April 1, 2026 • 40 mins
What drives a man to turn down half a million pounds at 18, test Mark Zuckerberg's sincerity over dinner, and wonder aloud if he can win a second Nobel Prize? For Demis Hassabis, co-founder and CEO of Google DeepMind, the answer is a lifelong pursuit of artificial general intelligence — and an unshakeable belief that the technology he's creating will change everything about what it means to be human. Oz speaks with journalist and author Sebastian Mallaby about his new book, The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence, tracing Demis's extraordinary journey from chess prodigy to the man at the center of the most consequential technological race of our time.
See omnystudio.com/listener for privacy information.
Listen
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Welcome to tech Stuff. I'm Os Voloscian, and today we
get the opportunity to go behind the curtain at Google's
Deep Mind. For almost three years, in the upstairs room
of a pub in North London, journalist Sebastian Malabi met
regularly with the company's CEO and co founder Demis has ABIs.
They spoke about artificial intelligence, philosophy, neuroscience, motivation and consequence,
(00:40):
all against the backdrop of an increasingly intense three way
race between Open Ai, Anthropic and Google to win the
race towards Agi. Sebasti, congratulations on your new book, The
Infinity Machine, and welcome to tex Stuff.
Speaker 2 (00:55):
Thank you ask going to be here.
Speaker 1 (00:56):
You begin the book with a quote from other scientists
who watch on Manhattan Project, who said, what we're creating
now is a monster whose influence is going to change history.
Yet it would be impossible not to see it through
the energy source which is now being made available, or
make scientists the most hated and the most wanted citizens
of any country where.
Speaker 2 (01:18):
Reliving that now with Ai argree, I mean I began
this project wanting to capture the tingling sensation of human beings,
like Demsis ABIs, creating the new version of atomic weapons. Right,
this incredibly powerful AI technology that has enormous upsides but
could also be very, very dangerous. And the surprise was
(01:38):
I didn't have to bring it up to them. They
brought it up to me. I mean, it's so much
on their minds. And so that's why I put this
quotation about the Manhattan Project at the start of the book,
because it's kind of the It sums up one of
the main threads, which is this, you know, scientists can't
resist inventing something which is exciting technically, and then they're
going to be the most hated and most wanted people
(02:01):
in the country.
Speaker 1 (02:02):
You described the question of motivation when it comes to
demis hanging in the air like the mushroom cloud over
Los Alamos, very arresting visual image. Why that image? And
what did you understand in the end about his motivation?
Speaker 2 (02:19):
I think when you look at that picture of the
mushroom cloud over Los Alamos, you're kind of thinking both wow,
but also why why did human beings do this? You know,
it's so destructive? Why did you do it? And I
guess you know, part of the threatened my book is
he had a series of ideas about how he could
(02:41):
build a eye and make it safe for humanity and
beneficial for humanity. And one by one these ideas become
unraveled as they collide with reality. So the story of
demisisabass in some ways, you know a story. I think
there's two categories have screw up in the world. Right,
Sometimes you get something where basically idiots are in charge,
(03:02):
they don't understand what they're doing, and they make a
humongous mistake Iran War for example. Right then you have
another much more interesting category of screw up, and that
is where intelligent people know from the beginning exactly what
they're doing. They can see the risks, they think they
can manage them, but then forces which are larger than them.
In this case with AI, it's a race dynamic between
(03:24):
multiple labs and multiple countries take over and they can
no longer control the technology that they've invented. And I
think those episodes where you couldn't just switch out the
individuals and have a better outcome, where the individual is good, sincerely, good, intelligent, thoughtful,
has foresight, and yet you still end up in a
(03:45):
bad place. That's what's really fascinating. When Demis sold his
company deep Mind to Google in twenty fourteen. There was
a condition which was this will never be used for weapons. Well,
you know now it's twenty twenty six it is being
used for weapons. And you know, so in time after
time he tried to draw lines in the sand, and
(04:05):
they've all been erased.
Speaker 1 (04:06):
It's harder and harder to say what these AI companies
are today. I mean, for a moment last year, Open
ai was the most popular social video app company in
the world, and now it doesn't do sour anymore. But like,
what is your working definition of what deep Mind is.
Speaker 2 (04:23):
I mean, it's a laboratory for the invention of machine intelligence.
And machine intelligence is a very compacious thing. You know,
it goes from text to video to images to a
system like Alpha fold, which divined all the shapes of
proteins in nature, and one demis the Nobel Price. So
it's a huge field, and indeed the creation of it
(04:46):
is a huge thing because you bring in experts in neuroscience,
experts in chemistry, experts in physics, exprests, computer science, ethical
experts who can philosophize about the personality that an AI
system should have. I mean, it's a very multi discip
binary thing, which is part of what makes the story fascinating.
Speaker 1 (05:02):
When did you first meet Demis Now?
Speaker 2 (05:04):
I first met him when I'd being interested in, you know,
technology generally. My last book was about Silicon Valley and
venture capital, and in the process of writing that book,
I would go to tech conferences in Europe and there
would be this sort of diminutive figure with a big,
big smile and sort of a kind of boyish charm,
really unassuming kind of guy with you know, a sort
(05:27):
of round neck sweater and his hair falling forward in
a fringe, and he would get up on the stage
with a big grin and sort of just almost as
if he was talking about how he was about to
wash the dishes after lunch. You know, he would in
a very plain spoken way, talk about well, when I
was a child, I had two ambitions. One must understand
(05:47):
all of science, and the other was to understand all
of philosophy. So I resolved this dilemma by deciding to
build AI, which would help me to understand both. And
so he had this mind blowing mental reach combined with
this totally approachable next door friend kind of attitude.
Speaker 1 (06:04):
And when did you have the idea to pitch him
on being a biographical subject for you.
Speaker 2 (06:10):
So after finishing my last book, The Power Law about
Venture Capital, I was thinking, and you know, this is
now mid twenty twenty two, what would be a good
next subject. And because I had met Demis several times
and I kind of followed what Deep Mind was up to.
I knew about the protein folding system, I knew about
i'lpha go, the go playing system before that, and so forth,
and I had a sense that it would probably go
(06:32):
from the fringe to the mainstream at some point in
the next year or so. And then it took me
a few months after that conversation inside my head to
get my act together, listen to every podcast that Demis
had ever done, read all his lectures, really think my
way into his brain, and then go and see him
to pitch him on giving me a ton of time,
because I need a lot of time with people if
(06:53):
I'm going to write a book with him. And I said, look,
you know, I want to write this book about you.
And it seems to me Demists that you may not
want a book about you, but you've said repeatedly in
all of your lectures that AI will be the most
important invention in all of human history demos. So that
means if you're the creator of this AI, you must
(07:15):
be one of the most important people in human history.
And if that's the case, you don't have a choice.
Somebody is going to write a book, right And furthermore,
you should welcome this because if you're going to invent
a technology that is going to disrupt people's lives so thoroughly,
you know your job will be different, how you raise
your children will be different, how you think yourself as
(07:35):
a human will be different because you now have this
different source of intelligence competing with you. You can't disrupt
people from head to toe and then not explain them
why you did it. You need to explain your motives, right,
And that's the project I'm proposing to you. And he
thought about it, and he's seen well disposed to this.
And then one week later chatchipt came out. Oh my goodness,
(07:59):
and expectation of the technology going from the fringe to
the mainstream happened a whole lot quicker than I expected.
Speaker 1 (08:05):
And I mean obviously a lot of the book of
your interests in the technology, how it might change the world,
the kind of financing and deal making shenanigans that made
Deep Mind in many ways what it is today. But
what at the personal side of an extraordinary biographical portrait
too very specific parents a prodigious talent for chess, which
(08:28):
he then gave up because he thought it wasn't in
some sense consequential enough. I mean, did you know when
you got it, you knew that Ai was the next thing,
and you know there he was. But did you know
what an extraordinary personal story he had before you really
got into it with him?
Speaker 2 (08:41):
No? I didn't. And in fact, I remember very clearly
to two early experiences in the first discussions. You know.
One was I was going to have this dinner I
told you about, and he told me to read a
book before I came to the dinner, and the book
was enders a Game. Now, this is a science fiction
story about a sort of diminutive boy genius hero who
(09:04):
has to save planet Earth from invading space aliens. And
he's at the end of the book he saves all
of humanity from space aliens. And Demis said to me, well,
I wanted you to read this book because I really
identify with that character Ender, And I'm thinking, wait, so
you're telling me you're the savior of humanity. I mean,
even if you think that Demis, maybe you shouldn't be
(09:24):
announcing it to the person who's about to write a
book about you. I mean, surely that's too messianic, to
over the top, too ridiculous. But he's right out there
with it. You mean, that is how he thinks, and
he's not ashamed to tell you. And so that was
pretty extraordinary. And then the second thing was I went
to see Shane Legg, his Scientific co founder, and he
told me the story about how I said, you know,
(09:44):
what was it like to work with Demis And he said, well,
you know, Demis has crazy determination. I said, well, what
do you mean. He said, well, you know, one day,
according to Demis, his dad said to him, during the
chess period of his life, listen, you're gonna go play
chest today. You know, you just have to try your best. Now,
when I say that to my son, I mean, you know,
(10:05):
it's fine to lose so long as you try your best.
The way that Demis apparently interpreted it, according to Shane,
was you have to try your absolute, absolute, absolute best.
And it's like running a race and at the end
of the marathon you fall over the tape and you're
on the ground and you have to be taken to
hospital because you're almost dead. And if you haven't been
taken to hospital, it means you didn't try hard enough.
(10:28):
That is what try your best meant to Demis aged
about ten or twelve, and I went to see Demis
the next time and I replayed this back to him
and said, is that really true? Is that how you
interpreted it? He said, oh, yeah, absolutely. You know you
have to give it every single drop all the time.
Speaker 1 (10:44):
And there's an amazing moment where Demis describes talking about
hearing nature or science screaming at him and him struggling
to hear and to understand.
Speaker 2 (10:54):
Yeah, that was the most extreme expression of his desire
to invent AI. So one day I was with him
in you know, Hampstead Heath, which is a park in
North London, not in the pub. Not in the pub
this time it was a nice day, so we went
for this cafe instead, and you know, there he was.
There was kind of a classic English scene. There was
somebody in front of me who was on his cell
phone you're doing some sort of sales job, and two
(11:16):
women behind me talking about their friend who had a
medical incident and had to get to hospital. So all
these quididian noises in the background, and there is demisis
Sabus looking at me, talking about the creation of this
godlike machine and saying that when he's up at two
in the morning at his desk at home thinking about this,
he can sort of feel reality summoning him, screaming at him,
(11:37):
understand me, understand me, and you know he would then
slam the table and say, look, Sebastian, this table, it's
made of atoms, buzzing around with electrons. Why should it
be solid? Why should that laptop you've got there, why
should it you know, pieces of sand and metal? How
could that turn into something which can think? I mean,
what's going on here? There must be some intelligent thoughts
designing all these things. And so he kind of basically
(12:01):
told me that inventing AI and understanding the universe is
like getting closer to what he thinks of as God.
Speaker 1 (12:08):
So he's a religious man.
Speaker 2 (12:09):
I don't know if he would agree with religious because
he doesn't go to organized religious services, but he's spiritual.
Speaker 1 (12:14):
I would say, interesting, And I mean that scene you
described could be a scene from Oppenheimer, right, I mean
it's so cinematic. Did you ever think is he doing
this for me? Or is he crazy? Or is it
just absolutely captivating? The energy and the sense of purpose
that he brings to this.
Speaker 2 (12:31):
He just exudes both energy and intelligence, but also storytelling,
natural talent. It's just amazing. I mean, you know, one
time I asked him about his first office in London
and Russell Square, which is a sort of storied square,
you know, near the British Museum and so forth. And
you know, normally, as a writer, you ask somebody to
recapture the emotion of opening their first office fifteen years ago.
(12:53):
It's fifteen years ago. They're going to say, oh, yeah,
it was cool. You know, that's all you'll get out
of them. But demis just flows with stories. He said, wow,
you know, I was in the attic that's where the
office was, and of course you had to come down
the stairs. They were all Rickety's so I came down
ding ding ding ding ding bang bang bang, And then
I come out on the square and there's these beautiful
trees in front of me. Beyond to the right. If
you just go three doors down Sebastian, that's where you
see the London mathematical society, where Turing invented the origins
(13:18):
of computer science, which we are now completing. And then
if you go beyond that to the level crossing black white,
black white crossing the street, the pedestrian crossing, that is
where the Hungarian nuclear scientist Zilad had the idea for
a nuclear chain reaction back in the nineteen thirties, which
led to the atom bomb. And of course we are
now creating the equivalent of the atom bomb with Ai.
(13:40):
What a subject, Yeah, I mean, he is such a storyteller.
Speaker 1 (13:44):
And I heard that he has a sense of humor
or perhaps a sense of humor about himself in some
ways as well. Didn't he say when he lost to
the table football competition in the office that his soul
was on fire?
Speaker 2 (13:54):
He did say that, yes, you know, one can mock
him for being too competitive and taking trips things like
table football seriously, but he actually really feels it.
Speaker 1 (14:03):
And we talked about the sort of mushroom cloud of motivation.
One of the things that doesn't seem to be so
motivating to him is money. I mean, there's a story
about the offer as a as an eighteen year old
of a half million pounds to join game development Studio right,
which he turned out right despite coming from me. I
know his mother had gone through homelessness her in her youth.
(14:24):
I mean, was that a hard decision for him? Why
did he make it?
Speaker 2 (14:27):
He said, it was completely easy in today's money. It
was well over a million dollars that he was being offered.
He was eighteen. As you say, his parents were not rich.
I mean, you know, any self respecting sort of Stanford character.
But at this point of fall, you know, taking the
money dropped out of Stanford and you know, written off
into the sunset with a loot. No, Demis is different.
Demis wanted to understand science. That was his primary motivation.
(14:51):
That's what he's up when he's up at two o'clock
in the morning. He's thinking, how do I understand nature?
And so he turned down the cash to go and
study computer science instead.
Speaker 1 (15:00):
Fast forward a few years and he meets Peter Teel,
who gives him a A plus for science fiction and
an F for business model. But nonetheless the size skives
some money.
Speaker 2 (15:13):
Well, actually there's a fraud and slip there. You said
an A plus for science fiction. I think you met
an A plus for science Oh, put.
Speaker 3 (15:19):
Science fiction, but science fiction maybe that would have been better,
because in fact Demis was spinning this vision and this
is twenty ten, right, He was saying, I'm going to invent.
Speaker 2 (15:30):
Very powerful AI. This is at a time when AI
literally couldn't recognize the photograph of a cat, like nothing
was working, and you have this character coming and say, oh,
I'm going to create artificial general intelligence. It was nuts.
So it kind of was science fiction.
Speaker 1 (15:45):
What was his entree to the world of technology investors
and when did deep Mind actually start as a company.
Speaker 2 (15:51):
Deep Mind started in twenty ten, having raised the money
from Peter tele The entree is very interesting because in
fact what happened was, you know, Demis had done a
small games company before, and he made some money. Wasn't
a terrific success, but nonetheless it wasn't total failure. And
he went back to the same investors. They all said,
you must be joking. There's no product if you're doing
(16:11):
AI and not putting money into that. So then he
had to think again, and his entree into the world
of Peter Teel and what's called the singularity summits, where
all these very early believers in AI would gather people
like Ray Kurtzweil, and they would dream about a future
of an AI that totally did not exist, and when
they got up on the stage, actually they did often
(16:33):
draw more on science fiction novels than on science when
they were kind of imagining a future with AI. And so,
in this strange cauldron of mythology and reality with all
kinds of weirdos trotting about, Demisa Sabots, who by this
point has a computer science degree and a PhD in
neuroscience as a proper scientist, shows up and he's asked
(16:56):
by a journalist what do you think of the Singularity conference?
Are you a Singularitarian? And he says it's a bit
Californian for me, and he's, oh, you could sort of
feel the kind of anxiety of being seen in this crowd.
But that's why you had to go to meet Peter Teele.
And then when he met Peter Tiel, he had this
clever trick. Peter Teel is a chess player. Demis is
(17:16):
a chess player. So rather than pitch Peter tile on
some idea about a company, and he said, well, I
think the interesting thing about chess is that the knight
and the Bishop are supremely well balanced. And it's in
that tension between those two pieces that much of the
joy of the game resides. So Peter T' is like, WHOA,
that's a conversation I want to pursue. And so that
(17:37):
got him and got Demssus Habits an invitation to Peter
Tele's house the next day, and then that's when he
pitched him on Deep Mind and got the money he
needed to start the company.
Speaker 1 (17:46):
And then first forty twenty thirteen and which the excerpt
in the Wall Street Journal of your book tells the
story of a birthday party for Elon Musk, replete with
all kinds of costumes and strange things and fake battlements.
But this is, perhaps, apart from the founding, the most
(18:06):
crucial moment in Deep Minds Genesis as a company.
Speaker 2 (18:09):
Right, yeah, that's right. So you know, by this point
Demis had raised three rounds of venture capital, including from
Elon Musk, and you know, there are various people could
come in, but it was a total pain in the neck.
He hated it. You know. He would sometimes have this expression,
I don't want this part of my brain to expand
he wanted to be doing science, and so what he
(18:31):
wanted was to be liberated from this hamster wheel of fundraising.
And along at this party, this birthday party that Elon
Musk had, along comes Larry Page from Google who's also there,
and says, let's go for a walk, and they walk
around the castle grounds, and in this bizarre setting, Larry
Page says to him, well, you know, you could spend
(18:52):
your career building another company like Google. That's fine, but
if you really want to do science, just join Google
and we'll give the resources, use our platform, and you'll
be able to do what you really love. And Demis
not only agreed with that pitch in the sense that yes,
he preferred to do science then to be a billionaire,
(19:13):
but he felt that Larry Page himself would have accepted
that pitch. That Larry Page cared about science, he could
have been a standard professor of computer science. So Demis
really identified with Larry Page, and that was why he
sold the Google.
Speaker 1 (19:26):
And Page had his eye on Demis or this was impulsive.
Had he planned out his chess game for this party
like Demis had three years before, he had.
Speaker 2 (19:34):
Totally planned the chess game. He'd been thinking for a
while about buying up nascent Ai companies, and he'd bought
the boutique founded by the Toronto professor Jeffrey Hinton together
with Ilias Askeva and one other person, and so he
was in a buying mode.
Speaker 1 (19:52):
That was twenty twelve, right the image net team exactly.
Speaker 2 (19:55):
He bought the image net team, and then the next
obvious person to buy was was and DeepMind, because they
had a different approach to AI. It wasn't just deep learning,
which is the ImageNet secret source, which is kind of
packet pattern recognition learning from data. It was also what's
called reinforcement learning, which is learning through trial and error
(20:17):
in a simulation. So you have a game like the
Atari games or go later on, and you try lots
of different the computer, try thoughts and move c suites,
run works and then learns through trial and error. And
in some ways, another strand in my book is the
interplay between deep learning on the one hand and reinforcement
learning on the other hand. And these two fields of
(20:39):
artificial intelligence, you know, have their different moments in the sun.
As the story progresses.
Speaker 1 (20:45):
Hinton talked He came on tex Stuff and talked about
how he ran an auction to sell image Net with Google, Microsoft,
and by Doo. But in the end he all he
really wanted was to go to Google for for Demis,
he was being courted as well by others, including a
dinner at Mark Zuckerberg's house in the I guess weeks
or months after this first meeting with Larry Page at
(21:06):
Elon Maas's birthday party, and he submitted Mark Zuckerberg to
a test at this dinner.
Speaker 2 (21:14):
Right, Yeah, that's right. So the test was a bit subtle. Predictably,
they sit down to dinner and Mark Zuckerberg, who's longing
to buy deep Mind to get one over Google.
Speaker 1 (21:25):
This was not recently, this was ten years ago.
Speaker 2 (21:27):
It is twenty thirteen. So Mark Zuckerberg says, well, I
think AI is the most important technology in human history.
It's extraordinary, and you know, I really hope you agree
to join me at Facebook because you know, we could
just do great things together. Blah blah blah blah. And
then you know, the conversation moves on, time goes by,
(21:48):
and then Demis slightly says, you know, three D printing
is extraordinary, and Zuckerberg goes, yeah, I agree, you know, incredible,
that's just going to unlock so many things. And then
a bit later, Demis says, you know, official reality, that
really is going to be transformative. As Zuckerberg's like, yeah,
it's transformative. It's so exciting. I'm so excited by that.
(22:08):
And then Demis's mind is wearing. Is said, Okay, he's
a bullshit artist. He does not believe that AI is
the most important thing. Ever, he does not get it.
I'm selling to Google, Forget, forget Facebook.
Speaker 1 (22:19):
Even though the more money on the table.
Speaker 2 (22:21):
Yeah, that's right. In fact, Facebook was offering to make
Demis a lot richer, but he was consistent throughout his career.
Demis in turning down the money not to go to
Cambridge University, turning down the money to sell to Facebook.
It's not about the money for him, it's really about
the science.
Speaker 1 (22:34):
And you mentioned him using his scientific method to see
two three years into the future. Instead, Facebook went with
Jan Lucun and gave him plenty of resources, and I
think he was trying to poach some of Demos's employees.
Demists told them that the Google deal was going to
happen and to sit and therefore to sit tight, and
they did. But you know, fast forward to twenty twenty
six and Jan Niklean hasentually been dumped from Meta, and
(22:58):
Demis is where he's sitting Google is he is? He
is he the successor to Sundar. Is he the you know,
the ego and the ID. I mean, what is his
role in Google today?
Speaker 2 (23:08):
Well, what his rold is today is to be the
chief executive of Google Deep Mind, which is the AI engine,
which is basically powering all the new products in Google.
So he's super important. Sundar is the chief executive of
Alburt and Google, and I would argue that the relationship
between Sundar and Demis is the most important relationship in
business anywhere at the moment. Because Sundar has Demiss back.
(23:32):
Sundar gives him the resources. Sundar takes care of the
kind of all that kind of corporate leadership staff that
Demis is good at, but it's really not what he
wants to do full time, and that gives Demis the
oxygen to pursue AI to the fullest of his abilities,
which are considerable you know, in the future, if Sundar
were to go, I don't think that's happening anytime soon,
(23:53):
by the way, but I think if he were to go,
you know, Demis would obviously be talked about as a candidate.
And it's a really interesting question because he is on
the one hand, somebody who is a leader, has vision,
can motivate people, would have the credibility to lead Google
as an AI company. I mean, how often do you
get somebody who's the CEO and also has a Nobel
(24:14):
Price That would be quite something. But at the same time,
Demis has a side him that wants to be a
pure scientist, that talks to me about you know, there's
too much noise in Silicon Valley. I want to go
and think I want to have a research professorship at Princeton.
That's where Oppenheimer went after the Manhattan Project. That's where
Einstein went, That's where I should be. You know. He
has that kind of you know, retreat to the idyll
(24:37):
of abstract contemplation side to him, and he's so good
at both of these things. It's what makes him exceptional.
I mean, if you mentioned Jan Lukun, you know, very
good scientist, but clearly not a great operator inside business.
You know, one could talk about Simultman, a great business operator,
but not a scientist. Dropped out of Steinford, doesn't have
a degree, you know, it's very rare to find both
(24:59):
in the same person.
Speaker 1 (25:09):
After the break is Demis an evil genius stay with us.
You mentioned earlier in the conversation this kind of journey
(25:33):
Demis had been on where one sort of safety mechanism
after another that he believed in fell away, and thus
this kind of metaphors about the atomic bomb. But ironically,
in some sense, the kind of safety to the wayside
(25:54):
race that we're in today with AI was kicked off
by Demis his desire for a safety board.
Speaker 2 (26:03):
Yes, that's a good irony, you're right. So what happened
was that Demis sold the company to Google in twenty fourteen,
and one of the conditions was there had to be
a safety oversight board and whereby Google would allow Deep
Mind to sort of appoint some you know, important philosophers
or other people of independent stature to make a final
(26:23):
decision on when AI would be deployed into the world.
And the idea was this is AI is too big
just to let the corporate board of Google do whatever
it wants with it. You know, there has to be
a check. So the first of these safety meetings was
arranged and Demis had the idea, we'll invite Elon Musk
to shair it, and he invited Rieed Hoffmann and various
(26:45):
other people and they all met at SpaceX and basically
what happens is Elon Musk sat there listening, absorbed all
the presentations from Deep Mind about their plans to build AI,
and a few months later he announces Open Ai, which
is going to be the rival company. And so all
of a sudden, this singleton vision, the idea that you know,
(27:07):
only one AI lab would shepherd AI into the world
on behalf of all humanity that just is by the wayside,
and you've now got two competing labs, and the race
dynamic begins to set in.
Speaker 1 (27:20):
How did demos feel about what Elon did?
Speaker 2 (27:23):
Betrayed? Elon had sat there listening to all his plans,
and he'd been invited to chair that meeting in good
faith to ensure safety for the world, which of course
is what at the time Elon was a big duma
and was constantly talking about AI safety and existential risk.
And so the idea that rather than uniting with deep
(27:44):
Mind and Google in a single effort to make the
technology safe, Elon must prefer to go off and start
a rival in Open AI to Demis. This was a
total betrayal. Of course, Elon thought of this as Demis
is dangerous, he's an evil genius, and therefore I need
to be the one because you know, all of these actors,
(28:07):
they basically say, I know that I'm a good person. Yeah,
if I'm the leader of the AI race, I will
make it safe because I'm good. But those other guys
over there, you can't trust those guys because you know whatever. Now,
if you quizzed Elon Musk about why did he say
that Demis was an Elon was it was an evil genius?
Speaker 1 (28:25):
Your term for a Freudian slip.
Speaker 2 (28:29):
Elon evil genius? Why was Demis an evil genius? Well,
the only good reason, or not a good reason, but
the reason was apparently Demis, in his game design days,
had worked on a game called Evil Genius. No, which
is a pretty thin basis on which to call him
an evil genius. But whatever I mean, they all had.
Speaker 1 (28:48):
Association Sean Elbows. So then that this is twenty this
is meeting is in twenty sixteen, twenty fifteen, twenty fifteen,
and when is the alpha go moment?
Speaker 2 (28:59):
Six? Okay, so coming out of that moment, when Elon
Musk decides to set up open. Ai Demis decides, well,
I'm just going to accelerate as fast as possible, and
the first thing he manages to score is this victory
over the Korean Go champion Lisa Dol And it's a
(29:19):
huge exhibition match in South Korea with all the media
in attendance, and it's kind of an it's not quite
chatchypt but it's a moment when Ai had what one
might call the Kasparov Deep Blue moment in nineteen ninety seven,
first time the human champion gets defeated, and then twenty sixteen,
(29:40):
so that's nineteen years later, the same thing happens with
Go and.
Speaker 1 (29:43):
Two hundred million people tune in and the defeated Korean
player apologizes to humanity. It's a huge moment, but it's
nothing like the Chatchipt moment six years later.
Speaker 2 (29:54):
Yeah, because Go people watched. Whereas chat ChiPT you used it.
It was personal, it was visceral.
Speaker 1 (30:01):
And within a week of you pitching damage on the book,
Chatchipt came out.
Speaker 2 (30:05):
That's right. And I went to see him right after
that and he said, you know, this is war. Those
guys have parked their tanks in our front yard actually
said on our lawn, but translating for American audience, in
our front yard. And so you could see that competitive
glint in his eye, and you knew he was going
(30:26):
to try and fight back.
Speaker 1 (30:27):
Was he self aware about the risk of using that
language even for himself given all these Manhattan Project analogies,
you know.
Speaker 2 (30:36):
He's a person with many different dimensions, and he's both
capable of worrying about safety and also using military metaphors
to express this determination to crush the opposition. And I
think actually it's going to be a business school case
study of how deep Mind made the comeback because they
emerged deep Mind the London lab with Google Brain the
(30:58):
Mountain View Google AI Lab. Normally, mergers are super difficult,
they don't work. And here was a merger you had
to do in the middle of an AI race which
had been kicked off by chatchept. You had eight time
zones between California and London. You had a record of
bitter rivalry between the AI scientists from Google and the
ones from deep Mind. And yet they pulled it off.
(31:20):
They did the merger, they blended the cultures, and within
two and a half years they had a model that
was outclassing open AI models.
Speaker 1 (31:27):
See that's just extraordinary. So I remember when the chetchipt
moment happened, and I would say, up until twenty twenty
the beginning of twenty twenty five, people were saying Google
is down and out, Google might be over. I mean,
you knew because you were reporting along the way that
probably wasn't true. But what the indications that you saw
that the rest of the world didn't that convinced you
(31:49):
along the way that demos and deep Mind might be
roaring back into Do you put them in first place? Now?
Speaker 2 (31:55):
I think it's sort of a pretty close race between
the Gemini models from Demis and then Claude is doing
really well at the moment the anthropic model. People love
it for coding and so forth. So you know, I'm
not sure that it's I think the race is still ongoing.
What I would say, though, is that you know, I'm
on record as having written in the New York Times.
Speaker 1 (32:15):
That around of money, right, probably run.
Speaker 2 (32:18):
Out of money. I mean, they may put it out,
but basically, in fact, since I wrote that piece, they
do seem to have focused their business quite a bit
by giving up on Soora for example, Saura was a
classic money losing idea. You know, it costs enormous amounts
to generate video, but people don't pay you to generate videos,
so quite rightly, they can do it. So maybe they
(32:40):
can cut costs enough to survive. But they have a
huge cash need and they do not have Google's deep
pockets behind them, unlike Demis.
Speaker 1 (32:50):
So Demis is kind of winning. But he said to you,
it doesn't necessarily feel like that, right, he said, this
is a paradoxical moment. It should feel amazing, but it
doesn't feel how I thought it would feel.
Speaker 2 (33:00):
Yeah, because early on he had this rather naive idea
that there would be one lab building AI and so
you could take your time about releasing the models, and
you know, if you were worried about safety, you could
just take another six months to test them. And now
you have this race, and you know, the Chinese have
plenty of models, and the other thing, it's not just
a race, it's actually also the open source nature of
(33:20):
these models, where they're being released out into the wild,
and some weird group can just download the model, have
it on their own computer, and then you can't pull
it back anymore. And so there was a big cyber
attack in Mexico recently where all of the electoral records
were stolen, and Anthropic realized that its clawed model was
(33:42):
being used. But because that model is proprietary, they could
immediately shut off access and stop the attack. You couldn't
do that with an open weight, open source model. And
yet we have open weight, you know, that's being put
out there, both by Meta and by the Chinese and
by Mistride in France. A lot of open source models
are out there, and so in many ways the way
(34:05):
AI is being deployed is frightening. The obvious safety measures
one might take are not happening. In addition to banning
open source, I think there should be much more powerful
sort of government oversight so that, just like with a pharmaceutical,
before you release it to be used in people, as
to go through clinical trials, so too, I think there
(34:25):
should be a sort of equivalent of the Food and
Drug Administration an AI agency that can actually veto the
release of really powerful models. And we don't have that,
and we should have that, and we should be negotiating
with China about doing it in both places at once,
because this is a global race and both sides have
to slow down. I was in China recently for eight
days because they always published books faster. So I was
(34:49):
meeting AI leaders, both from industry and from academia, and
I was surprised by how much they do talk about safety.
So I think there is a discussion to be had
with it Chinese about safety, but the US administration of
this moment doesn't want to do that.
Speaker 1 (35:06):
I mean, coming back to the Manhattan Project again, Demis
has said, I think that he thinks this may end
in a bunker and what does he mean by that?
And he has he primed himself psychologically for an Ai
Hiroshima that he may feel in some sense responsible for.
Speaker 2 (35:25):
Yeah, I mean when I was doing the research, interfering
not just Demis, but all the scientists that he works with,
you know, one hundred or something of them. In deep mind,
I would hear this, these references to the bunker come up,
and I assumed it wasn't literally, you know, a real
thing that Demis wanted to disappear into a bunker at
the moment when he thought the AI models were coming
dangerously powerful. And I would have these dinners every six
(35:47):
months with a friend who had been at deep mind,
but had left and I tested this on him one
evening and I said, yeah, surely this is just a
metaphor bunker. He can't be serious. This guy said, well, actually,
you know, I had my bag packed. It was serious.
There was actually this vision that AI would become so
(36:09):
powerful that bad guys would try and get it off you.
So you had to hide in some place a bit
like Los Animals and develop in sort of isolation and secret,
and also be isolated because you needed maximum focus on
the science to get it right when you were at
this moment of maximum danger because the model was suddenly
very powerful, and that was his vision. Now I think
(36:30):
today he doesn't believe that anymore, because we're so far
from a single lab, you know, midwifing AI. So I
think now he's more inclined to speak of some version
of the Center for European Nuclear Research SAN, which is
a sort of technical agency that oversees nuclear power on
a multinational basis. I think you would like some sort
(36:53):
of global body to impose rules on what kind of
AI should be let out into the wild. But you know,
at the same time he knows that politically that's not
on the cards, and he has a sense of timing
about when you should raise these issues, and so you know,
whereas Dariama Day took on the Pentagon by trying to
(37:17):
assert safety principles and then just got rolled, I think Demis,
when he does that, is going to feel that he's
got the door is half open and he can give
it a push and we'll see. You know. Of course,
sometimes people keep their capital drive for so long that
they never use it, but we'll see. If the moment
comes when he does use it will be very interesting.
Speaker 1 (37:37):
Especially just to close, there was a great review of
your book in the Financial Times which ends with this,
Whether and how Demis ever achieves AGI will form the
defining chapters of his extraordinary and unfinished biography. What did
you think about that? And what is the next chapter
for him? And will you write another follow up book?
(37:57):
Do you think?
Speaker 2 (37:57):
You know? I tend not to write follow ups about
the same thing the same person. I prefer to plow
fresh ground. But look, I mean, you know, Demis t
haming fifty this year. He's got a lot of runway.
I'm sure he'll do more incredible things in the future.
So probably I am offering an interim report. But the advantage,
you know, if you wait. I did this before with
Alan Greenspann. I wrote the definitive biography after he retired,
(38:20):
and by that time, you know, people are interested, but
less so than when he's stood in the seat. I
think capturing a portrait of you know, the most interesting
figure in artificial intelligence in real time while he's still
in the steat and he's still doing it is sometimes
the fun of it, right, I mean, who wants to
wait for the definitive biography in twenty years time?
Speaker 1 (38:40):
Well, next for him, you know, I think he's.
Speaker 2 (38:41):
Going to carry on running Google Deep Mind. There's going
to be more agentic models coming out this year. There
will be you know, world models and more robotics coming
There will probably be much more AI for science, both
in terms of drug discovery and in terms of you know,
material sciences, chemistry and so forth. So I think, you know,
(39:02):
one day, I remember, towards the end of my time
interviewing him, he showed up at the pub and he
had a backpack and he pished something out of it,
and he got this little box out and he said,
I got to show you this, and he opened the
box up and inside was the Nobel prize medal and
either at that meeting or another one. He said to me,
(39:24):
I wonder if I can get another one. He's not
over yet.
Speaker 1 (39:28):
It's a vasty malabi. Thank you.
Speaker 2 (39:29):
It's been great fun to talk.
Speaker 1 (39:46):
For tech Stuff. I'm as Volosciin. This episode was produced
by Eliza Dennis and Melissa Slaughter. It was executive produced
by me Julian Nutter and Kate Osborne for Kaleidoscope and
Katrina Novel for Ihart Podcasts. The engineer is p Bowman
and Jack Insley makes this episode. Kyle Murdoch wrote our
theme song. Please rate, review, and reach out to us
(40:07):
at tech Stuff podcast at gmail dot com. We also
love to hear what you think our panels should cover
next time
Welcome to tech Stuff. I'm Os Voloscian, and today we
get the opportunity to go behind the curtain at Google's
Deep Mind. For almost three years, in the upstairs room
of a pub in North London, journalist Sebastian Malabi met
regularly with the company's CEO and co founder Demis has ABIs.
They spoke about artificial intelligence, philosophy, neuroscience, motivation and consequence,
(00:40):
all against the backdrop of an increasingly intense three way
race between Open Ai, Anthropic and Google to win the
race towards Agi. Sebasti, congratulations on your new book, The
Infinity Machine, and welcome to tex Stuff.
Speaker 2 (00:55):
Thank you ask going to be here.
Speaker 1 (00:56):
You begin the book with a quote from other scientists
who watch on Manhattan Project, who said, what we're creating
now is a monster whose influence is going to change history.
Yet it would be impossible not to see it through
the energy source which is now being made available, or
make scientists the most hated and the most wanted citizens
of any country where.
Speaker 2 (01:18):
Reliving that now with Ai argree, I mean I began
this project wanting to capture the tingling sensation of human beings,
like Demsis ABIs, creating the new version of atomic weapons. Right,
this incredibly powerful AI technology that has enormous upsides but
could also be very, very dangerous. And the surprise was
(01:38):
I didn't have to bring it up to them. They
brought it up to me. I mean, it's so much
on their minds. And so that's why I put this
quotation about the Manhattan Project at the start of the book,
because it's kind of the It sums up one of
the main threads, which is this, you know, scientists can't
resist inventing something which is exciting technically, and then they're
going to be the most hated and most wanted people
(02:01):
in the country.
Speaker 1 (02:02):
You described the question of motivation when it comes to
demis hanging in the air like the mushroom cloud over
Los Alamos, very arresting visual image. Why that image? And
what did you understand in the end about his motivation?
Speaker 2 (02:19):
I think when you look at that picture of the
mushroom cloud over Los Alamos, you're kind of thinking both wow,
but also why why did human beings do this? You know,
it's so destructive? Why did you do it? And I
guess you know, part of the threatened my book is
he had a series of ideas about how he could
(02:41):
build a eye and make it safe for humanity and
beneficial for humanity. And one by one these ideas become
unraveled as they collide with reality. So the story of
demisisabass in some ways, you know a story. I think
there's two categories have screw up in the world. Right,
Sometimes you get something where basically idiots are in charge,
(03:02):
they don't understand what they're doing, and they make a
humongous mistake Iran War for example. Right then you have
another much more interesting category of screw up, and that
is where intelligent people know from the beginning exactly what
they're doing. They can see the risks, they think they
can manage them, but then forces which are larger than them.
In this case with AI, it's a race dynamic between
(03:24):
multiple labs and multiple countries take over and they can
no longer control the technology that they've invented. And I
think those episodes where you couldn't just switch out the
individuals and have a better outcome, where the individual is good, sincerely, good, intelligent, thoughtful,
has foresight, and yet you still end up in a
(03:45):
bad place. That's what's really fascinating. When Demis sold his
company deep Mind to Google in twenty fourteen. There was
a condition which was this will never be used for weapons. Well,
you know now it's twenty twenty six it is being
used for weapons. And you know, so in time after
time he tried to draw lines in the sand, and
(04:05):
they've all been erased.
Speaker 1 (04:06):
It's harder and harder to say what these AI companies
are today. I mean, for a moment last year, Open
ai was the most popular social video app company in
the world, and now it doesn't do sour anymore. But like,
what is your working definition of what deep Mind is.
Speaker 2 (04:23):
I mean, it's a laboratory for the invention of machine intelligence.
And machine intelligence is a very compacious thing. You know,
it goes from text to video to images to a
system like Alpha fold, which divined all the shapes of
proteins in nature, and one demis the Nobel Price. So
it's a huge field, and indeed the creation of it
(04:46):
is a huge thing because you bring in experts in neuroscience,
experts in chemistry, experts in physics, exprests, computer science, ethical
experts who can philosophize about the personality that an AI
system should have. I mean, it's a very multi discip
binary thing, which is part of what makes the story fascinating.
Speaker 1 (05:02):
When did you first meet Demis Now?
Speaker 2 (05:04):
I first met him when I'd being interested in, you know,
technology generally. My last book was about Silicon Valley and
venture capital, and in the process of writing that book,
I would go to tech conferences in Europe and there
would be this sort of diminutive figure with a big,
big smile and sort of a kind of boyish charm,
really unassuming kind of guy with you know, a sort
(05:27):
of round neck sweater and his hair falling forward in
a fringe, and he would get up on the stage
with a big grin and sort of just almost as
if he was talking about how he was about to
wash the dishes after lunch. You know, he would in
a very plain spoken way, talk about well, when I
was a child, I had two ambitions. One must understand
(05:47):
all of science, and the other was to understand all
of philosophy. So I resolved this dilemma by deciding to
build AI, which would help me to understand both. And
so he had this mind blowing mental reach combined with
this totally approachable next door friend kind of attitude.
Speaker 1 (06:04):
And when did you have the idea to pitch him
on being a biographical subject for you.
Speaker 2 (06:10):
So after finishing my last book, The Power Law about
Venture Capital, I was thinking, and you know, this is
now mid twenty twenty two, what would be a good
next subject. And because I had met Demis several times
and I kind of followed what Deep Mind was up to.
I knew about the protein folding system, I knew about
i'lpha go, the go playing system before that, and so forth,
and I had a sense that it would probably go
(06:32):
from the fringe to the mainstream at some point in
the next year or so. And then it took me
a few months after that conversation inside my head to
get my act together, listen to every podcast that Demis
had ever done, read all his lectures, really think my
way into his brain, and then go and see him
to pitch him on giving me a ton of time,
because I need a lot of time with people if
(06:53):
I'm going to write a book with him. And I said, look,
you know, I want to write this book about you.
And it seems to me Demists that you may not
want a book about you, but you've said repeatedly in
all of your lectures that AI will be the most
important invention in all of human history demos. So that
means if you're the creator of this AI, you must
(07:15):
be one of the most important people in human history.
And if that's the case, you don't have a choice.
Somebody is going to write a book, right And furthermore,
you should welcome this because if you're going to invent
a technology that is going to disrupt people's lives so thoroughly,
you know your job will be different, how you raise
your children will be different, how you think yourself as
(07:35):
a human will be different because you now have this
different source of intelligence competing with you. You can't disrupt
people from head to toe and then not explain them
why you did it. You need to explain your motives, right,
And that's the project I'm proposing to you. And he
thought about it, and he's seen well disposed to this.
And then one week later chatchipt came out. Oh my goodness,
(07:59):
and expectation of the technology going from the fringe to
the mainstream happened a whole lot quicker than I expected.
Speaker 1 (08:05):
And I mean obviously a lot of the book of
your interests in the technology, how it might change the world,
the kind of financing and deal making shenanigans that made
Deep Mind in many ways what it is today. But
what at the personal side of an extraordinary biographical portrait
too very specific parents a prodigious talent for chess, which
(08:28):
he then gave up because he thought it wasn't in
some sense consequential enough. I mean, did you know when
you got it, you knew that Ai was the next thing,
and you know there he was. But did you know
what an extraordinary personal story he had before you really
got into it with him?
Speaker 2 (08:41):
No? I didn't. And in fact, I remember very clearly
to two early experiences in the first discussions. You know.
One was I was going to have this dinner I
told you about, and he told me to read a
book before I came to the dinner, and the book
was enders a Game. Now, this is a science fiction
story about a sort of diminutive boy genius hero who
(09:04):
has to save planet Earth from invading space aliens. And
he's at the end of the book he saves all
of humanity from space aliens. And Demis said to me, well,
I wanted you to read this book because I really
identify with that character Ender, And I'm thinking, wait, so
you're telling me you're the savior of humanity. I mean,
even if you think that Demis, maybe you shouldn't be
(09:24):
announcing it to the person who's about to write a
book about you. I mean, surely that's too messianic, to
over the top, too ridiculous. But he's right out there
with it. You mean, that is how he thinks, and
he's not ashamed to tell you. And so that was
pretty extraordinary. And then the second thing was I went
to see Shane Legg, his Scientific co founder, and he
told me the story about how I said, you know,
(09:44):
what was it like to work with Demis And he said, well,
you know, Demis has crazy determination. I said, well, what
do you mean. He said, well, you know, one day,
according to Demis, his dad said to him, during the
chess period of his life, listen, you're gonna go play
chest today. You know, you just have to try your best. Now,
when I say that to my son, I mean, you know,
(10:05):
it's fine to lose so long as you try your best.
The way that Demis apparently interpreted it, according to Shane,
was you have to try your absolute, absolute, absolute best.
And it's like running a race and at the end
of the marathon you fall over the tape and you're
on the ground and you have to be taken to
hospital because you're almost dead. And if you haven't been
taken to hospital, it means you didn't try hard enough.
(10:28):
That is what try your best meant to Demis aged
about ten or twelve, and I went to see Demis
the next time and I replayed this back to him
and said, is that really true? Is that how you
interpreted it? He said, oh, yeah, absolutely. You know you
have to give it every single drop all the time.
Speaker 1 (10:44):
And there's an amazing moment where Demis describes talking about
hearing nature or science screaming at him and him struggling
to hear and to understand.
Speaker 2 (10:54):
Yeah, that was the most extreme expression of his desire
to invent AI. So one day I was with him
in you know, Hampstead Heath, which is a park in
North London, not in the pub. Not in the pub
this time it was a nice day, so we went
for this cafe instead, and you know, there he was.
There was kind of a classic English scene. There was
somebody in front of me who was on his cell
phone you're doing some sort of sales job, and two
(11:16):
women behind me talking about their friend who had a
medical incident and had to get to hospital. So all
these quididian noises in the background, and there is demisis
Sabus looking at me, talking about the creation of this
godlike machine and saying that when he's up at two
in the morning at his desk at home thinking about this,
he can sort of feel reality summoning him, screaming at him,
(11:37):
understand me, understand me, and you know he would then
slam the table and say, look, Sebastian, this table, it's
made of atoms, buzzing around with electrons. Why should it
be solid? Why should that laptop you've got there, why
should it you know, pieces of sand and metal? How
could that turn into something which can think? I mean,
what's going on here? There must be some intelligent thoughts
designing all these things. And so he kind of basically
(12:01):
told me that inventing AI and understanding the universe is
like getting closer to what he thinks of as God.
Speaker 1 (12:08):
So he's a religious man.
Speaker 2 (12:09):
I don't know if he would agree with religious because
he doesn't go to organized religious services, but he's spiritual.
Speaker 1 (12:14):
I would say, interesting, And I mean that scene you
described could be a scene from Oppenheimer, right, I mean
it's so cinematic. Did you ever think is he doing
this for me? Or is he crazy? Or is it
just absolutely captivating? The energy and the sense of purpose
that he brings to this.
Speaker 2 (12:31):
He just exudes both energy and intelligence, but also storytelling,
natural talent. It's just amazing. I mean, you know, one
time I asked him about his first office in London
and Russell Square, which is a sort of storied square,
you know, near the British Museum and so forth. And
you know, normally, as a writer, you ask somebody to
recapture the emotion of opening their first office fifteen years ago.
(12:53):
It's fifteen years ago. They're going to say, oh, yeah,
it was cool. You know, that's all you'll get out
of them. But demis just flows with stories. He said, wow,
you know, I was in the attic that's where the
office was, and of course you had to come down
the stairs. They were all Rickety's so I came down
ding ding ding ding ding bang bang bang, And then
I come out on the square and there's these beautiful
trees in front of me. Beyond to the right. If
you just go three doors down Sebastian, that's where you
see the London mathematical society, where Turing invented the origins
(13:18):
of computer science, which we are now completing. And then
if you go beyond that to the level crossing black white,
black white crossing the street, the pedestrian crossing, that is
where the Hungarian nuclear scientist Zilad had the idea for
a nuclear chain reaction back in the nineteen thirties, which
led to the atom bomb. And of course we are
now creating the equivalent of the atom bomb with Ai.
(13:40):
What a subject, Yeah, I mean, he is such a storyteller.
Speaker 1 (13:44):
And I heard that he has a sense of humor
or perhaps a sense of humor about himself in some
ways as well. Didn't he say when he lost to
the table football competition in the office that his soul
was on fire?
Speaker 2 (13:54):
He did say that, yes, you know, one can mock
him for being too competitive and taking trips things like
table football seriously, but he actually really feels it.
Speaker 1 (14:03):
And we talked about the sort of mushroom cloud of motivation.
One of the things that doesn't seem to be so
motivating to him is money. I mean, there's a story
about the offer as a as an eighteen year old
of a half million pounds to join game development Studio right,
which he turned out right despite coming from me. I
know his mother had gone through homelessness her in her youth.
(14:24):
I mean, was that a hard decision for him? Why
did he make it?
Speaker 2 (14:27):
He said, it was completely easy in today's money. It
was well over a million dollars that he was being offered.
He was eighteen. As you say, his parents were not rich.
I mean, you know, any self respecting sort of Stanford character.
But at this point of fall, you know, taking the
money dropped out of Stanford and you know, written off
into the sunset with a loot. No, Demis is different.
Demis wanted to understand science. That was his primary motivation.
(14:51):
That's what he's up when he's up at two o'clock
in the morning. He's thinking, how do I understand nature?
And so he turned down the cash to go and
study computer science instead.
Speaker 1 (15:00):
Fast forward a few years and he meets Peter Teel,
who gives him a A plus for science fiction and
an F for business model. But nonetheless the size skives
some money.
Speaker 2 (15:13):
Well, actually there's a fraud and slip there. You said
an A plus for science fiction. I think you met
an A plus for science Oh, put.
Speaker 3 (15:19):
Science fiction, but science fiction maybe that would have been better,
because in fact Demis was spinning this vision and this
is twenty ten, right, He was saying, I'm going to invent.
Speaker 2 (15:30):
Very powerful AI. This is at a time when AI
literally couldn't recognize the photograph of a cat, like nothing
was working, and you have this character coming and say, oh,
I'm going to create artificial general intelligence. It was nuts.
So it kind of was science fiction.
Speaker 1 (15:45):
What was his entree to the world of technology investors
and when did deep Mind actually start as a company.
Speaker 2 (15:51):
Deep Mind started in twenty ten, having raised the money
from Peter tele The entree is very interesting because in
fact what happened was, you know, Demis had done a
small games company before, and he made some money. Wasn't
a terrific success, but nonetheless it wasn't total failure. And
he went back to the same investors. They all said,
you must be joking. There's no product if you're doing
(16:11):
AI and not putting money into that. So then he
had to think again, and his entree into the world
of Peter Teel and what's called the singularity summits, where
all these very early believers in AI would gather people
like Ray Kurtzweil, and they would dream about a future
of an AI that totally did not exist, and when
they got up on the stage, actually they did often
(16:33):
draw more on science fiction novels than on science when
they were kind of imagining a future with AI. And so,
in this strange cauldron of mythology and reality with all
kinds of weirdos trotting about, Demisa Sabots, who by this
point has a computer science degree and a PhD in
neuroscience as a proper scientist, shows up and he's asked
(16:56):
by a journalist what do you think of the Singularity conference?
Are you a Singularitarian? And he says it's a bit
Californian for me, and he's, oh, you could sort of
feel the kind of anxiety of being seen in this crowd.
But that's why you had to go to meet Peter Teele.
And then when he met Peter Tiel, he had this
clever trick. Peter Teel is a chess player. Demis is
(17:16):
a chess player. So rather than pitch Peter tile on
some idea about a company, and he said, well, I
think the interesting thing about chess is that the knight
and the Bishop are supremely well balanced. And it's in
that tension between those two pieces that much of the
joy of the game resides. So Peter T' is like, WHOA,
that's a conversation I want to pursue. And so that
(17:37):
got him and got Demssus Habits an invitation to Peter
Tele's house the next day, and then that's when he
pitched him on Deep Mind and got the money he
needed to start the company.
Speaker 1 (17:46):
And then first forty twenty thirteen and which the excerpt
in the Wall Street Journal of your book tells the
story of a birthday party for Elon Musk, replete with
all kinds of costumes and strange things and fake battlements.
But this is, perhaps, apart from the founding, the most
(18:06):
crucial moment in Deep Minds Genesis as a company.
Speaker 2 (18:09):
Right, yeah, that's right. So you know, by this point
Demis had raised three rounds of venture capital, including from
Elon Musk, and you know, there are various people could
come in, but it was a total pain in the neck.
He hated it. You know. He would sometimes have this expression,
I don't want this part of my brain to expand
he wanted to be doing science, and so what he
(18:31):
wanted was to be liberated from this hamster wheel of fundraising.
And along at this party, this birthday party that Elon
Musk had, along comes Larry Page from Google who's also there,
and says, let's go for a walk, and they walk
around the castle grounds, and in this bizarre setting, Larry
Page says to him, well, you know, you could spend
(18:52):
your career building another company like Google. That's fine, but
if you really want to do science, just join Google
and we'll give the resources, use our platform, and you'll
be able to do what you really love. And Demis
not only agreed with that pitch in the sense that yes,
he preferred to do science then to be a billionaire,
(19:13):
but he felt that Larry Page himself would have accepted
that pitch. That Larry Page cared about science, he could
have been a standard professor of computer science. So Demis
really identified with Larry Page, and that was why he
sold the Google.
Speaker 1 (19:26):
And Page had his eye on Demis or this was impulsive.
Had he planned out his chess game for this party
like Demis had three years before, he had.
Speaker 2 (19:34):
Totally planned the chess game. He'd been thinking for a
while about buying up nascent Ai companies, and he'd bought
the boutique founded by the Toronto professor Jeffrey Hinton together
with Ilias Askeva and one other person, and so he
was in a buying mode.
Speaker 1 (19:52):
That was twenty twelve, right the image net team exactly.
Speaker 2 (19:55):
He bought the image net team, and then the next
obvious person to buy was was and DeepMind, because they
had a different approach to AI. It wasn't just deep learning,
which is the ImageNet secret source, which is kind of
packet pattern recognition learning from data. It was also what's
called reinforcement learning, which is learning through trial and error
(20:17):
in a simulation. So you have a game like the
Atari games or go later on, and you try lots
of different the computer, try thoughts and move c suites,
run works and then learns through trial and error. And
in some ways, another strand in my book is the
interplay between deep learning on the one hand and reinforcement
learning on the other hand. And these two fields of
(20:39):
artificial intelligence, you know, have their different moments in the sun.
As the story progresses.
Speaker 1 (20:45):
Hinton talked He came on tex Stuff and talked about
how he ran an auction to sell image Net with Google, Microsoft,
and by Doo. But in the end he all he
really wanted was to go to Google for for Demis,
he was being courted as well by others, including a
dinner at Mark Zuckerberg's house in the I guess weeks
or months after this first meeting with Larry Page at
(21:06):
Elon Maas's birthday party, and he submitted Mark Zuckerberg to
a test at this dinner.
Speaker 2 (21:14):
Right, Yeah, that's right. So the test was a bit subtle. Predictably,
they sit down to dinner and Mark Zuckerberg, who's longing
to buy deep Mind to get one over Google.
Speaker 1 (21:25):
This was not recently, this was ten years ago.
Speaker 2 (21:27):
It is twenty thirteen. So Mark Zuckerberg says, well, I
think AI is the most important technology in human history.
It's extraordinary, and you know, I really hope you agree
to join me at Facebook because you know, we could
just do great things together. Blah blah blah blah. And
then you know, the conversation moves on, time goes by,
(21:48):
and then Demis slightly says, you know, three D printing
is extraordinary, and Zuckerberg goes, yeah, I agree, you know, incredible,
that's just going to unlock so many things. And then
a bit later, Demis says, you know, official reality, that
really is going to be transformative. As Zuckerberg's like, yeah,
it's transformative. It's so exciting. I'm so excited by that.
(22:08):
And then Demis's mind is wearing. Is said, Okay, he's
a bullshit artist. He does not believe that AI is
the most important thing. Ever, he does not get it.
I'm selling to Google, Forget, forget Facebook.
Speaker 1 (22:19):
Even though the more money on the table.
Speaker 2 (22:21):
Yeah, that's right. In fact, Facebook was offering to make
Demis a lot richer, but he was consistent throughout his career.
Demis in turning down the money not to go to
Cambridge University, turning down the money to sell to Facebook.
It's not about the money for him, it's really about
the science.
Speaker 1 (22:34):
And you mentioned him using his scientific method to see
two three years into the future. Instead, Facebook went with
Jan Lucun and gave him plenty of resources, and I
think he was trying to poach some of Demos's employees.
Demists told them that the Google deal was going to
happen and to sit and therefore to sit tight, and
they did. But you know, fast forward to twenty twenty
six and Jan Niklean hasentually been dumped from Meta, and
(22:58):
Demis is where he's sitting Google is he is? He
is he the successor to Sundar. Is he the you know,
the ego and the ID. I mean, what is his
role in Google today?
Speaker 2 (23:08):
Well, what his rold is today is to be the
chief executive of Google Deep Mind, which is the AI engine,
which is basically powering all the new products in Google.
So he's super important. Sundar is the chief executive of
Alburt and Google, and I would argue that the relationship
between Sundar and Demis is the most important relationship in
business anywhere at the moment. Because Sundar has Demiss back.
(23:32):
Sundar gives him the resources. Sundar takes care of the
kind of all that kind of corporate leadership staff that
Demis is good at, but it's really not what he
wants to do full time, and that gives Demis the
oxygen to pursue AI to the fullest of his abilities,
which are considerable you know, in the future, if Sundar
were to go, I don't think that's happening anytime soon,
(23:53):
by the way, but I think if he were to go,
you know, Demis would obviously be talked about as a candidate.
And it's a really interesting question because he is on
the one hand, somebody who is a leader, has vision,
can motivate people, would have the credibility to lead Google
as an AI company. I mean, how often do you
get somebody who's the CEO and also has a Nobel
(24:14):
Price That would be quite something. But at the same time,
Demis has a side him that wants to be a
pure scientist, that talks to me about you know, there's
too much noise in Silicon Valley. I want to go
and think I want to have a research professorship at Princeton.
That's where Oppenheimer went after the Manhattan Project. That's where
Einstein went, That's where I should be. You know. He
has that kind of you know, retreat to the idyll
(24:37):
of abstract contemplation side to him, and he's so good
at both of these things. It's what makes him exceptional.
I mean, if you mentioned Jan Lukun, you know, very
good scientist, but clearly not a great operator inside business.
You know, one could talk about Simultman, a great business operator,
but not a scientist. Dropped out of Steinford, doesn't have
a degree, you know, it's very rare to find both
(24:59):
in the same person.
Speaker 1 (25:09):
After the break is Demis an evil genius stay with us.
You mentioned earlier in the conversation this kind of journey
(25:33):
Demis had been on where one sort of safety mechanism
after another that he believed in fell away, and thus
this kind of metaphors about the atomic bomb. But ironically,
in some sense, the kind of safety to the wayside
(25:54):
race that we're in today with AI was kicked off
by Demis his desire for a safety board.
Speaker 2 (26:03):
Yes, that's a good irony, you're right. So what happened
was that Demis sold the company to Google in twenty fourteen,
and one of the conditions was there had to be
a safety oversight board and whereby Google would allow Deep
Mind to sort of appoint some you know, important philosophers
or other people of independent stature to make a final
(26:23):
decision on when AI would be deployed into the world.
And the idea was this is AI is too big
just to let the corporate board of Google do whatever
it wants with it. You know, there has to be
a check. So the first of these safety meetings was
arranged and Demis had the idea, we'll invite Elon Musk
to shair it, and he invited Rieed Hoffmann and various
(26:45):
other people and they all met at SpaceX and basically
what happens is Elon Musk sat there listening, absorbed all
the presentations from Deep Mind about their plans to build AI,
and a few months later he announces Open Ai, which
is going to be the rival company. And so all
of a sudden, this singleton vision, the idea that you know,
(27:07):
only one AI lab would shepherd AI into the world
on behalf of all humanity that just is by the wayside,
and you've now got two competing labs, and the race
dynamic begins to set in.
Speaker 1 (27:20):
How did demos feel about what Elon did?
Speaker 2 (27:23):
Betrayed? Elon had sat there listening to all his plans,
and he'd been invited to chair that meeting in good
faith to ensure safety for the world, which of course
is what at the time Elon was a big duma
and was constantly talking about AI safety and existential risk.
And so the idea that rather than uniting with deep
(27:44):
Mind and Google in a single effort to make the
technology safe, Elon must prefer to go off and start
a rival in Open AI to Demis. This was a
total betrayal. Of course, Elon thought of this as Demis
is dangerous, he's an evil genius, and therefore I need
to be the one because you know, all of these actors,
(28:07):
they basically say, I know that I'm a good person. Yeah,
if I'm the leader of the AI race, I will
make it safe because I'm good. But those other guys
over there, you can't trust those guys because you know whatever. Now,
if you quizzed Elon Musk about why did he say
that Demis was an Elon was it was an evil genius?
Speaker 1 (28:25):
Your term for a Freudian slip.
Speaker 2 (28:29):
Elon evil genius? Why was Demis an evil genius? Well,
the only good reason, or not a good reason, but
the reason was apparently Demis, in his game design days,
had worked on a game called Evil Genius. No, which
is a pretty thin basis on which to call him
an evil genius. But whatever I mean, they all had.
Speaker 1 (28:48):
Association Sean Elbows. So then that this is twenty this
is meeting is in twenty sixteen, twenty fifteen, twenty fifteen,
and when is the alpha go moment?
Speaker 2 (28:59):
Six? Okay, so coming out of that moment, when Elon
Musk decides to set up open. Ai Demis decides, well,
I'm just going to accelerate as fast as possible, and
the first thing he manages to score is this victory
over the Korean Go champion Lisa Dol And it's a
(29:19):
huge exhibition match in South Korea with all the media
in attendance, and it's kind of an it's not quite
chatchypt but it's a moment when Ai had what one
might call the Kasparov Deep Blue moment in nineteen ninety seven,
first time the human champion gets defeated, and then twenty sixteen,
(29:40):
so that's nineteen years later, the same thing happens with
Go and.
Speaker 1 (29:43):
Two hundred million people tune in and the defeated Korean
player apologizes to humanity. It's a huge moment, but it's
nothing like the Chatchipt moment six years later.
Speaker 2 (29:54):
Yeah, because Go people watched. Whereas chat ChiPT you used it.
It was personal, it was visceral.
Speaker 1 (30:01):
And within a week of you pitching damage on the book,
Chatchipt came out.
Speaker 2 (30:05):
That's right. And I went to see him right after
that and he said, you know, this is war. Those
guys have parked their tanks in our front yard actually
said on our lawn, but translating for American audience, in
our front yard. And so you could see that competitive
glint in his eye, and you knew he was going
(30:26):
to try and fight back.
Speaker 1 (30:27):
Was he self aware about the risk of using that
language even for himself given all these Manhattan Project analogies,
you know.
Speaker 2 (30:36):
He's a person with many different dimensions, and he's both
capable of worrying about safety and also using military metaphors
to express this determination to crush the opposition. And I
think actually it's going to be a business school case
study of how deep Mind made the comeback because they
emerged deep Mind the London lab with Google Brain the
(30:58):
Mountain View Google AI Lab. Normally, mergers are super difficult,
they don't work. And here was a merger you had
to do in the middle of an AI race which
had been kicked off by chatchept. You had eight time
zones between California and London. You had a record of
bitter rivalry between the AI scientists from Google and the
ones from deep Mind. And yet they pulled it off.
(31:20):
They did the merger, they blended the cultures, and within
two and a half years they had a model that
was outclassing open AI models.
Speaker 1 (31:27):
See that's just extraordinary. So I remember when the chetchipt
moment happened, and I would say, up until twenty twenty
the beginning of twenty twenty five, people were saying Google
is down and out, Google might be over. I mean,
you knew because you were reporting along the way that
probably wasn't true. But what the indications that you saw
that the rest of the world didn't that convinced you
(31:49):
along the way that demos and deep Mind might be
roaring back into Do you put them in first place? Now?
Speaker 2 (31:55):
I think it's sort of a pretty close race between
the Gemini models from Demis and then Claude is doing
really well at the moment the anthropic model. People love
it for coding and so forth. So you know, I'm
not sure that it's I think the race is still ongoing.
What I would say, though, is that you know, I'm
on record as having written in the New York Times.
Speaker 1 (32:15):
That around of money, right, probably run.
Speaker 2 (32:18):
Out of money. I mean, they may put it out,
but basically, in fact, since I wrote that piece, they
do seem to have focused their business quite a bit
by giving up on Soora for example, Saura was a
classic money losing idea. You know, it costs enormous amounts
to generate video, but people don't pay you to generate videos,
so quite rightly, they can do it. So maybe they
(32:40):
can cut costs enough to survive. But they have a
huge cash need and they do not have Google's deep
pockets behind them, unlike Demis.
Speaker 1 (32:50):
So Demis is kind of winning. But he said to you,
it doesn't necessarily feel like that, right, he said, this
is a paradoxical moment. It should feel amazing, but it
doesn't feel how I thought it would feel.
Speaker 2 (33:00):
Yeah, because early on he had this rather naive idea
that there would be one lab building AI and so
you could take your time about releasing the models, and
you know, if you were worried about safety, you could
just take another six months to test them. And now
you have this race, and you know, the Chinese have
plenty of models, and the other thing, it's not just
a race, it's actually also the open source nature of
(33:20):
these models, where they're being released out into the wild,
and some weird group can just download the model, have
it on their own computer, and then you can't pull
it back anymore. And so there was a big cyber
attack in Mexico recently where all of the electoral records
were stolen, and Anthropic realized that its clawed model was
(33:42):
being used. But because that model is proprietary, they could
immediately shut off access and stop the attack. You couldn't
do that with an open weight, open source model. And
yet we have open weight, you know, that's being put
out there, both by Meta and by the Chinese and
by Mistride in France. A lot of open source models
are out there, and so in many ways the way
(34:05):
AI is being deployed is frightening. The obvious safety measures
one might take are not happening. In addition to banning
open source, I think there should be much more powerful
sort of government oversight so that, just like with a pharmaceutical,
before you release it to be used in people, as
to go through clinical trials, so too, I think there
(34:25):
should be a sort of equivalent of the Food and
Drug Administration an AI agency that can actually veto the
release of really powerful models. And we don't have that,
and we should have that, and we should be negotiating
with China about doing it in both places at once,
because this is a global race and both sides have
to slow down. I was in China recently for eight
days because they always published books faster. So I was
(34:49):
meeting AI leaders, both from industry and from academia, and
I was surprised by how much they do talk about safety.
So I think there is a discussion to be had
with it Chinese about safety, but the US administration of
this moment doesn't want to do that.
Speaker 1 (35:06):
I mean, coming back to the Manhattan Project again, Demis
has said, I think that he thinks this may end
in a bunker and what does he mean by that?
And he has he primed himself psychologically for an Ai
Hiroshima that he may feel in some sense responsible for.
Speaker 2 (35:25):
Yeah, I mean when I was doing the research, interfering
not just Demis, but all the scientists that he works with,
you know, one hundred or something of them. In deep mind,
I would hear this, these references to the bunker come up,
and I assumed it wasn't literally, you know, a real
thing that Demis wanted to disappear into a bunker at
the moment when he thought the AI models were coming
dangerously powerful. And I would have these dinners every six
(35:47):
months with a friend who had been at deep mind,
but had left and I tested this on him one
evening and I said, yeah, surely this is just a
metaphor bunker. He can't be serious. This guy said, well, actually,
you know, I had my bag packed. It was serious.
There was actually this vision that AI would become so
(36:09):
powerful that bad guys would try and get it off you.
So you had to hide in some place a bit
like Los Animals and develop in sort of isolation and secret,
and also be isolated because you needed maximum focus on
the science to get it right when you were at
this moment of maximum danger because the model was suddenly
very powerful, and that was his vision. Now I think
(36:30):
today he doesn't believe that anymore, because we're so far
from a single lab, you know, midwifing AI. So I
think now he's more inclined to speak of some version
of the Center for European Nuclear Research SAN, which is
a sort of technical agency that oversees nuclear power on
a multinational basis. I think you would like some sort
(36:53):
of global body to impose rules on what kind of
AI should be let out into the wild. But you know,
at the same time he knows that politically that's not
on the cards, and he has a sense of timing
about when you should raise these issues, and so you know,
whereas Dariama Day took on the Pentagon by trying to
(37:17):
assert safety principles and then just got rolled, I think Demis,
when he does that, is going to feel that he's
got the door is half open and he can give
it a push and we'll see. You know. Of course,
sometimes people keep their capital drive for so long that
they never use it, but we'll see. If the moment
comes when he does use it will be very interesting.
Speaker 1 (37:37):
Especially just to close, there was a great review of
your book in the Financial Times which ends with this,
Whether and how Demis ever achieves AGI will form the
defining chapters of his extraordinary and unfinished biography. What did
you think about that? And what is the next chapter
for him? And will you write another follow up book?
(37:57):
Do you think?
Speaker 2 (37:57):
You know? I tend not to write follow ups about
the same thing the same person. I prefer to plow
fresh ground. But look, I mean, you know, Demis t
haming fifty this year. He's got a lot of runway.
I'm sure he'll do more incredible things in the future.
So probably I am offering an interim report. But the advantage,
you know, if you wait. I did this before with
Alan Greenspann. I wrote the definitive biography after he retired,
(38:20):
and by that time, you know, people are interested, but
less so than when he's stood in the seat. I
think capturing a portrait of you know, the most interesting
figure in artificial intelligence in real time while he's still
in the steat and he's still doing it is sometimes
the fun of it, right, I mean, who wants to
wait for the definitive biography in twenty years time?
Speaker 1 (38:40):
Well, next for him, you know, I think he's.
Speaker 2 (38:41):
Going to carry on running Google Deep Mind. There's going
to be more agentic models coming out this year. There
will be you know, world models and more robotics coming
There will probably be much more AI for science, both
in terms of drug discovery and in terms of you know,
material sciences, chemistry and so forth. So I think, you know,
(39:02):
one day, I remember, towards the end of my time
interviewing him, he showed up at the pub and he
had a backpack and he pished something out of it,
and he got this little box out and he said,
I got to show you this, and he opened the
box up and inside was the Nobel prize medal and
either at that meeting or another one. He said to me,
(39:24):
I wonder if I can get another one. He's not
over yet.
Speaker 1 (39:28):
It's a vasty malabi. Thank you.
Speaker 2 (39:29):
It's been great fun to talk.
Speaker 1 (39:46):
For tech Stuff. I'm as Volosciin. This episode was produced
by Eliza Dennis and Melissa Slaughter. It was executive produced
by me Julian Nutter and Kate Osborne for Kaleidoscope and
Katrina Novel for Ihart Podcasts. The engineer is p Bowman
and Jack Insley makes this episode. Kyle Murdoch wrote our
theme song. Please rate, review, and reach out to us
(40:07):
at tech Stuff podcast at gmail dot com. We also
love to hear what you think our panels should cover
next time
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Girlfriends: Trust Me Babe
When a group of women from all over the country realise they all dated the same prolific romance scammer they vow to bring him to justice. In this brand new season of global number 1 hit podcast, The Girlfriends, Anna Sinfield meets a group of funny, feisty, determined women who all had the misfortune of dating a mysterious man named Derek Alldred. Trust Me Babe is a story about the protective forces of gossip, gut instinct, and trusting your besties and the group of women who took matters into their own hands to take down a fraudster when no one else would listen. If you’re affected by any of the themes in this show, our charity partners NO MORE have available resources at https://www.nomore.org. To learn more about romance scams, and to access specialised support, visit https://fightcybercrime.org/ The Girlfriends: Trust Me Babe is produced by Novel for iHeartPodcasts. For more from Novel, visit https://novel.audio/. You can listen to new episodes of The Girlfriends: Trust Me Babe completely ad-free and 1 week early with an iHeart True Crime+ subscription, available exclusively on Apple Podcasts. Open your Apple Podcasts app, search for “iHeart True Crime+, and subscribe today!