DeepMind’s Demis Hassabis on its breakthrough scientific discoveries | WIRED Live

I'm delighted, this morning,

to be joined by Demis Hassabis,

the CEO and Co-founder of DeepMind.

Demis has led an absolutely fascinating career.

He's the a former chess prodigy,

the recipients of a double first

at the University of Cambridge,

a five time World Mind Sports Olympiad champion,

an MIT in Harvard alumnus,

and a teenage entrepreneur.

Today, DeepMind is one of the world's

leading AI research companies,

and it's best known for developing AlphaGo,

the first program to beat a world champion at Go.

DeepMind's, published over a thousand research papers,

including more than a dozen in nature and science,

and achieve breakthrough results

in many challenging AI domains.

Demis, welcome, great to have you with us today,

thanks for coming in in person.

No problem, it's great to be here.

Thank you so much, so last time we met,

we talked a lot about the way

that DeepMind is now moving into science

and is now thinking about how it can impact,

how AI can impact scientific research.

So I'm interested, really,

can you just describe that sort of, that move to us.

Yeah, well, I mean, when we started DeepMind

and actually even before DeepMind, for me,

the ultimate vision of building AI was to try

and use it as a tool

to understand the world around us better.

[Interviewer] Yeah.

That's what I was dreaming about

when my teenage years, when I first got into AI

and I've been working towards that my whole career

and obviously DeepMind for the last 10 years.

And what's exciting now is that maybe we've got

to the point finally

where our algorithms are powerful enough

and mature enough that we can actually apply it

to big scientific challenges.

[Interviewer] Yeah.

And maybe help accelerate scientific discovery.

So, you know, we are kind of most famous

for our work on games, things like AlphaGo as you mentioned,

but really that was always just a proving ground

for developing and testing

and sort of proving out these algorithms efficiently

and then the idea was to transfer them

to things like science.

And how do you think

that scientific research is best organized

in order to get optimal results?

Well, I think there's, you know, there's different ways

to try and organize scientific research.

The main one, academia--

[Interviewer] Yeah.

Which, you know, obviously I spent quite a lot

of time in is mostly bottom up, I would say.

So the creativity sort of bubbles up from PhD students,

postdocs and so on.

And it's kind of a creative chaos, let's say.

[Interviewer] Yeah.

And then on the other hand,

you've got companies, startups, the best startups,

which are mostly top down.

And the great thing about those

is they come with a lot of energy and pace and focus

and intensity that you get

with the best startups, as you know well.

And I've always wondered why,

and I've been lucky enough to be in both worlds,

why you couldn't combine the best of those two worlds

and have a kind of, you know,

blue sky research group or division,

but with the same intensity you get

from a startup kind of mentality

and that's what we tried to do at DeepMind.

So in academia generally, you know,

you hire the smartest

or you recruit the smartest people you possibly can.

You put 'em in a lab.

You say, see you five years and you close the door.

for the best, close the door,

and when you open it five years later

and maybe you have something, maybe you don't.

[Interviewer] Yeah.

And, you know, it's not very coordinated, so it's not

that efficient in many ways

because you, as you say, in the top places,

top universities, you have some

of the smartest people in the world there.

So the ingredients are there

and you give them time to think and so on,

but there's no sort of macro coordination between, you know,

beyond the kind of small lab level.

Like each lab is maybe coordinated,

but there's no coordination on a bigger level than that.

And, in fact, it's designed for there not to be.

[Interviewer] Yeah.

And so, you know, it makes it hard to really go

after massive breakthroughs in an intense way

over multi-years, especially if it's interdisciplinary.

[Interviewer] Yeah.

So there's some things like that that are quite hard

to do actually in academia, I would say.

I Think the interdisciplinary bit's interesting

because as far as I can understand,

what you're doing at DeepMind is trying to build teams

that have, you know, multiple sort of areas,

domains of expertise.

Absolutely, so obviously our core thing

is machine learning.

[Interviewer] Yeah.

But we also have neuroscientists.

We have mathematicians, physicists,

and, you know, and, of course, engineering.

[Interviewer] Yeah.

And engineering is our kind of workbench if you like.

If, you know, the engineering is,

it makes an empirical science,

and obviously we have some of the world's top engineers,

but in academia, for example,

if you're working in computer science, there isn't actually,

weirdly, there isn't a career path for engineers.

You have to be a research scientist

and then you've got the, you know,

traditional PhD student, postdoc, and then you've gotta try

and make it to assistant prof.

But if you want to be a career engineer in academia,

there isn't really a sort of defined career path.

So you obviously you end up losing some

of the best engineers out of that.

[Interviewer] Yeah.

So there's many, many sort of strange things like that

that would be odd to somebody who's not used to academia.

Like why does it work like that?

It's just kind of the way it's always worked.

And how do you get those individuals

who have deep expertise of a particular domain

to kind of communicate with each other?

Yeah, because a biologist isn't always gonna be able

to speak to a chemist or a mathematician

or a computer scientist.

[Interviewer] Right.

That's the hard, absolute hardest thing.

So, first of all, you need to hire people with curiosity.

[Interviewer] Yeah.

And also I would say a bit of humbleness

because it takes some humility to approach someone else.

Say you are a world expert in one of those domains, right,

but you are, you know, relative beginner

in these other domains and, you know,

it takes some vulnerability

and humility to go to someone else who's super expert

in the other domain and kind of, you know,

explain you don't know that much about that

when you are used to being the one that explains your area.

[Interviewer] Sure.

And, and I feel like

that's one big reason why there aren't

that many true interdisciplinary people

because it's quite hard to the ego

to go and do that.

I mean, I experienced that when I had a, you know,

whole first career in compete science,

and then I went back to university

to my PhD in neuroscience.

Yeah. And I was starting as a beginner again

and at, you know, sort of the bottom of the pile

after seven years of running my own games company

and so it's quite a, you know, it takes a certain sort

of psyche to be able to deal with that.

And then on the other thing is what we look for actually

at DeepMind is, and I try and hire for,

is what I call sort of affectionately glue people,

which is people that really are sitting in that intersection

of one or more subject areas or disciplines

and can do that translation, spot the connections

and you don't need everyone to be like that.

You just need a few people like that

and who can, who can operate at the level

where they can understand

and explain things to the other world class experts

in their more narrow domains

and then make the connections for them.

Yeah.

And so I say, ah, you know,

you should talk to this person really,

because what you are actually talking about is similar.

And that's one of the things I do actually

at Deep Mind is to try and be a generalist

and make those connections.

Yeah, that I was about, say there are many glue people out

there, they must be quite hard to find it.

They're really hard.

I would say we have, you know,

maybe out of a thousand people,

like a couple of dozen of those.

[Interviewer] Right.

So they're very rare.

[Interviewer] Yeah.

And it is rare because it's a bit like a decathlete.

They, you know, decathlon sort of...

It's hard, the whole way that academia's structured.

Yeah.

It's hard for those interdisciplinary people

because usually when you get evaluated for a position

or a grant,

normally you, you are judged

by an expert panel in your nominal subject.

and they don't care

or know about these other extraneous things

that you know about,

which are these interdisciplinary part.

So you sort of have to compete

on the narrow domain with the other people

who are only doing that,

whilst keeping your general interest going.

So it's essentially, it's quite a lot more work

and a lot harder to do that.

So one of the areas that you really working hard on,

I know at the moment we talked about it when we last met,

you know, to discuss

the wide feature was protein folding.

[Demis] Yeah.

Just give us a sense maybe if the audience

on why you felt that particular challenge was one

that you wanted to sort of like apply,

you know, DeepMind's resources.

Sure, well, when we go after big problem like that,

there's a lot of evaluation we do beforehand.

So partly one of the main thing, first of starting points is

to make sure the problem is a big enough

impact if you were to solve it.

If you're gonna spend 3, 4, 5 years

trying to solve something, you better make sure that

it's something that would unlock a lot of new potential.

And secondly,

we also look for things like properties of the problem.

Do they suit the types of algorithms we're building?

So the kinds of things we look for is there,

is there enough training data?

or is there even better,

is there simulations we can create more synthetic data from?

Is there a clear objective function

that you're trying to optimize?

So, you know, something that some clear metric

that your AI system can hill climb towards.

And then we also look for things

like clear external benchmarks, maybe a competition

or something that runs by the community

that you're going into

that you can clearly measure your progress against,

and protein folding ticked all of those boxes.

So it, you know, it has some data,

it has some great competitions,

this thing called CASP every two years,

which is a community run competition where you have

to predict the structure of a protein

before it's revealed by experimentalists,

so it's a fantastic competition.

And in terms of how important it is, I mean,

proteins are essential to every function in your body.

So it is key for disease

and so if we could understand the shape they fold into,

then maybe we could accelerate things like drug discovery.

So clearly if we could crack that problem,

it should have a lot of downstream impact.

You mentioned CASP,

I know there's one going on at the moment.

Clearly you can't sort of like talk about, you know,

any kind of, you know, insider knowledge you might have

about what's going on,

but like that's a really important proving point, isn't it?

Because it shows the research going in the right direction.

That's right, and CASP is this incredible,

it's in the 14th edition now,

so it's been going since mid-90s

[Interviewer] Yeah

And yes, actually the results are gonna come out

next week and, you know, we think we've done pretty well,

but obviously, you know, it's confidential

until the, the routes results are officially announced.

But it's a great example, I think in science

of one of these really rigorous benchmarks

that have been run and it's fantastic for push, you know,

allowing the field to progress

and making sure that you really are making progress

towards the ultimate goal.

Not, not kind of kidding yourself on the way somehow

that you know, that your internal metrics are one thing.

You obviously, you are always measuring

how good are your algorithms internally

and your own benchmarks.

[Interviewer] Yeah.

But there's nothing better

than like an external measure

that's judged by an independent assessors

to really test your metal.

Well, we're gonna be looking out for those results next.

Yeah.

We can't have a conversation in 2020,

obviously without talking about Coronavirus.

So I'd love to get your thoughts on

whether DeepMind's been working

on this particular challenge

and how machine learning as far as you know,

it has kind of played out and had an impact

on drug discovery or you know,

what we understand about the pandemic.

Yeah, I mean, so us particularly, I've been,

I mean I've been working on it both professionally

at DeepMind and also on a personal level

as a scientist.

With DeepMind,

what we've done is actually very early on,

the Covid virus was sequenced

by I think some Chinese researchers, genetically sequenced.

And so what we did is we used an earlier version

of AlphaGo, which we thought was pretty good already

in March to basically look at the structures,

some of the proteins in the virus that were understudied.

A few of them, we already know the structure

experimentally, but some of them we didn't.

And we put our best predictions out,

open source that to community straight away

so that people could use it

and maybe target against that.

But we didn't make a big claim about that

because at the time, obviously the cast competition

hadn't been run yet.

that was run over the summer, so we hadn't had external

validation of how good our system was.

We thought it was pretty good,

but we couldn't,

we didn't want people necessarily to rely on that

until it was externally validated.

[Interviewer] Yeah.

So, God forbid if something like that happens again,

you know, in a future pandemic 3, 4, 5 years,

I would imagine AI playing a much bigger part than it did

this time around, where I think in some senses

it came a a little bit too early

for AI to be at the forefront

of obviously what the great scientific effort

that's going on.

And we've seen a lot of, you know,

new hope in the last few weeks with the vaccines and so on.

Obviously the whole scientific communities come together

and I think AI will be just one component of something like

that in the future, but maybe, you know,

a much more important component

than it was this time around.

And then on a personal level, you know, I've worked

with the Royal Society to try and put out studies

and things like masks and the effect of opening schools

and other things bring together

some of the top scientists in the UK

to help write white papers to advise government.

So let's talk a little bit about this discussion

around algorithmic sort of bias and data bias.

Obviously, AI systems can amplify this in some ways.

How do you think about that in terms of what, you know,

DeepMind can do to sort of combat that?

Yeah, so this is a hugely important question.

Obviously, we have a whole research team looking at this.

You know, we kind of put it under the rubric

of fairness, bias and interpretability.

These are all key things that we need

to understand about systems before you deploy them.

A lot of what we do is pure research.

So it's not, you know,

I think it becomes more pressing

when you actually create a product out of your research.

and then it's used in the real world to do something

that affects people's lives.

So a lot more research has to be done on that,

we work on that.

We also collaborate with places like Turing Institute

and other places to sort of look into that further.

and I think the next phase

of development is gonna be building analysis tools ,

and visualization tools to look inside these sort of,

so-called black boxes of these neural network systems

And understanding better what they do.

And I think there we can take inspiration from

how we analyze the brain using neuroscience

and things like FMI machines and so on.

What's the equivalent of that for AI systems?

And just on the optimistic side of, of this, this problem is

that of course humans are biased in many ways ourselves.

And if we build it, the AI system's wrong,

it will pl you know,

they will amplify the biases we already have

or designers and other people already have.

If we build it right,

then potentially the AI systems could be less biased

than we are as individuals.

So I think it's sort of, it's a bifurcation.

If we do well with it,

I think it could actually make the societal problem

of fairness and bias better than it currently is.

So clearly we've had a global pandemic.

AI is also a kind of like a global technology

and to some degree there's a lot of competition going on

to sort of, you know, advance research

as quickly as possible.

We have obviously China and the United States.

The UK it is gonna be very important

obviously for us, you know,

as we kind of embark

on our post Brexit sort of like journey.

Where do you think we are in the UK

and how best can we kind of push deep tech forward,

and really become a center of excellence, do you think?

I mean,

I think that the UK you know,

we've always punched well above our weight

in terms of pure research.

So we've got world class universities, you know,

we always have several in the top 10 in the world.

And if you look at other measures like Nobel prizes

or citations and big papers,

we do phenomenally well.

I think that the, the UK needs to really cherish that

and build on that for starters.

And that includes like post Brexit,

making sure we're still plugged into

all the European sort of frameworks on research frameworks

'cause we generally get more grants

than we put money in.

So that needs to continue somehow

and also work, keep welcoming the top talent here.

In terms of AI, obviously, you know, DeepMind's still here.

We got where we are located here

with pretty much all of our staff.

There are many, you know,

I think we've helped create a big ecosystem in the UK now

of AI startups.

So I think we do very well on that.

So I think there's a lot going well,

we just need to carry on building on that momentum

and investing in things like scholarships

for underrepresented groups to broaden access

to AI technologies

and to deepen where we're strong, you know,

in certain universities, maybe increased grants

to those and so on.

From a government level, we are doing that as DeepMind,

like sponsoring a lot of master scholarships

and also I do that philanthropically

as well in my own personal way.

So I think all of those things are gonna add.

And I think in the future, you know,

you've got the UK you've got places like Canada, France,

where there's actually a lot of good AI work going on

that maybe it could act as a counterweight

to the two superpowers.

I think in general, let's say you take Canada, UK,

and France together, probably, you know,

that that is as there's a much going on

in those three countries put together

as there is in, in those two superpowers.

So I think there's a lot of opportunity for the UK actually

to have a big voice at the table,

big say on how this goes.

It's good to hear that optimism.

So let's maybe take a look

at some of the questions coming in from the audience is.

We've been talking a lot this morning about sustainability.

How do you think AI and DeepMind in particular can kind

of have an impact on clean tech?

Yeah.

Which seems to be an area that, you know,

is really developing very quickly.

Yeah, I think it's one of my passion areas, actually.

I think AI has a lot to play

and we've done quite a few think projects already.

Actually our more sort of applied projects

to the better known ones are

applying our AI systems actually very similar ones to,

to AlphaGo to control the cooling systems in data centers.

Yeah, and data centers use, you know,

large amount of energy.

Actually Google's ones are mostly renewable energy now,

which is great, but even still, we save 30%

of the energy used in the data center

by just more efficiently operating the cooling systems.

You know, there's hundreds of different switches and pumps

and fans and things you can turn on.

and AI is perfectly suited to managing that.

And then we've developed that further recently

as a service on cloud, Google cloud.

We call it,

you know, sort of building adaptive controls.

So big industrial buildings, office buildings,

I guess not so much now during Covid,

but in normal times use a huge amount of energy.

Right, again, with the, you know, air conditioning

and heating and that stuff.

And it's the same kind of principle

that we use in the data centers,

but now in terms of controlling all the building systems

that govern, you know,

climate temperature and all that sort of stuff

and that saves a huge amount of obviously money,

but also energy

by more efficiently running all those systems.

Sure, another question coming in from the audience,

there was one about neuropsychology.

How influential is it when developing AI solutions

and an area that dear to your heart.

Yeah, so it's actually very influential

on a certain type of level.

So what we're not trying to do when we say we are

neuroscience inspired is we're not trying to copy

how the brain works, right,

like reverse engineer,

some people are trying to do that,

but we don't think that's the right way to build AI

because, of course, there's different massive differences

obviously between a carbon based system like our brains

and a silicon based system like a computer.

[Interviewer] Sure.

So the implementation details will be different,

but what you are interested in is the principles

of general intelligence.

So the architecture, the representations and the algorithms.

And so we call that systems neuroscience

and that's where the level that we look at neuroscience

and psychology at to get inspiration.

Okay, thank you,

Another question around how you balance, you know,

the work for Google

and Google products, clearly you have to deliver sort

of products for Google,

how best to balance that relationship?

Yeah, very big,

I mean, you can think of it

as they're our biggest client in a way, right.

If you think about all the things, different things we do

and so it's obviously

we've actually had a hundred product launches

now within Google.

So when, when you use Google devices, you may not know it,

but most of the time you'll be using,

some of our tech will be under the hood.

Probably the best example of that is WaveNet,

which is the world's best text-to-speech system.

and we developed that a few years ago and scaled it up.

And now pretty much any device you speak to on Android

or Google device anywhere, the voice that's speaking back

to you will be WaveNet our technology,

so that's just one example.

On the other hand,

obviously, we do a lot of other partnerships

and there are a lot of other work that we do that's outside

of Google, especially in the scientific realm,

so it's quite a nice balance.

And the, and one nice thing about Google is that

when you build some research

and you do discover something new, you can straight away,

put it in a product that reaches a billion people.

So that's fantastic for impact

and also for getting, you know,

back sort of real world information about how,

how good are your outcomes really,

you know, once they go out of the lab.

Yeah, absolutely,

once people actually using the product,

Using it at scale.

Absolutely.

Yeah.

So a final question from the audience

that synthetic data could be tricky.

How do you manage that avoid issue, including biases?

Yeah, well, synthetic data, actually,

interestingly, I mean, we are experts in synthetic data

because we started with games and virtual worlds

and where we generate all of our own data, actually,

and one reason I did that was when we were startup,

obviously a small startup, we had no customers and no data.

So how, how would we compete with, you know,

big companies that had obviously all their consumer data?

And one of the answers was to use games

and then generate that ourselves with like Atari games

and run the simulators and then generate our own data.

[Interviewer] Yeah, yeah.

So that was one of the,

I guess the founding principles of Deep Mind.

One of the reasons we chose games obviously is

also my background in games.

And so,

but of course, you know, as we've now matured,

you've gotta be careful with synthetic data.

You're not generating in, in by, in a biased way,

but just like we discussed with fairness and bias,

it's also the optimistic side of that is

that if you generate synthetic data,

you can mathematically analyze it

and make sure that it's not biased,

which is, which you can't really do with real data

because you, you just have whatever data you have.

[Interviewer] Yeah.

So I think synthetic data has the potential

if you analyze it in the right way to be more balanced.

Okay. One final question,

your a chess prodigy,

have you been watching The Queen's Gambit?

And if so, what do you think of it?

I have, it seems like me

and then the rest of the world

and I watched it actually as soon as it came out.

Of course, how could I not given it was about chess

and it was such a great story.

I thought it was fantastic

and I actually thought it was extremely realistic portrayal

in many ways of what it's like

to be a child chess prodigy and the exhilaration of it,

but also the pressure perhaps minus the psychedelic drugs.

But, I mean, and the chess was extremely accurate,

which is unusual in films,

but I guess it's because Garry Kasparov,

I found out afterwards, was a consultant on it, so.

[Interviewer] Oh, I see.

So I guess he made sure

that the chess part was really accurate.

Well, but I recommend it.

Well, if people out there

aren't watching the DeepMind documentary on Netflix,

The Queen's Gambit is the second best.

Demis, thank you so much for joining us,

really delightful to have you here at Wired Live

Thank you.