lex Fridman (27:49.280)

But it would be nice to do almost like AlphaCode

Lex Fridman (27:51.720)

so programming rules.

Lex Fridman (27:53.960)

So generating rules that automate even that part

Demis Hassabis (27:59.000)

of the generation of rules.

Lex Fridman (28:00.440)

So I have thought about systems actually

Demis Hassabis (28:02.960)

that I think would be amazing for a games designer.

Lex Fridman (28:05.680)

If you could have a system that takes your game,

Demis Hassabis (28:09.200)

plays it tens of millions of times, maybe overnight,

Lex Fridman (28:11.960)

and then self balances the rules better.

Lex Fridman (28:13.840)

So it tweaks the rules and maybe the equations

Lex Fridman (28:18.080)

and the parameters so that the game is more balanced,

Demis Hassabis (28:22.680)

the units in the game or some of the rules could be tweaked.

Lex Fridman (28:26.280)

So it's a bit of like giving a base set

Lex Fridman (28:28.320)

and then allowing Monte Carlo Tree Search

Lex Fridman (28:30.800)

or something like that to sort of explore it.

Lex Fridman (28:33.360)

And I think that would be super powerful tool actually

Lex Fridman (28:37.080)

for balancing, auto balancing a game,

Demis Hassabis (28:39.720)

which usually takes thousands of hours

Lex Fridman (28:42.120)

from hundreds of human games testers normally

Demis Hassabis (28:44.520)

to balance a game like StarCraft,

Lex Fridman (28:47.480)

which is Blizzard are amazing at balancing their games,

Lex Fridman (28:50.640)

but it takes them years and years and years.

Lex Fridman (28:52.600)

So one could imagine at some point

Demis Hassabis (28:54.120)

when this stuff becomes efficient enough

Lex Fridman (28:57.080)

to you might be able to do that like overnight.

Lex Fridman (28:59.560)

Do you think a game that is optimal designed by an AI system

Lex Fridman (29:05.000)

would look very much like a planet earth?

Demis Hassabis (29:09.640)

Maybe, maybe it's only the sort of game

Lex Fridman (29:11.640)

I would love to make is, and I've tried in my games career,

Demis Hassabis (29:16.040)

the games design career, my first big game

Lex Fridman (29:18.560)

was designing a theme park, an amusement park.

Demis Hassabis (29:21.440)

Then with games like Republic, I tried to have games

Lex Fridman (29:25.200)

where we designed whole cities and allowed you to play in.

Demis Hassabis (29:28.480)

So, and of course people like Will Wright

Lex Fridman (29:30.320)

have written games like SimEarth,

Demis Hassabis (29:32.640)

trying to simulate the whole of earth, pretty tricky,

Lex Fridman (29:35.200)

but I think.

Demis Hassabis (29:36.040)

SimEarth, I haven't actually played that one.

Lex Fridman (29:37.600)

So what is it?

Lex Fridman (29:38.440)

Does it incorporate of evolution or?

Lex Fridman (29:40.320)

Yeah, it has evolution and it sort of tries to,

Demis Hassabis (29:43.280)

it sort of treats it as an entire biosphere,

Lex Fridman (29:45.320)

but from quite high level.

Demis Hassabis (29:47.240)

So.

Lex Fridman (29:48.080)

It'd be nice to be able to sort of zoom in,

Demis Hassabis (29:50.280)

zoom out and zoom in.

Lex Fridman (29:51.320)

Exactly, exactly.

Lex Fridman (29:52.160)

So obviously it couldn't do, that was in the 90s.

Lex Fridman (29:53.440)

I think he wrote that in the 90s.

Lex Fridman (29:54.920)

So it couldn't, it wasn't able to do that,

Lex Fridman (29:57.560)

but that would be obviously the ultimate sandbox game.

Demis Hassabis (2:00:00.880)

from say 10, 20 years ago.

Lex Fridman (2:00:02.680)

Back 10, 20 years ago, it would have been

Demis Hassabis (2:00:05.480)

a whole day of research, individual research or programming,

Lex Fridman (2:00:10.900)

doing some experiment, neuroscience,

Demis Hassabis (2:00:12.580)

computer science experiment,

Lex Fridman (2:00:14.080)

reading lots of research papers.

Lex Fridman (2:00:16.640)

And then perhaps at nighttime,

Lex Fridman (2:00:19.720)

reading science fiction books or playing some games.

Lex Fridman (2:00:25.440)

But lots of focus, so like deep focused work

Lex Fridman (2:00:28.360)

on whether it's programming or reading research papers.

Demis Hassabis (2:00:32.440)

Yes, so that would be lots of deep focus work.

Lex Fridman (2:00:35.300)

These days for the last sort of, I guess, five to 10 years,

Demis Hassabis (2:00:39.560)

I've actually got quite a structure

Lex Fridman (2:00:41.020)

that works very well for me now,

Demis Hassabis (2:00:42.360)

which is that I'm a complete night owl, always have been.

Lex Fridman (2:00:46.140)

So I optimize for that.

Lex Fridman (2:00:47.680)

So I'll basically do a normal day's work,

Lex Fridman (2:00:50.760)

get into work about 11 o clock

Lex Fridman (2:00:52.560)

and sort of do work to about seven in the office.

Lex Fridman (2:00:56.400)

And I will arrange back to back meetings

Demis Hassabis (2:00:58.960)

for the entire time of that.

Lex Fridman (2:01:00.920)

And with as many, meet as many people as possible.

Lex Fridman (2:01:03.200)

So that's my collaboration management part of the day.

Lex Fridman (2:01:06.500)

Then I go home, spend time with the family and friends,

Demis Hassabis (2:01:10.680)

have dinner, relax a little bit.

Lex Fridman (2:01:13.600)

And then I start a second day of work.

Demis Hassabis (2:01:15.240)

I call it my second day of work around 10 p.m., 11 p.m.

Lex Fridman (2:01:18.500)

And that's the time to about the small hours of the morning,

Demis Hassabis (2:01:21.260)

four or five in the morning, where I will do my thinking

Lex Fridman (2:01:24.760)

and reading and research, writing research papers.

Demis Hassabis (2:01:29.000)

Sadly, I don't have time to code anymore,

Lex Fridman (2:01:30.960)

but it's not efficient to do that these days,

Demis Hassabis (2:01:34.900)

given the amount of time I have.

Lex Fridman (2:01:37.120)

But that's when I do, you know,

Demis Hassabis (2:01:38.360)

maybe do the long kind of stretches

Lex Fridman (2:01:40.760)

of thinking and planning.

Lex Fridman (2:01:42.440)

And then probably, you know, using email, other things,

Lex Fridman (2:01:45.280)

I would set, I would fire off a lot of things to my team

Demis Hassabis (2:01:47.880)

to deal with the next morning.

Lex Fridman (2:01:49.360)

But actually thinking about this overnight,

Demis Hassabis (2:01:51.640)

we should go for this project

Lex Fridman (2:01:53.200)

or arrange this meeting the next day.

Demis Hassabis (2:01:54.880)

When you're thinking through a problem,

Lex Fridman (2:01:56.120)

are you talking about a sheet of paper with a pen?

Lex Fridman (2:01:58.160)

Is there some structured process?

Lex Fridman (2:02:01.040)

I still like pencil and paper best for working out things,

Lex Fridman (2:02:04.360)

but these days it's just so efficient

Lex Fridman (2:02:06.720)

to read research papers just on the screen.

Demis Hassabis (2:02:08.720)

I still often print them out, actually.

Lex Fridman (2:02:10.220)

I still prefer to mark out things.

Lex Fridman (2:02:12.540)

And I find it goes into the brain better

Lex Fridman (2:02:14.880)

and sticks in the brain better

Demis Hassabis (2:02:16.000)

when you're still using physical pen and pencil and paper.

Lex Fridman (2:02:19.440)

So you take notes with the...

Demis Hassabis (2:02:20.800)

I have lots of notes, electronic ones,

Lex Fridman (2:02:22.440)

and also whole stacks of notebooks that I use at home, yeah.

Demis Hassabis (2:02:27.640)

On some of these most challenging next steps, for example,

Lex Fridman (2:02:30.320)

stuff none of us know about that you're working on,

Demis Hassabis (2:02:33.800)

you're thinking,

Lex Fridman (2:02:35.580)

there's some deep thinking required there, right?

Lex Fridman (2:02:37.640)

Like what is the right problem?

Lex Fridman (2:02:39.420)

What is the right approach?

Demis Hassabis (2:02:41.280)

Because you're gonna have to invest a huge amount of time

Lex Fridman (2:02:43.920)

for the whole team.

Demis Hassabis (2:02:44.800)

They're going to have to pursue this thing.

Lex Fridman (2:02:46.760)

What's the right way to do it?

Lex Fridman (2:02:48.560)

Is RL gonna work here or not?

Lex Fridman (2:02:50.040)

Yes.

Lex Fridman (2:02:50.880)

What's the right thing to try?

Lex Fridman (2:02:53.120)

What's the right benchmark to use?

Lex Fridman (2:02:55.120)

Do we need to construct a benchmark from scratch?

Lex Fridman (2:02:57.320)

All those kinds of things.

Demis Hassabis (2:02:58.200)

Yes.

Lex Fridman (2:02:59.040)

So I think of all those kinds of things

Demis Hassabis (2:03:00.200)

in the nighttime phase, but also much more,

Lex Fridman (2:03:03.480)

I find I've always found the quiet hours of the morning

Demis Hassabis (2:03:07.660)

when everyone's asleep, it's super quiet outside.

Lex Fridman (2:03:11.420)

I love that time.

Demis Hassabis (2:03:12.280)

It's the golden hours,

Lex Fridman (2:03:13.360)

like between one and three in the morning.

Demis Hassabis (2:03:16.480)

Put some music on, some inspiring music on,

Lex Fridman (2:03:18.880)

and then think these deep thoughts.

Lex Fridman (2:03:21.600)

So that's when I would read my philosophy books

Lex Fridman (2:03:24.240)

and Spinoza's, my recent favorite can, all these things.

Lex Fridman (2:03:28.820)

And I read about a great scientist of history,

Lex Fridman (2:03:33.660)

how they did things, how they thought things.

Lex Fridman (2:03:35.640)

So that's when you do all your creative,

Lex Fridman (2:03:37.240)

that's when I do all my creative thinking.

Lex Fridman (2:03:39.160)

And it's good, I think people recommend

Lex Fridman (2:03:41.840)

you do your sort of creative thinking in one block.

Lex Fridman (2:03:45.120)

And the way I organize the day,

Lex Fridman (2:03:47.160)

that way I don't get interrupted.

Demis Hassabis (2:03:48.560)

There's obviously no one else is up at those times.

Lex Fridman (2:03:51.460)

So I can go, I can sort of get super deep

Lex Fridman (2:03:55.880)

and super into flow.

Lex Fridman (2:03:57.560)

The other nice thing about doing it nighttime wise

Demis Hassabis (2:03:59.640)

is if I'm really onto something

Lex Fridman (2:04:02.760)

or I've got really deep into something,

Demis Hassabis (2:04:04.940)

I can choose to extend it

Lex Fridman (2:04:06.840)

and I'll go into six in the morning, whatever.

Lex Fridman (2:04:09.000)

And then I'll just pay for it the next day.

Lex Fridman (2:04:10.760)

So I'll be a bit tired and I won't be my best,

Lex Fridman (2:04:12.960)

but that's fine.

Lex Fridman (2:04:13.900)

I can decide looking at my schedule the next day

Lex Fridman (2:04:16.840)

and given where I'm at with this particular thought

Lex Fridman (2:04:19.360)

or creative idea that I'm gonna pay that cost the next day.

Lex Fridman (2:04:22.840)

So I think that's more flexible than morning people

Lex Fridman (2:04:26.220)

who do that, they get up at four in the morning.

Demis Hassabis (2:04:28.780)

They can also do those golden hours then,

Lex Fridman (2:04:31.000)

but then their start of their scheduled day

Demis Hassabis (2:04:32.640)

starts at breakfast, 8 a.m.,

Lex Fridman (2:04:34.440)

whatever they have their first meeting.

Lex Fridman (2:04:36.040)

And then it's hard, you have to reschedule a day

Lex Fridman (2:04:37.880)

if you're in flow.

Lex Fridman (2:04:38.960)

So I don't have to do that.

Lex Fridman (2:04:39.800)

So that could be a true special thread of thoughts

Demis Hassabis (2:04:41.880)

that you're too passionate about.

Lex Fridman (2:04:45.160)

This is where some of the greatest ideas

Demis Hassabis (2:04:46.740)

could potentially come is when you just lose yourself

Lex Fridman (2:04:49.320)

late into the night.

Lex Fridman (2:04:51.360)

And for the meetings, I mean, you're loading in

Lex Fridman (2:04:53.860)

really hard problems in a very short amount of time.

Lex Fridman (2:04:56.520)

So you have to do some kind of first principles thinking

Lex Fridman (2:04:58.800)

here, it's like, what's the problem?

Lex Fridman (2:05:00.160)

What's the state of things?

Lex Fridman (2:05:01.360)

What's the right next steps?

Demis Hassabis (2:05:03.120)

You have to get really good at context switching,

Lex Fridman (2:05:05.120)

which is one of the hardest things,

Demis Hassabis (2:05:07.200)

because especially as we do so many things,

Lex Fridman (2:05:09.020)

if you include all the scientific things we do,

Demis Hassabis (2:05:10.800)

scientific fields we're working in,

Lex Fridman (2:05:12.600)

these are complex fields in themselves.

Lex Fridman (2:05:15.380)

And you have to sort of keep abreast of that.

Lex Fridman (2:05:18.960)

But I enjoy it.

Demis Hassabis (2:05:20.000)

I've always been a sort of generalist in a way.

Lex Fridman (2:05:23.840)

And that's actually what happened in my games career

Demis Hassabis (2:05:25.600)

after chess.

Lex Fridman (2:05:27.880)

One of the reasons I stopped playing chess

Demis Hassabis (2:05:29.260)

was because I got into computers,

Lex Fridman (2:05:30.320)

but also I started realizing there were many other

Demis Hassabis (2:05:32.280)

great games out there to play too.

Lex Fridman (2:05:33.880)

So I've always been that way inclined, multidisciplinary.

Lex Fridman (2:05:36.920)

And there's too many interesting things in the world

Lex Fridman (2:05:39.120)

to spend all your time just on one thing.

Lex Fridman (2:05:41.680)

So you mentioned Spinoza, gotta ask the big, ridiculously

Lex Fridman (2:05:45.640)

big question about life.

Lex Fridman (2:05:47.640)

What do you think is the meaning of this whole thing?

Lex Fridman (2:05:50.480)

Why are we humans here?

Demis Hassabis (2:05:52.560)

You've already mentioned that perhaps the universe

Lex Fridman (2:05:55.120)

created us.

Lex Fridman (2:05:56.720)

Is that why you think we're here?

Lex Fridman (2:05:58.920)

To understand how the universe works?

Demis Hassabis (2:06:00.120)

Yeah, I think my answer to that would be,

Lex Fridman (2:06:02.080)

and at least the life I'm living,

Demis Hassabis (2:06:03.960)

is to gain and understand the knowledge,

Lex Fridman (2:06:08.120)

to gain knowledge and understand the universe.

Demis Hassabis (2:06:10.600)

That's what I think, I can't see any higher purpose

Lex Fridman (2:06:13.560)

than that if you think back to the classical Greeks,

Demis Hassabis (2:06:15.720)

the virtue of gaining knowledge.

Lex Fridman (2:06:17.560)

It's, I think it's one of the few true virtues

Demis Hassabis (2:06:20.440)

is to understand the world around us

Lex Fridman (2:06:23.600)

and the context and humanity better.

Lex Fridman (2:06:25.680)

And I think if you do that, you become more compassionate

Lex Fridman (2:06:29.140)

and more understanding yourself and more tolerant

Lex Fridman (2:06:32.080)

and all these, I think all these other things

Lex Fridman (2:06:33.580)

may flow from that.

Lex Fridman (2:06:34.760)

And to me, understanding the nature of reality,

Lex Fridman (2:06:37.640)

that is the biggest question.

Lex Fridman (2:06:38.760)

What is going on here is sometimes the colloquial way I say.

Lex Fridman (2:06:41.400)

What is really going on here?

Demis Hassabis (2:06:43.600)

It's so mysterious.

Lex Fridman (2:06:44.900)

I feel like we're in some huge puzzle.

Lex Fridman (2:06:47.040)

And it's, but the world is also seems to be,

Lex Fridman (2:06:49.960)

the universe seems to be structured in a way.

Demis Hassabis (2:06:52.880)

You know, why is it structured in a way

Lex Fridman (2:06:54.280)

that science is even possible?

Demis Hassabis (2:06:55.840)

That, you know, methods, the scientific method works,

Lex Fridman (2:06:58.160)

things are repeatable.

Demis Hassabis (2:07:00.240)

It feels like it's almost structured in a way

Lex Fridman (2:07:02.560)

to be conducive to gaining knowledge.

Lex Fridman (2:07:05.000)

So I feel like, and you know,

Lex Fridman (2:07:06.480)

why should computers be even possible?

Demis Hassabis (2:07:07.960)

Wasn't that amazing that computational electronic devices

Lex Fridman (2:07:11.880)

can be possible, and they're made of sand,

Demis Hassabis (2:07:15.300)

our most common element that we have,

Lex Fridman (2:07:17.280)

you know, silicon on the Earth's crust.

Demis Hassabis (2:07:19.960)

It could have been made of diamond or something,

Lex Fridman (2:07:21.480)

then we would have only had one computer.

Lex Fridman (2:07:23.800)

So a lot of things are kind of slightly suspicious to me.

Lex Fridman (2:07:26.560)

It sure as heck sounds, this puzzle sure as heck sounds

Demis Hassabis (2:07:29.220)

like something we talked about earlier,

Lex Fridman (2:07:30.760)

what it takes to design a game that's really fun to play

Demis Hassabis (2:07:35.120)

for prolonged periods of time.

Lex Fridman (2:07:36.620)

And it does seem like this puzzle, like you mentioned,

Demis Hassabis (2:07:40.420)

the more you learn about it,

Lex Fridman (2:07:42.280)

the more you realize how little you know.

Lex Fridman (2:07:44.860)

So it humbles you, but excites you

Lex Fridman (2:07:46.820)

by the possibility of learning more.

Demis Hassabis (2:07:49.020)

It's one heck of a puzzle we got going on here.

Lex Fridman (2:07:53.560)

So like I mentioned, of all the people in the world,

Demis Hassabis (2:07:56.420)

you're very likely to be the one who creates the AGI system

Lex Fridman (2:08:02.580)

that achieves human level intelligence and goes beyond it.

Lex Fridman (2:08:06.320)

So if you got a chance and very well,

Lex Fridman (2:08:08.360)

you could be the person that goes into the room

Demis Hassabis (2:08:10.340)

with the system and have a conversation.

Lex Fridman (2:08:13.140)

Maybe you only get to ask one question.

Lex Fridman (2:08:15.260)

If you do, what question would you ask her?

Lex Fridman (2:08:19.460)

I would probably ask, what is the true nature of reality?

Demis Hassabis (2:08:23.660)

I think that's the question.

Lex Fridman (2:08:24.560)

I don't know if I'd understand the answer

Demis Hassabis (2:08:25.980)

because maybe it would be 42 or something like that,

Lex Fridman (2:08:28.540)

but that's the question I would ask.

Lex Fridman (2:08:32.420)

And then there'll be a deep sigh from the systems,

Lex Fridman (2:08:34.820)

like, all right, how do I explain to this human?

Demis Hassabis (2:08:37.460)

All right, let me, I don't have time to explain.

Lex Fridman (2:08:41.860)

Maybe I'll draw you a picture that it is.

Lex Fridman (2:08:44.660)

I mean, how do you even begin to answer that question?

Lex Fridman (2:08:51.280)

Well, I think it would.

Lex Fridman (2:08:52.780)

What would you think the answer could possibly look like?

Lex Fridman (2:08:55.680)

I think it could start looking like

Demis Hassabis (2:08:59.940)

more fundamental explanations of physics

Lex Fridman (2:09:02.060)

would be the beginning.

Demis Hassabis (2:09:03.900)

More careful specification of that,

Lex Fridman (2:09:05.740)

taking you, walking us through by the hand

Demis Hassabis (2:09:07.700)

as to what one would do to maybe prove those things out.

Lex Fridman (2:09:10.620)

Maybe giving you glimpses of what things

Demis Hassabis (2:09:13.700)

you totally miss in the physics of today.

Lex Fridman (2:09:15.740)

Exactly, exactly.

Demis Hassabis (2:09:16.700)

Just here's glimpses of, no, like there's a much,

Lex Fridman (2:09:22.260)

a much more elaborate world

Demis Hassabis (2:09:23.640)

or a much simpler world or something.

Lex Fridman (2:09:26.780)

A much deeper, maybe simpler explanation of things,

Demis Hassabis (2:09:30.260)

right, than the standard model of physics,

Lex Fridman (2:09:31.900)

which we know doesn't work, but we still keep adding to.

Demis Hassabis (2:09:34.860)

So, and that's how I think the beginning

Lex Fridman (2:09:37.940)

of an explanation would look.

Lex Fridman (2:09:38.940)

And it would start encompassing many of the mysteries

Lex Fridman (2:09:41.260)

that we have wondered about for thousands of years,

Demis Hassabis (2:09:43.380)

like consciousness, dreaming, life, and gravity,

Lex Fridman (2:09:47.900)

all of these things.

Demis Hassabis (2:09:48.820)

Yeah, giving us glimpses of explanations for those things.

Lex Fridman (2:09:52.620)

Well, Damasir, one of the special human beings

Demis Hassabis (2:09:57.180)

in this giant puzzle of ours,

Lex Fridman (2:09:59.080)

and it's a huge honor that you would take a pause

Demis Hassabis (2:10:01.020)

from the bigger puzzle to solve this small puzzle

Lex Fridman (2:10:03.260)

of a conversation with me today.

Demis Hassabis (2:10:04.780)

It's truly an honor and a pleasure.

Lex Fridman (2:10:06.300)

Thank you so much.

Demis Hassabis (2:10:07.140)

Thank you, I really enjoyed it.

Lex Fridman (2:10:07.960)

Thanks, Lex.

Demis Hassabis (2:10:09.100)

Thanks for listening to this conversation

Lex Fridman (2:10:10.580)

with Damas Ashabis.

Demis Hassabis (2:10:11.980)

To support this podcast,

Lex Fridman (2:10:13.180)

please check out our sponsors in the description.

Lex Fridman (2:10:15.820)

And now, let me leave you with some words

Lex Fridman (2:10:17.900)

from Edgar Dykstra.

Demis Hassabis (2:10:20.380)

Computer science is no more about computers

Lex Fridman (2:10:23.500)

than astronomy is about telescopes.

Demis Hassabis (2:10:26.260)

Thank you for listening, and hope to see you next time.

Lex Fridman (30:00.520)

Of course.

Lex Fridman (30:01.480)

On that topic, do you think we're living in a simulation?

Lex Fridman (30:04.760)

Yes, well, so, okay.

Lex Fridman (30:06.160)

So I.

Lex Fridman (30:07.000)

We're gonna jump around from the absurdly philosophical

Demis Hassabis (30:09.280)

to the technical.

Lex Fridman (30:10.120)

Sure, sure, very, very happy to.

Lex Fridman (30:11.880)

So I think my answer to that question

Lex Fridman (30:13.800)

is a little bit complex because there is simulation theory,

Demis Hassabis (30:17.640)

which obviously Nick Bostrom,

Lex Fridman (30:18.800)

I think famously first proposed.

Lex Fridman (30:21.680)

And I don't quite believe it in that sense.

Lex Fridman (30:24.720)

So in the sense that are we in some sort of computer game

Demis Hassabis (30:29.600)

or have our descendants somehow recreated earth

Lex Fridman (30:34.000)

in the 21st century and some,

Demis Hassabis (30:36.520)

for some kind of experimental reason.

Lex Fridman (30:38.480)

I think that, but I do think that we,

Demis Hassabis (30:41.880)

that we might be, that the best way to understand physics

Lex Fridman (30:45.600)

and the universe is from a computational perspective.

Lex Fridman (30:49.320)

So understanding it as an information universe

Lex Fridman (30:52.440)

and actually information being the most fundamental unit

Demis Hassabis (30:56.200)

of reality rather than matter or energy.

Lex Fridman (30:59.920)

So a physicist would say, you know, matter or energy,

Demis Hassabis (31:02.400)

you know, E equals MC squared.

Lex Fridman (31:03.760)

These are the things that are the fundamentals

Demis Hassabis (31:06.440)

of the universe.

Lex Fridman (31:07.400)

I'd actually say information,

Demis Hassabis (31:09.880)

which of course itself can be,

Lex Fridman (31:11.760)

can specify energy or matter, right?

Demis Hassabis (31:13.560)

Matter is actually just, you know,

Lex Fridman (31:14.920)

we're just out the way our bodies

Lex Fridman (31:16.880)

and the molecules in our body are arranged as information.

Lex Fridman (31:19.720)

So I think information may be the most fundamental way

Demis Hassabis (31:23.080)

to describe the universe.

Lex Fridman (31:24.960)

And therefore you could say we're in some sort of simulation

Demis Hassabis (31:28.280)

because of that.

Lex Fridman (31:29.880)

But I don't, I do, I'm not,

Demis Hassabis (31:31.040)

I'm not really a subscriber to the idea that, you know,

Lex Fridman (31:34.200)

these are sort of throw away billions of simulations around.

Demis Hassabis (31:36.920)

I think this is actually very critical and possibly unique,

Lex Fridman (31:40.640)

this simulation.

Demis Hassabis (31:41.760)

This particular one.

Lex Fridman (31:42.600)

Yes.

Lex Fridman (31:43.440)

And you just mean treating the universe as a computer

Lex Fridman (31:48.760)

that's processing and modifying information

Demis Hassabis (31:52.240)

is a good way to solve the problems of physics,

Lex Fridman (31:54.880)

of chemistry, of biology,

Lex Fridman (31:57.160)

and perhaps of humanity and so on.

Lex Fridman (31:59.720)

Yes, I think understanding physics

Demis Hassabis (32:02.240)

in terms of information theory

Lex Fridman (32:04.840)

might be the best way to really understand

Demis Hassabis (32:07.880)

what's going on here.

Lex Fridman (32:09.360)

From our understanding of a universal Turing machine,

Demis Hassabis (32:13.560)

from our understanding of a computer,

Lex Fridman (32:15.280)

do you think there's something outside

Demis Hassabis (32:17.400)

of the capabilities of a computer

Lex Fridman (32:19.440)

that is present in our universe?

Demis Hassabis (32:21.000)

You have a disagreement with Roger Penrose

Lex Fridman (32:23.560)

about the nature of consciousness.

Demis Hassabis (32:25.920)

He thinks that consciousness is more

Lex Fridman (32:27.760)

than just a computation.

Lex Fridman (32:30.080)

Do you think all of it, the whole shebangs,

Lex Fridman (32:32.680)

can be a computation?

Demis Hassabis (32:34.000)

Yeah, I've had many fascinating debates

Lex Fridman (32:35.840)

with Sir Roger Penrose,

Lex Fridman (32:37.680)

and obviously he's famously,

Lex Fridman (32:39.680)

and I read, you know, Emperors of the New Mind

Lex Fridman (32:41.520)

and his books, his classical books,

Lex Fridman (32:45.400)

and they were pretty influential in the 90s.

Lex Fridman (32:47.800)

And he believes that there's something more,

Lex Fridman (32:50.960)

something quantum that is needed

Demis Hassabis (32:53.040)

to explain consciousness in the brain.

Lex Fridman (32:55.840)

I think about what we're doing actually at DeepMind

Lex Fridman (32:58.320)

and what my career is being,

Lex Fridman (32:59.920)

we're almost like Turing's champion.

Lex Fridman (33:01.920)

So we are pushing Turing machines or classical computation

Lex Fridman (33:05.360)

to the limits.

Lex Fridman (33:06.200)

What are the limits of what classical computing can do?

Lex Fridman (33:09.440)

Now, and at the same time,

Demis Hassabis (33:11.760)

I've also studied neuroscience to see,

Lex Fridman (33:14.240)

and that's why I did my PhD in,

Demis Hassabis (33:15.520)

was to see, also to look at, you know,

Lex Fridman (33:17.720)

is there anything quantum in the brain

Lex Fridman (33:19.240)

from a neuroscience or biological perspective?

Lex Fridman (33:21.360)

And so far, I think most neuroscientists

Lex Fridman (33:24.560)

and most mainstream biologists and neuroscientists

Lex Fridman (33:26.440)

would say there's no evidence of any quantum systems

Demis Hassabis (33:29.480)

or effects in the brain.

Lex Fridman (33:30.800)

As far as we can see, it can be mostly explained

Demis Hassabis (33:33.000)

by classical theories.

Lex Fridman (33:35.880)

So, and then, so there's sort of the search

Demis Hassabis (33:39.280)

from the biology side.

Lex Fridman (33:40.600)

And then at the same time,

Demis Hassabis (33:42.120)

there's the raising of the water, the bar,

Lex Fridman (33:44.960)

from what classical Turing machines can do.

Demis Hassabis (33:48.240)

And, you know, including our new AI systems.

Lex Fridman (33:51.680)

And as you alluded to earlier, you know,

Demis Hassabis (33:55.040)

I think AI, especially in the last decade plus,

Lex Fridman (33:57.760)

has been a continual story now of surprising events

Lex Fridman (34:02.360)

and surprising successes,

Lex Fridman (34:03.920)

knocking over one theory after another

Demis Hassabis (34:05.800)

of what was thought to be impossible, you know,

Lex Fridman (34:07.760)

from Go to protein folding and so on.

Lex Fridman (34:10.080)

And so I think I would be very hesitant

Lex Fridman (34:14.760)

to bet against how far the universal Turing machine

Lex Fridman (34:19.520)

and classical computation paradigm can go.

Lex Fridman (34:23.400)

And my betting would be that all of,

Demis Hassabis (34:26.720)

certainly what's going on in our brain,

Lex Fridman (34:29.080)

can probably be mimicked or approximated

Demis Hassabis (34:32.160)

on a classical machine,

Lex Fridman (34:34.720)

not requiring something metaphysical or quantum.

Lex Fridman (34:38.400)

And we'll get there with some of the work with AlphaFold,

Lex Fridman (34:41.720)

which I think begins the journey of modeling

Demis Hassabis (34:45.080)

this beautiful and complex world of biology.

Lex Fridman (34:48.160)

So you think all the magic of the human mind

Demis Hassabis (34:50.160)

comes from this, just a few pounds of mush,

Lex Fridman (34:54.280)

of biological computational mush,

Demis Hassabis (34:57.480)

that's akin to some of the neural networks,

Lex Fridman (35:01.560)

not directly, but in spirit

Demis Hassabis (35:03.800)

that DeepMind has been working with.

Lex Fridman (35:06.200)

Well, look, I think it's, you say it's a few, you know,

Demis Hassabis (35:08.680)

of course it's, this is the,

Lex Fridman (35:09.680)

I think the biggest miracle of the universe

Demis Hassabis (35:11.520)

is that it is just a few pounds of mush in our skulls.

Lex Fridman (35:15.000)

And yet it's also our brains are the most complex objects

Demis Hassabis (35:18.560)

that we know of in the universe.

Lex Fridman (35:20.240)

So there's something profoundly beautiful

Lex Fridman (35:22.360)

and amazing about our brains.

Lex Fridman (35:23.920)

And I think that it's an incredibly,

Demis Hassabis (35:28.640)

incredible efficient machine.

Lex Fridman (35:30.720)

And it's, you know, phenomenon basically.

Lex Fridman (35:35.520)

And I think that building AI,

Lex Fridman (35:37.480)

one of the reasons I wanna build AI,

Lex Fridman (35:38.920)

and I've always wanted to is,

Lex Fridman (35:40.440)

I think by building an intelligent artifact like AI,

Lex Fridman (35:43.800)

and then comparing it to the human mind,

Lex Fridman (35:46.480)

that will help us unlock the uniqueness

Lex Fridman (35:49.560)

and the true secrets of the mind

Lex Fridman (35:50.960)

that we've always wondered about since the dawn of history,

Demis Hassabis (35:53.480)

like consciousness, dreaming, creativity, emotions,

Lex Fridman (35:59.160)

what are all these things, right?

Demis Hassabis (36:00.760)

We've wondered about them since the dawn of humanity.

Lex Fridman (36:04.200)

And I think one of the reasons,

Demis Hassabis (36:05.920)

and, you know, I love philosophy and philosophy of mind is,

Lex Fridman (36:08.760)

we found it difficult is there haven't been the tools

Demis Hassabis (36:11.200)

for us to really, other than introspection,

Lex Fridman (36:13.680)

from very clever people in history,

Demis Hassabis (36:15.880)

very clever philosophers,

Lex Fridman (36:17.200)

to really investigate this scientifically.

Lex Fridman (36:19.360)

But now suddenly we have a plethora of tools.

Lex Fridman (36:21.720)

Firstly, we have all of the neuroscience tools,

Demis Hassabis (36:23.240)

fMRI machines, single cell recording, all of this stuff,

Lex Fridman (36:25.920)

but we also have the ability, computers and AI,

Demis Hassabis (36:29.000)

to build intelligent systems.

Lex Fridman (36:31.640)

So I think that, you know,

Demis Hassabis (36:34.720)

I think it is amazing what the human mind does.

Lex Fridman (36:37.320)

And I'm kind of in awe of it really.

Lex Fridman (36:41.120)

And I think it's amazing that with our human minds,

Lex Fridman (36:44.440)

we're able to build things like computers

Lex Fridman (36:46.760)

and actually even, you know,

Lex Fridman (36:48.280)

think and investigate about these questions.

Demis Hassabis (36:49.880)

I think that's also a testament to the human mind.

Lex Fridman (36:52.720)

Yeah.

Demis Hassabis (36:53.560)

The universe built the human mind

Lex Fridman (36:56.200)

that now is building computers that help us understand

Demis Hassabis (36:59.600)

both the universe and our own human mind.

Lex Fridman (37:01.480)

That's right.

Demis Hassabis (37:02.320)

This is actually it.

Lex Fridman (37:03.140)

I mean, I think that's one, you know,

Demis Hassabis (37:03.980)

one could say we are,

Lex Fridman (37:05.760)

maybe we're the mechanism by which the universe

Demis Hassabis (37:08.160)

is going to try and understand itself.

Lex Fridman (37:09.840)

Yeah.

Demis Hassabis (37:10.680)

It's beautiful.

Lex Fridman (37:13.160)

So let's go to the basic building blocks of biology

Demis Hassabis (37:16.960)

that I think is another angle at which you can start

Lex Fridman (37:20.200)

to understand the human mind, the human body,

Demis Hassabis (37:22.280)

which is quite fascinating,

Lex Fridman (37:23.400)

which is from the basic building blocks,

Demis Hassabis (37:26.640)

start to simulate, start to model

Lex Fridman (37:28.960)

how from those building blocks,

Demis Hassabis (37:30.480)

you can construct bigger and bigger, more complex systems,

Lex Fridman (37:33.080)

maybe one day the entirety of the human biology.

Lex Fridman (37:35.820)

So here's another problem that thought

Lex Fridman (37:39.680)

to be impossible to solve, which is protein folding.

Lex Fridman (37:42.720)

And Alpha Fold or specifically Alpha Fold 2 did just that.

Lex Fridman (37:48.840)

It solved protein folding.

Demis Hassabis (37:50.320)

I think it's one of the biggest breakthroughs,

Lex Fridman (37:53.400)

certainly in the history of structural biology,

Lex Fridman (37:55.140)

but in general in science,

Lex Fridman (38:00.240)

maybe from a high level, what is it and how does it work?

Lex Fridman (38:04.840)

And then we can ask some fascinating questions after.

Lex Fridman (38:08.700)

Sure.

Lex Fridman (38:09.980)

So maybe to explain it to people not familiar

Lex Fridman (38:12.880)

with protein folding is, you know,

Demis Hassabis (38:14.400)

first of all, explain proteins, which is, you know,

Lex Fridman (38:16.980)

proteins are essential to all life.

Demis Hassabis (38:18.840)

Every function in your body depends on proteins.

Lex Fridman (38:21.520)

Sometimes they're called the workhorses of biology.

Lex Fridman (38:23.920)

And if you look into them and I've, you know,

Lex Fridman (38:25.340)

obviously as part of Alpha Fold,

Demis Hassabis (38:26.660)

I've been researching proteins and structural biology

Lex Fridman (38:30.200)

for the last few years, you know,

Demis Hassabis (38:31.760)

they're amazing little bio nano machines proteins.

Lex Fridman (38:34.760)

They're incredible if you actually watch little videos

Demis Hassabis (38:36.460)

of how they work, animations of how they work.

Lex Fridman (38:39.000)

And proteins are specified by their genetic sequence

Demis Hassabis (38:42.600)

called the amino acid sequence.

Lex Fridman (38:44.280)

So you can think of it as their genetic makeup.

Lex Fridman (38:47.040)

And then in the body in nature,

Lex Fridman (38:50.080)

they fold up into a 3D structure.

Lex Fridman (38:53.360)

So you can think of it as a string of beads

Lex Fridman (38:55.320)

and then they fold up into a ball.

Demis Hassabis (38:57.160)

Now, the key thing is you want to know

Lex Fridman (38:59.100)

what that 3D structure is because the structure,

Demis Hassabis (39:02.480)

the 3D structure of a protein is what helps to determine

Lex Fridman (39:06.120)

what does it do, the function it does in your body.

Lex Fridman (39:08.580)

And also if you're interested in drugs or disease,

Lex Fridman (39:12.320)

you need to understand that 3D structure

Demis Hassabis (39:13.980)

because if you want to target something

Lex Fridman (39:15.840)

with a drug compound about to block something

Demis Hassabis (39:18.640)

the protein's doing, you need to understand

Lex Fridman (39:21.120)

where it's gonna bind on the surface of the protein.

Lex Fridman (39:23.440)

So obviously in order to do that,

Lex Fridman (39:24.940)

you need to understand the 3D structure.

Lex Fridman (39:26.720)

So the structure is mapped to the function.

Lex Fridman (39:28.640)

The structure is mapped to the function

Lex Fridman (39:29.880)

and the structure is obviously somehow specified

Lex Fridman (39:32.560)

by the amino acid sequence.

Lex Fridman (39:34.840)

And that's the, in essence, the protein folding problem is,

Lex Fridman (39:37.420)

can you just from the amino acid sequence,

Demis Hassabis (39:39.620)

the one dimensional string of letters,

Lex Fridman (39:42.560)

can you immediately computationally predict

Lex Fridman (39:45.600)

the 3D structure?

Lex Fridman (39:47.120)

And this has been a grand challenge in biology

Demis Hassabis (39:50.020)

for over 50 years.

Lex Fridman (39:51.500)

So I think it was first articulated by Christian Anfinsen,

Demis Hassabis (39:54.360)

a Nobel prize winner in 1972,

Lex Fridman (39:57.040)

as part of his Nobel prize winning lecture.

Lex Fridman (39:59.240)

And he just speculated this should be possible

Lex Fridman (40:01.860)

to go from the amino acid sequence to the 3D structure,

Lex Fridman (40:04.960)

but he didn't say how.

Lex Fridman (40:06.060)

So it's been described to me as equivalent

Demis Hassabis (40:09.440)

to Fermat's last theorem, but for biology.

Lex Fridman (40:12.320)

You should, as somebody that very well might win

Demis Hassabis (40:15.120)

the Nobel prize in the future.

Lex Fridman (40:16.560)

But outside of that, you should do more

Demis Hassabis (40:19.240)

of that kind of thing.

Lex Fridman (40:20.080)

In the margin, just put random things

Demis Hassabis (40:22.160)

that will take like 200 years to solve.

Lex Fridman (40:24.440)

Set people off for 200 years.

Demis Hassabis (40:26.000)

It should be possible.

Lex Fridman (40:27.720)

And just don't give any details.

Demis Hassabis (40:29.040)

Exactly.

Lex Fridman (40:29.880)

I think everyone exactly should be,

Demis Hassabis (40:31.500)

I'll have to remember that for future.

Lex Fridman (40:33.520)

So yeah, so he set off, you know,

Demis Hassabis (40:34.800)

with this one throwaway remark, just like Fermat,

Lex Fridman (40:37.040)

you know, he set off this whole 50 year field really

Demis Hassabis (40:42.640)

of computational biology.

Lex Fridman (40:44.400)

And they had, you know, they got stuck.

Demis Hassabis (40:46.240)

They hadn't really got very far with doing this.

Lex Fridman (40:48.520)

And until now, until AlphaFold came along,

Lex Fridman (40:52.500)

this is done experimentally, right?

Lex Fridman (40:54.320)

Very painstakingly.

Lex Fridman (40:55.500)

So the rule of thumb is, and you have to like

Lex Fridman (40:57.440)

crystallize the protein, which is really difficult.

Demis Hassabis (40:59.820)

Some proteins can't be crystallized like membrane proteins.

Lex Fridman (41:03.060)

And then you have to use very expensive electron microscopes

Demis Hassabis (41:05.940)

or X ray crystallography machines.

Lex Fridman (41:08.200)

Really painstaking work to get the 3D structure

Lex Fridman (41:10.680)

and visualize the 3D structure.

Lex Fridman (41:12.400)

So the rule of thumb in experimental biology

Demis Hassabis (41:14.840)

is that it takes one PhD student,

Lex Fridman (41:16.860)

their entire PhD to do one protein.

Lex Fridman (41:20.320)

And with AlphaFold 2, we were able to predict

Lex Fridman (41:23.440)

the 3D structure in a matter of seconds.

Lex Fridman (41:26.400)

And so we were, you know, over Christmas,

Lex Fridman (41:28.700)

we did the whole human proteome

Demis Hassabis (41:30.240)

or every protein in the human body or 20,000 proteins.

Lex Fridman (41:33.280)

So the human proteomes like the equivalent

Demis Hassabis (41:34.760)

of the human genome, but on protein space.

Lex Fridman (41:37.560)

And sort of revolutionized really

Lex Fridman (41:40.240)

what a structural biologist can do.

Lex Fridman (41:43.300)

Because now they don't have to worry

Demis Hassabis (41:45.720)

about these painstaking experimental,

Lex Fridman (41:47.960)

should they put all of that effort in or not?

Demis Hassabis (41:49.560)

They can almost just look up the structure

Lex Fridman (41:51.120)

of their proteins like a Google search.

Lex Fridman (41:53.280)

And so there's a data set on which it's trained

Lex Fridman (41:56.880)

and how to map this amino acid sequence.

Demis Hassabis (41:58.800)

First of all, it's incredible that a protein,

Lex Fridman (42:00.760)

this little chemical computer is able to do

Demis Hassabis (42:02.480)

that computation itself in some kind of distributed way

Lex Fridman (42:05.720)

and do it very quickly.

Demis Hassabis (42:07.800)

That's a weird thing.

Lex Fridman (42:08.840)

And they evolve that way because, you know,

Demis Hassabis (42:10.480)

in the beginning, I mean, that's a great invention,

Lex Fridman (42:13.200)

just the protein itself.

Lex Fridman (42:14.760)

And then there's, I think, probably a history

Lex Fridman (42:18.240)

of like they evolved to have many of these proteins

Lex Fridman (42:22.740)

and those proteins figure out how to be computers themselves

Lex Fridman (42:26.600)

in such a way that you can create structures

Demis Hassabis (42:28.560)

that can interact in complexes with each other

Lex Fridman (42:30.540)

in order to form high level functions.

Demis Hassabis (42:32.660)

I mean, it's a weird system that they figured it out.

Lex Fridman (42:35.520)

Well, for sure.

Demis Hassabis (42:36.360)

I mean, you know, maybe we should talk

Lex Fridman (42:37.640)

about the origins of life too,

Lex Fridman (42:39.000)

but proteins themselves, I think are magical

Lex Fridman (42:41.180)

and incredible, as I said, little bio nano machines.

Lex Fridman (42:45.760)

And actually Leventhal, who was another scientist,

Lex Fridman (42:50.280)

a contemporary of Amphinson, he coined this Leventhal,

Lex Fridman (42:55.120)

what became known as Leventhal's paradox,

Lex Fridman (42:56.820)

which is exactly what you're saying.

Demis Hassabis (42:58.320)

He calculated roughly an average protein,

Lex Fridman (43:01.580)

which is maybe 2000 amino acids base as long,

Demis Hassabis (43:05.080)

is can fold in maybe 10 to the power 300

Lex Fridman (43:09.960)

different confirmations.

Lex Fridman (43:11.480)

So there's 10 to the power 300 different ways

Lex Fridman (43:13.320)

that protein could fold up.

Lex Fridman (43:14.800)

And yet somehow in nature, physics solves this,

Lex Fridman (43:18.160)

solves this in a matter of milliseconds.

Lex Fridman (43:20.520)

So proteins fold up in your body in, you know,

Lex Fridman (43:23.080)

sometimes in fractions of a second.

Lex Fridman (43:25.600)

So physics is somehow solving that search problem.

Lex Fridman (43:29.080)

And just to be clear, in many of these cases,

Demis Hassabis (43:31.200)

maybe you can correct me if I'm wrong,

Lex Fridman (43:33.040)

there's often a unique way for that sequence to form itself.

Lex Fridman (43:37.680)

So among that huge number of possibilities,

Lex Fridman (43:41.240)

it figures out a way how to stably,

Demis Hassabis (43:45.320)

in some cases there might be a misfunction, so on,

Lex Fridman (43:47.800)

which leads to a lot of the disorders and stuff like that.

Lex Fridman (43:50.040)

But most of the time it's a unique mapping

Lex Fridman (43:52.720)

and that unique mapping is not obvious.

Demis Hassabis (43:54.820)

No, exactly.

Lex Fridman (43:55.660)

Which is what the problem is.

Demis Hassabis (43:57.120)

Exactly, so there's a unique mapping usually in a healthy,

Lex Fridman (44:00.720)

if it's healthy, and as you say in disease,

Lex Fridman (44:04.040)

so for example, Alzheimer's,

Lex Fridman (44:05.400)

one conjecture is that it's because of misfolded protein,

Demis Hassabis (44:09.000)

a protein that folds in the wrong way, amyloid beta protein.

Lex Fridman (44:12.040)

So, and then because it folds in the wrong way,

Demis Hassabis (44:14.560)

it gets tangled up, right, in your neurons.

Lex Fridman (44:17.600)

So it's super important to understand

Demis Hassabis (44:20.560)

both healthy functioning and also disease

Lex Fridman (44:23.600)

is to understand, you know, what these things are doing

Lex Fridman (44:26.480)

and how they're structuring.

Lex Fridman (44:27.600)

Of course, the next step is sometimes proteins change shape

Demis Hassabis (44:30.540)

when they interact with something.

Lex Fridman (44:32.160)

So they're not just static necessarily in biology.

Demis Hassabis (44:37.200)

Maybe you can give some interesting,

Lex Fridman (44:39.780)

so beautiful things to you about these early days

Demis Hassabis (44:43.260)

of AlphaFold, of solving this problem,

Lex Fridman (44:46.160)

because unlike games, this is real physical systems

Demis Hassabis (44:51.280)

that are less amenable to self play type of mechanisms.

Lex Fridman (44:55.640)

Sure.

Demis Hassabis (44:56.460)

The size of the data set is smaller

Lex Fridman (44:58.440)

than you might otherwise like,

Lex Fridman (44:59.760)

so you have to be very clever about certain things.

Lex Fridman (45:01.800)

Is there something you could speak to

Lex Fridman (45:04.800)

what was very hard to solve

Lex Fridman (45:06.680)

and what are some beautiful aspects about the solution?

Demis Hassabis (45:09.920)

Yeah, I would say AlphaFold is the most complex

Lex Fridman (45:12.800)

and also probably most meaningful system

Demis Hassabis (45:14.600)

we've built so far.

Lex Fridman (45:15.860)

So it's been an amazing time actually in the last,

Demis Hassabis (45:18.400)

you know, two, three years to see that come through

Lex Fridman (45:20.520)

because as we talked about earlier, you know,

Demis Hassabis (45:23.200)

games is what we started on

Lex Fridman (45:25.480)

building things like AlphaGo and AlphaZero,

Lex Fridman (45:27.900)

but really the ultimate goal was to,

Lex Fridman (45:30.400)

not just to crack games,

Demis Hassabis (45:31.520)

it was just to build,

Lex Fridman (45:33.120)

use them to bootstrap general learning systems

Demis Hassabis (45:35.320)

we could then apply to real world challenges.

Lex Fridman (45:37.440)

Specifically, my passion is scientific challenges

Demis Hassabis (45:40.640)

like protein folding.

Lex Fridman (45:41.920)

And then AlphaFold of course

Demis Hassabis (45:43.280)

is our first big proof point of that.

Lex Fridman (45:45.360)

And so, you know, in terms of the data

Lex Fridman (45:49.040)

and the amount of innovations that had to go into it,

Lex Fridman (45:50.920)

we, you know, it was like

Demis Hassabis (45:52.280)

more than 30 different component algorithms

Lex Fridman (45:54.480)

needed to be put together to crack the protein folding.

Demis Hassabis (45:57.960)

I think some of the big innovations were that

Lex Fridman (46:00.800)

kind of building in some hard coded constraints

Demis Hassabis (46:04.220)

around physics and evolutionary biology

Lex Fridman (46:07.760)

to constrain sort of things like the bond angles

Demis Hassabis (46:11.640)

in the protein and things like that,

Lex Fridman (46:15.400)

a lot, but not to impact the learning system.

Lex Fridman (46:18.040)

So still allowing the system to be able to learn

Lex Fridman (46:21.000)

the physics itself from the examples that we had.

Lex Fridman (46:25.540)

And the examples, as you say,

Lex Fridman (46:26.640)

there are only about 150,000 proteins,

Demis Hassabis (46:28.840)

even after 40 years of experimental biology,

Lex Fridman (46:31.240)

only around 150,000 proteins have been,

Demis Hassabis (46:33.880)

the structures have been found out about.

Lex Fridman (46:35.920)

So that was our training set,

Demis Hassabis (46:37.120)

which is much less than normally we would like to use,

Lex Fridman (46:41.120)

but using various tricks, things like self distillation.

Lex Fridman (46:43.840)

So actually using AlphaFold predictions,

Lex Fridman (46:48.280)

some of the best predictions

Demis Hassabis (46:49.480)

that it thought was highly confident in,

Lex Fridman (46:51.000)

we put them back into the training set, right?

Demis Hassabis (46:53.320)

To make the training set bigger,

Lex Fridman (46:55.440)

that was critical to AlphaFold working.

Lex Fridman (46:58.400)

So there was actually a huge number

Lex Fridman (47:00.160)

of different innovations like that,

Demis Hassabis (47:02.720)

that were required to ultimately crack the problem.

Lex Fridman (47:06.080)

AlphaFold one, what it produced was a distrogram.

Lex Fridman (47:09.720)

So a kind of a matrix of the pairwise distances

Lex Fridman (47:13.600)

between all of the molecules in the protein.

Lex Fridman (47:17.880)

And then there had to be a separate optimization process

Lex Fridman (47:20.440)

to create the 3D structure.

Lex Fridman (47:23.640)

And what we did for AlphaFold two

Lex Fridman (47:25.120)

is make it truly end to end.

Lex Fridman (47:26.920)

So we went straight from the amino acid sequence of bases

Lex Fridman (47:31.720)

to the 3D structure directly

Demis Hassabis (47:33.860)

without going through this intermediate step.

Lex Fridman (47:36.080)

And in machine learning, what we've always found is

Demis Hassabis (47:38.600)

that the more end to end you can make it,

Lex Fridman (47:40.920)

the better the system.

Lex Fridman (47:42.160)

And it's probably because in the end,

Lex Fridman (47:46.160)

the system's better at learning what the constraints are

Demis Hassabis (47:48.560)

than we are as the human designers of specifying it.

Lex Fridman (47:51.920)

So anytime you can let it flow end to end

Lex Fridman (47:54.040)

and actually just generate what it is

Lex Fridman (47:55.400)

you're really looking for, in this case, the 3D structure,

Demis Hassabis (47:58.440)

you're better off than having this intermediate step,

Lex Fridman (48:00.560)

which you then have to handcraft the next step for.

Lex Fridman (48:03.360)

So it's better to let the gradients and the learning

Lex Fridman (48:06.160)

flow all the way through the system from the end point,

Demis Hassabis (48:09.000)

the end output you want to the inputs.

Lex Fridman (48:10.880)

So that's a good way to start on a new problem.

Demis Hassabis (48:13.040)

Handcraft a bunch of stuff,

Lex Fridman (48:14.360)

add a bunch of manual constraints

Demis Hassabis (48:16.640)

with a small end to end learning piece

Lex Fridman (48:18.640)

or a small learning piece and grow that learning piece

Demis Hassabis (48:21.560)

until it consumes the whole thing.

Lex Fridman (48:22.840)

That's right.

Lex Fridman (48:23.680)

And so you can also see,

Lex Fridman (48:25.320)

this is a bit of a method we've developed

Demis Hassabis (48:26.960)

over doing many sort of successful alpha,

Lex Fridman (48:29.640)

we call them alpha X projects, right?

Lex Fridman (48:32.200)

And the easiest way to see that is the evolution

Lex Fridman (48:34.600)

of alpha go to alpha zero.

Lex Fridman (48:36.720)

So alpha go was a learning system,

Lex Fridman (48:39.640)

but it was specifically trained to only play go, right?

Demis Hassabis (48:42.280)

So, and what we wanted to do with first version of alpha go

Lex Fridman (48:45.360)

is just get to world champion performance

Lex Fridman (48:47.520)

no matter how we did it, right?

Lex Fridman (48:49.200)

And then of course, alpha go zero,

Demis Hassabis (48:51.400)

we remove the need to use human games as a starting point,

Lex Fridman (48:55.280)

right?

Lex Fridman (48:56.120)

So it could just play against itself

Lex Fridman (48:57.960)

from random starting point from the beginning.

Lex Fridman (49:00.280)

So that removed the need for human knowledge about go.

Lex Fridman (49:03.720)

And then finally alpha zero then generalized it

Lex Fridman (49:05.960)

so that any things we had in there, the system,

Lex Fridman (49:08.920)

including things like symmetry of the go board were removed.

Lex Fridman (49:12.240)

So the alpha zero could play from scratch

Lex Fridman (49:14.600)

any two player game and then mu zero,

Demis Hassabis (49:16.440)

which is the final, our latest version

Lex Fridman (49:18.360)

of that set of things was then extending it

Lex Fridman (49:20.680)

so that you didn't even have to give it

Lex Fridman (49:22.120)

the rules of the game.

Demis Hassabis (49:23.200)

It would learn that for itself.

Lex Fridman (49:24.880)

So it could also deal with computer games

Demis Hassabis (49:26.600)

as well as board games.

Lex Fridman (49:27.760)

So that line of alpha go, alpha go zero, alpha zero,

Demis Hassabis (49:30.400)

mu zero, that's the full trajectory

Lex Fridman (49:33.480)

of what you can take from imitation learning

Demis Hassabis (49:37.200)

to full self supervised learning.

Lex Fridman (49:40.440)

Yeah, exactly.

Lex Fridman (49:41.640)

And learning the entire structure

Lex Fridman (49:44.720)

of the environment you're put in from scratch, right?

Lex Fridman (49:47.640)

And bootstrapping it through self play yourself.

Lex Fridman (49:51.840)

But the thing is it would have been impossible, I think,

Demis Hassabis (49:53.720)

or very hard for us to build alpha zero

Lex Fridman (49:55.960)

or mu zero first out of the box.

Demis Hassabis (49:58.600)

Even psychologically, because you have to believe

Lex Fridman (50:01.400)

in yourself for a very long time.

Demis Hassabis (50:03.040)

You're constantly dealing with doubt

Lex Fridman (50:04.640)

because a lot of people say that it's impossible.

Demis Hassabis (50:06.680)

Exactly, so it's hard enough just to do go.

Lex Fridman (50:08.640)

As you were saying, everyone thought that was impossible

Demis Hassabis (50:10.920)

or at least a decade away from when we did it

Lex Fridman (50:14.160)

back in 2015, 2016.

Lex Fridman (50:17.320)

And so yes, it would have been psychologically

Lex Fridman (50:20.960)

probably very difficult as well as the fact

Demis Hassabis (50:22.960)

that of course we learn a lot by building alpha go first.

Lex Fridman (50:26.400)

Right, so I think this is why I call AI

Demis Hassabis (50:28.520)

an engineering science.

Lex Fridman (50:29.880)

It's one of the most fascinating science disciplines,

Lex Fridman (50:32.280)

but it's also an engineering science in the sense

Lex Fridman (50:34.200)

that unlike natural sciences, the phenomenon you're studying

Demis Hassabis (50:38.200)

doesn't exist out in nature.

Lex Fridman (50:39.440)

You have to build it first.

Lex Fridman (50:40.880)

So you have to build the artifact first,

Lex Fridman (50:42.480)

and then you can study and pull it apart and how it works.

Demis Hassabis (50:46.480)

This is tough to ask you this question

Lex Fridman (50:50.000)

because you probably will say it's everything,

Lex Fridman (50:51.480)

but let's try to think through this

Lex Fridman (50:54.360)

because you're in a very interesting position

Demis Hassabis (50:56.480)

where DeepMind is a place of some of the most brilliant

Lex Fridman (50:59.520)

ideas in the history of AI,

Lex Fridman (51:01.760)

but it's also a place of brilliant engineering.

Lex Fridman (51:05.880)

So how much of solving intelligence,

Demis Hassabis (51:08.040)

this big goal for DeepMind,

Lex Fridman (51:09.880)

how much of it is science?

Lex Fridman (51:12.120)

How much is engineering?

Lex Fridman (51:13.320)

So how much is the algorithms?

Lex Fridman (51:14.720)

How much is the data?

Lex Fridman (51:16.160)

How much is the hardware compute infrastructure?

Lex Fridman (51:19.840)

How much is it the software compute infrastructure?

Lex Fridman (51:23.960)

What else is there?

Lex Fridman (51:24.800)

How much is the human infrastructure?

Lex Fridman (51:27.200)

And like just the humans interacting certain kinds of ways

Demis Hassabis (51:30.280)

in all the space of all those ideas.

Lex Fridman (51:31.720)

And how much is maybe like philosophy?

Lex Fridman (51:33.640)

How much, what's the key?

Lex Fridman (51:35.160)

If you were to sort of look back,

Demis Hassabis (51:40.680)

like if we go forward 200 years and look back,

Lex Fridman (51:43.200)

what was the key thing that solved intelligence?

Lex Fridman (51:46.320)

Is it the ideas or the engineering?

Lex Fridman (51:47.800)

I think it's a combination.

Demis Hassabis (51:49.040)

First of all, of course,

Lex Fridman (51:49.880)

it's a combination of all those things,

Lex Fridman (51:51.360)

but the ratios of them changed over time.

Lex Fridman (51:54.760)

So even in the last 12 years,

Lex Fridman (51:57.480)

so we started DeepMind in 2010,

Lex Fridman (51:59.400)

which is hard to imagine now because 2010,

Demis Hassabis (52:01.920)

it's only 12 short years ago,

Lex Fridman (52:03.400)

but nobody was talking about AI.

Demis Hassabis (52:05.600)

I don't know if you remember back to your MIT days,

Lex Fridman (52:07.600)

no one was talking about it.

Demis Hassabis (52:08.720)

I did a postdoc at MIT back around then.

Lex Fridman (52:11.080)

And it was sort of thought of as a,

Demis Hassabis (52:12.880)

well, look, we know AI doesn't work.

Lex Fridman (52:14.200)

We tried this hard in the 90s at places like MIT,

Demis Hassabis (52:17.040)

mostly using logic systems and old fashioned,

Lex Fridman (52:19.880)

sort of good old fashioned AI, we would call it now.

Demis Hassabis (52:22.600)

People like Minsky and Patrick Winston,

Lex Fridman (52:25.320)

and you know all these characters, right?

Lex Fridman (52:26.720)

And used to debate a few of them.

Lex Fridman (52:28.280)

And they used to think I was mad thinking about

Demis Hassabis (52:30.120)

that some new advance could be done with learning systems.

Lex Fridman (52:32.360)

And I was actually pleased to hear that

Demis Hassabis (52:34.720)

because at least you know you're on a unique track

Lex Fridman (52:36.960)

at that point, right?

Demis Hassabis (52:37.840)

Even if all of your professors are telling you you're mad.

Lex Fridman (52:41.880)

And of course in industry,

Demis Hassabis (52:43.840)

we couldn't get, it was difficult to get two cents together,

Lex Fridman (52:47.680)

which is hard to imagine now as well,

Demis Hassabis (52:48.920)

given that it's the biggest sort of buzzword in VCs

Lex Fridman (52:51.560)

and fundraisings easy and all these kinds of things today.

Lex Fridman (52:54.720)

So back in 2010, it was very difficult.

Lex Fridman (52:57.720)

And the reason we started then,

Lex Fridman (52:59.360)

and Shane and I used to discuss

Lex Fridman (53:02.480)

what were the sort of founding tenants of DeepMind.

Lex Fridman (53:04.920)

And it was various things.

Lex Fridman (53:06.120)

One was algorithmic advances.

Lex Fridman (53:08.680)

So deep learning, you know,

Lex Fridman (53:09.760)

Jeff Hinton and Co had just sort of invented that

Demis Hassabis (53:12.360)

in academia, but no one in industry knew about it.

Lex Fridman (53:15.200)

We love reinforcement learning.

Demis Hassabis (53:16.640)

We thought that could be scaled up.

Lex Fridman (53:18.240)

But also understanding about the human brain

Demis Hassabis (53:20.160)

had advanced quite a lot in the decade prior

Lex Fridman (53:23.920)

with fMRI machines and other things.

Lex Fridman (53:25.440)

So we could get some good hints about architectures

Lex Fridman (53:28.840)

and algorithms and sort of representations maybe

Demis Hassabis (53:32.480)

that the brain uses.

Lex Fridman (53:33.400)

So at a systems level, not at a implementation level.

Lex Fridman (53:37.760)

And then the other big things were compute and GPUs, right?

Lex Fridman (53:41.040)

So we could see a compute was going to be really useful

Lex Fridman (53:44.160)

and had got to a place where it become commoditized

Lex Fridman (53:46.960)

mostly through the games industry

Lex Fridman (53:48.560)

and that could be taken advantage of.

Lex Fridman (53:50.760)

And then the final thing was also mathematical

Lex Fridman (53:52.800)

and theoretical definitions of intelligence.

Lex Fridman (53:54.960)

So things like AIXI, AIXE,

Demis Hassabis (53:57.560)

which Shane worked on with his supervisor, Marcus Hutter,

Lex Fridman (54:00.160)

which is this sort of theoretical proof really

Demis Hassabis (54:03.360)

of universal intelligence,

Lex Fridman (54:05.280)

which is actually a reinforcement learning system

Demis Hassabis (54:08.000)

in the limit.

Lex Fridman (54:08.840)

I mean, it assumes infinite compute and infinite memory

Demis Hassabis (54:10.640)

in the way, you know, like a Turing machine proves.

Lex Fridman (54:12.920)

But I was also waiting to see something like that too,

Demis Hassabis (54:15.840)

to, you know, like Turing machines and computation theory

Lex Fridman (54:19.440)

that people like Turing and Shannon came up with

Demis Hassabis (54:21.520)

underpins modern computer science.

Lex Fridman (54:24.800)

You know, I was waiting for a theory like that

Demis Hassabis (54:26.400)

to sort of underpin AGI research.

Lex Fridman (54:28.880)

So when I, you know, met Shane

Lex Fridman (54:30.120)

and saw he was working on something like that,

Lex Fridman (54:32.000)

you know, that to me was a sort of final piece

Demis Hassabis (54:33.680)

of the jigsaw.

Lex Fridman (54:34.560)

So in the early days, I would say that ideas

Demis Hassabis (54:38.320)

were the most important.

Lex Fridman (54:40.040)

You know, for us, it was deep reinforcement learning,

Demis Hassabis (54:42.440)

scaling up deep learning.

Lex Fridman (54:44.600)

Of course, we've seen transformers.

Lex Fridman (54:46.240)

So huge leaps, I would say, you know, three or four

Lex Fridman (54:48.920)

from, if you think from 2010 till now,

Demis Hassabis (54:51.520)

huge evolutions, things like AlphaGo.

Lex Fridman (54:53.680)

And maybe there's a few more still needed.

Lex Fridman (54:57.920)

But as we get closer to AI, AGI,

Lex Fridman (55:02.000)

I think engineering becomes more and more important

Lex Fridman (55:04.600)

and data because scale and of course the recent,

Lex Fridman (55:07.800)

you know, results of GPT3 and all the big language models

Lex Fridman (55:10.440)

and large models, including our ones,

Lex Fridman (55:12.800)

has shown that scale and large models

Demis Hassabis (55:16.000)

are clearly gonna be a necessary,

Lex Fridman (55:18.080)

but perhaps not sufficient part of an AGI solution.

Lex Fridman (55:21.960)

And throughout that, like you said,

Lex Fridman (55:24.560)

and I'd like to give you a big thank you.

Demis Hassabis (55:26.720)

You're one of the pioneers in this is sticking by ideas

Lex Fridman (55:30.640)

like reinforcement learning, that this can actually work

Demis Hassabis (55:34.560)

given actually limited success in the past.

Lex Fridman (55:38.480)

And also, which we still don't know,

Lex Fridman (55:41.480)

but proudly having the best researchers in the world

Lex Fridman (55:46.760)

and talking about solving intelligence.

Lex Fridman (55:49.400)

So talking about whatever you call it,

Lex Fridman (55:50.920)

AGI or something like this, speaking of MIT,

Demis Hassabis (55:54.720)

that's just something you wouldn't bring up.

Lex Fridman (55:57.240)

Not maybe you did in like 40, 50 years ago,

Lex Fridman (56:03.560)

but that was, AI was a place where you do tinkering,

Lex Fridman (56:09.320)

very small scale, not very ambitious projects.

Lex Fridman (56:12.560)

And maybe the biggest ambitious projects

Lex Fridman (56:16.160)

were in the space of robotics

Lex Fridman (56:17.480)

and doing like the DARPA challenge.

Lex Fridman (56:19.200)

But the task of solving intelligence and believing you can,

Demis Hassabis (56:23.400)

that's really, really powerful.

Lex Fridman (56:24.560)

So in order for engineering to do its work,

Demis Hassabis (56:27.680)

to have great engineers, build great systems,

Lex Fridman (56:30.960)

you have to have that belief,

Demis Hassabis (56:32.360)

that threads throughout the whole thing

Lex Fridman (56:33.920)

that you can actually solve

Demis Hassabis (56:35.040)

some of these impossible challenges.

Lex Fridman (56:36.640)

Yeah, that's right.

Lex Fridman (56:37.480)

And back in 2010, our mission statement and still is today,

Lex Fridman (56:42.280)

it was used to be solving step one, solve intelligence,

Demis Hassabis (56:45.600)

step two, use it to solve everything else.

Lex Fridman (56:47.520)

So if you can imagine pitching that to a VC in 2010,

Demis Hassabis (56:51.120)

the kind of looks we got,

Lex Fridman (56:52.680)

we managed to find a few kooky people to back us,

Lex Fridman (56:55.880)

but it was tricky.

Lex Fridman (56:57.680)

And it got to the point where we wouldn't mention it

Demis Hassabis (57:00.160)

to any of our professors because they would just eye roll

Lex Fridman (57:03.120)

and think we committed career suicide.

Lex Fridman (57:05.760)

And so it was, there's a lot of things that we had to do,

Lex Fridman (57:10.040)

but we always believed it.

Lex Fridman (57:11.560)

And one reason, by the way,

Lex Fridman (57:13.240)

one reason I've always believed in reinforcement learning

Demis Hassabis (57:16.160)

is that if you look at neuroscience,

Lex Fridman (57:19.120)

that is the way that the primate brain learns.

Demis Hassabis (57:22.720)

One of the main mechanisms is the dopamine system

Lex Fridman (57:24.880)

implements some form of TD learning.

Demis Hassabis (57:26.440)

It was a very famous result in the late 90s

Lex Fridman (57:29.680)

where they saw this in monkeys

Lex Fridman (57:31.320)

and as a propagating prediction error.

Lex Fridman (57:34.520)

So again, in the limit,

Demis Hassabis (57:36.800)

this is what I think you can use neuroscience for is,

Lex Fridman (57:39.480)

at mathematics, when you're doing something as ambitious

Demis Hassabis (57:43.160)

as trying to solve intelligence

Lex Fridman (57:44.560)

and it's blue sky research, no one knows how to do it,

Demis Hassabis (57:47.760)

you need to use any evidence

Lex Fridman (57:50.160)

or any source of information you can

Demis Hassabis (57:52.120)

to help guide you in the right direction

Lex Fridman (57:54.280)

or give you confidence you're going in the right direction.

Lex Fridman (57:56.680)

So that was one reason we pushed so hard on that.

Lex Fridman (57:59.840)

And just going back to your earlier question

Demis Hassabis (58:01.840)

about organization, the other big thing

Lex Fridman (58:04.280)

that I think we innovated with at DeepMind

Demis Hassabis (58:06.000)

to encourage invention and innovation

Lex Fridman (58:10.320)

was the multidisciplinary organization we built

Lex Fridman (58:12.920)

and we still have today.

Lex Fridman (58:14.160)

So DeepMind originally was a confluence

Demis Hassabis (58:16.680)

of the most cutting edge knowledge in neuroscience

Lex Fridman (58:19.400)

with machine learning, engineering and mathematics, right?

Lex Fridman (58:22.840)

And gaming.

Lex Fridman (58:24.400)

And then since then we've built that out even further.

Lex Fridman (58:26.760)

So we have philosophers here and ethicists,

Lex Fridman (58:30.280)

but also other types of scientists, physicists and so on.

Lex Fridman (58:33.160)

And that's what brings together,

Lex Fridman (58:35.160)

I tried to build a sort of new type of Bell Labs,

Lex Fridman (58:38.760)

but in its golden era, right?

Lex Fridman (58:41.200)

And a new expression of that to try and foster

Demis Hassabis (58:45.680)

this incredible sort of innovation machine.

Lex Fridman (58:48.480)

So talking about the humans in the machine,

Demis Hassabis (58:50.600)

DeepMind itself is a learning machine

Lex Fridman (58:53.080)

with lots of amazing human minds in it

Demis Hassabis (58:55.600)

coming together to try and build these learning systems.

Lex Fridman (59:00.360)

If we return to the big ambitious dream of AlphaFold,

Demis Hassabis (59:04.960)

that may be the early steps on a very long journey

Lex Fridman (59:08.400)

in biology, do you think the same kind of approach

Demis Hassabis (59:14.200)

can use to predict the structure and function

Lex Fridman (59:16.400)

of more complex biological systems?

Lex Fridman (59:18.720)

So multi protein interaction,

Lex Fridman (59:21.480)

and then, I mean, you can go out from there,

Demis Hassabis (59:24.400)

just simulating bigger and bigger systems

Lex Fridman (59:26.920)

that eventually simulate something like the human brain

Demis Hassabis (59:29.560)

or the human body, just the big mush,

Lex Fridman (59:32.560)

the mess of the beautiful, resilient mess of biology.

Lex Fridman (59:36.480)

Do you see that as a long term vision?

Lex Fridman (59:39.600)

I do, and I think, if you think about

Lex Fridman (59:42.560)

what are the top things I wanted to apply AI to

Lex Fridman (59:45.680)

once we had powerful enough systems,

Demis Hassabis (59:47.680)

biology and curing diseases and understanding biology

Lex Fridman (59:52.240)

was right up there, top of my list.

Demis Hassabis (59:54.120)

That's one of the reasons I personally pushed that myself

Lex Fridman (59:56.760)

and with AlphaFold, but I think AlphaFold,