Garry Kasparov: “Deep Thinking” | Talks at Google

Garry Kasparov: “Deep Thinking” | Talks at Google


[APPLAUSE] DEMIS HASSABIS: Well, thank
you, everyone, for coming. It’s a really special
privilege and honor for me, actually,
to talk to Garry, in my opinion, my
humble opinion, the greatest chess
player of all time. And you know, I’ve
really enjoyed his book, which I reviewed recently. And you know, I was impressed
with Garry’s understanding of artificial intelligence and
the latest advances in that. So you know, it’s going to
be great to talk about that as well as chess today. GARRY KASPAROV: Yes. DEMIS HASSABIS: So welcome
to Google and DeepMind. GARRY KASPAROV: Thank you
very much for your review. It offered me all the
protection against all the tech guys that tried to criticize
me for not being an expert. DEMIS HASSABIS: Well, I’m
glad that that can be of use. So before we get talking
about Deep Blue match– I’m sure everyone’s
going to want to hear about your insights on
that and also machine learning more generally. I wanted to begin by asking
you about growing up as a chess champion in the Soviet Union. Did you always want
to be a chess player, world champion at chess? Did you consider anything else? Or were you, from a
very young age, decided that this was your path? GARRY KASPAROV: I
learned how to play chess when I was five or six. I’m sorry. I couldn’t give you
an exact moment. Nobody was there
to tweet about it. It was late ’68,
maybe early ’69, watching my parents trying
to solve chess problem. And I loved the
game at first sight. And ever since, I’m still
in awe with the game. And I could feel that was
a match made in heaven. And everybody around
also could see that chess was a perfect fit for my mind. DEMIS HASSABIS: Skills, yeah. GARRY KASPAROV: Yeah. DEMIS HASSABIS:
And, actually, you talk about, in the
book, about chess informing all of your thinking
and the rest of your life, right? What do you mean by that? What skills can you see yourself
using in the rest of your life that you learned from chess? GARRY KASPAROV:
Naturally, if you are engaged in a competitive
sport at such an early age, you see many things just as a
reflection of your chess games, your engagement. Because you have to
play, you have to win. It means you have to
change certain habits and certain customs. And what was important
for me, that’s what I learned from my mother,
is that my game was not just about winning. It was also about
making a difference. And that’s what
helped me to make a transition later on in
my life from playing chess, being number one chess
player for 20 years, to other things
that I’m doing now. Not pretending that
I could be number one and repeat my
outstanding achievements in the game of chess,
but still recognizing that I’m quite useful. Because I’m trying to
bring my chess experience, what I learned from the game
of chess, my analytical skills, to make the
difference elsewhere. DEMIS HASSABIS: So maybe we
should talk about, obviously, the heart of your book,
which is the Deep Blue match. You know, I was
fascinated to see– having gone through the
AlphaGo matches ourselves from the other side– your take on it from the
player’s perspective. GARRY KASPAROV: It’s
an interesting story. Because when we played
our first match– and I always want
to remind people that there were two matches. And I won the first one. DEMIS HASSABIS: Yes, exactly. We should make that very clear. GARRY KASPAROV: Yes, yes, yes. DEMIS HASSABIS: I was
going to say, exactly. GARRY KASPAROV: Now, the
first match in Philadelphia, it was organized as
quite a low-level event. The corporation was
not even involved. It was ACM behind it. And the team always
wanted to challenge me. And I had an experience of
playing against them in 1989 when they had Deep Thought, the
prototype from Carnegie Mellon that they brought to IBM. It’s turned into a
Deep Blue project. And everything has
changed after game one. By the way, if we’re talking
about a watershed moment, that was in February
1996 in Philadelphia when I lost game
one of that match. Because the rest is, you may
argue, a matter of technique, matter of time. It was like signing on the wall. If the machine can beat a world
champion in one game, then– DEMIS HASSABIS: Eventually. GARRY KASPAROV: Eventually. DEMIS HASSABIS: Yeah. GARRY KASPAROV: I fought back. I won the match. I won game two and
game five and game six. But it was pretty clear that the
rest would be a matter of time. And the first game had
some kind of a record following on internet. I think the numbers they
had, there were even higher numbers later on in Atlanta
when IBM ran the website there. And suddenly, the corporation,
Lou Gerstner and his team, discovered the huge potential
at a rather low cost. And while the rest of the
match was still played just about purely chess,
then it turned into a big corporate endeavor. And look, it’s water
under the bridge. 20 years ago, I lost the match. But I think, and I
discuss it in the book, I made many mistakes
in preparation. And one of the biggest
mistakes, and that’s why I was so upset
with myself, is that I didn’t treat
IBM, Deep Blue, as just an opponent the way
I treated Anatoly Karpov or Vishy Anand or Nigel Short. For me, I was still part of
a great social and scientific experiment of the end of the
20th century, something that could help us to understand
more about how we humans make decisions, how machines
can play with us. It was not just
winning or losing– big mistake. DEMIS HASSABIS: Yeah. GARRY KASPAROV: Now,
for IBM, it was just about winning or losing. DEMIS HASSABIS: Yeah. GARRY KASPAROV: Yeah. And one of the big
mistakes that I made while signing the
contract, you always have to read the fine print. DEMIS HASSABIS: Yeah. And say, hey, that
rematch clause is– GARRY KASPAROV: And
because when people ask me whether IBM cheated– no. They just bent the
rules in their favor. They followed the letter but
not the spirit of the agreement. And, for instance,
one of the big issues after our first match
in Philadelphia for me was how can I prepare if
I didn’t have any games? This is the normal
way to prepare. You look at the games
of your opponent. And Deep Blue in
Philadelphia was a black box. I had no idea what
it was capable of. There were so few games
that the machine played against other computers,
but it was not the machine that faced me. Now one year–
more than one year, 15 months– between the
first match and the second match, and I was
under the impression that I’ll be treated fairly. Because after my first
victory in Philadelphia, I went to Yorktown Heights. I sat with that team. They had similar regiments for
all IBM labs around the world. So the atmosphere
was very friendly. DEMIS HASSABIS: Yeah. And then it changed. GARRY KASPAROV: And then
I expected the games to be provided. And then they said, wait a
second, but did you read? The game’s played in
official tournaments. And, of course, Deep
Blue hasn’t played a single game outside the lab,
which means that in May 1997, I faced another
black box that I knew was much stronger than it
was before, but I still no idea what to expect. I knew they had a
professional team. So they made a
massive preparation. And I have to also admit my
preparation was quite lousy. Because, again, only
just before the match, a week before the match,
I realized how difficult the challenge could be. But also, one of
the key elements of this contract that
I totally overlooked was about machines rebooting. That’s a big issue. DEMIS HASSABIS: Yeah. GARRY KASPAROV: You
understand what it is? Here, I don’t have to explain. But the general audience
doesn’t understand it. The moment you
reboot the computer, you will never be able to
reconstruct the game, which means that is a whole idea that
the match is fair and square. And you can always go back
and see why Deep Blue made this move or that move. It’s over. And also the– DEMIS HASSABIS: Yeah. Maybe people didn’t
realize that. I didn’t realize it till I
read the book that in fact several times– GARRY KASPAROV: They rebooted. DEMIS HASSABIS: –that
Deep Blue crashed. And then they rebooted it. GARRY KASPAROV: We don’t
know that it crashed. DEMIS HASSABIS: And then they
came up with a different move in a couple of situations. GARRY KASPAROV:
Look, no, no, no. It doesn’t matter
what the move was. It’s just the crash– if you played a match, anything
but the problems was agreed– crash. You lost the game. Heart attack. Sorry, go to see a doctor. [LAUGHTER] DEMIS HASSABIS: Should
have just been a loss. GARRY KASPAROV: No. They have Ken Thompson,
the great computer expert, who was there in
Philadelphia helping me. He was also in
front of the screen. But on the screen, you could
see the Deep Blue communication back to the programmer. But you didn’t see whether
they said anything– DEMIS HASSABIS: The other way. GARRY KASPAROV: –the other way. Again, I don’t know. But it definitely created a
lot of tension in the match. And, you know, after
losing game two– which that’s another story– I was very upset. And I demanded logs. And by the way, if they
wanted to play a fair game, all they had to do
is produce the logs to prove that my suspicions
were not well-founded. They didn’t do it. DEMIS HASSABIS: Yeah. GARRY KASPAROV: They just
wanted me just to be inflamed. Because they realized that
while Deep Blue was not that strong at that time, I
still think I was stronger. DEMIS HASSABIS: Yeah. You were definitely stronger
I would have said, too. GARRY KASPAROV: Now,
20 years later, you can look at the games. You know, you can take a
chess engine on your laptop, and you’ll find out
that many mistakes were made from both sides. I mean, one of
the most amazing– not even game two,
but game five– the end game, reaching the end
game, I was slightly better. And everybody at
that time in 1997 believed that was a brilliant
escape by Deep Blue. Now, in 30 seconds
to one minute, it depends on the strength of
the speed of your computer, chess engines like
Stockfish or Komodo will tell you that the
end game was a draw. Deep Blue made a bad
mistake, and then that missed the win, which no one saw
in 1997, including Deep Blue. So that tells you that’s
the draw of the strength. And I think that if we played
the short match, the rubber match, I still had a
good chance of winning. Again, it wouldn’t change any
sort of long-term outcome. But at that– DEMIS HASSABIS: How long do you
think you could have held them off for at your full potential? GARRY KASPAROV: Maybe
two, three years. DEMIS HASSABIS: Yeah. That’s what I would
have guessed, too. GARRY KASPAROV: Deep Blue
maximum two, three years. I played two more matches with
Deep Fritz and Deep Junior in 2003– both ended in a draw. So that was a balance
in the next five years. But in 1997, they realized that
if putting pressure on only one human player in a match, they
could achieve the result. If you cannot make
your player stronger, you can definitely inflame the
other player and took him off balance. DEMIS HASSABIS: So sort of
moving more to the present day now, how have chess
computers changed chess? Do you think it’s
for the better? It’s just different? What do you think
about that evolution? GARRY KASPAROV: It’s
something that you said that it’s quite striking. Because you said is it
for better or worse. It’s happening, period. The technology is
neither good nor bad. It’s agnostic. You know, you can do many great
things with your mobile phone. But you can also create
a terrorist network. So it’s happening, and
we just have to adjust. And as for the
game of chess, it’s different, because the young
generation of chess players, they learn very
differently from us. I remember I had books. And not so many new books you
can buy in the Soviet Union. Every book was cherished. And I had my notebooks. When I went to
the top and played world championship matches, I
had also notebooks and recorded my analysis. And I treasured them. I remember I had a couple
of quite thick notebooks with analysis. And they were just top secret. And I believed in 1985, in 1986,
1987, that was a real treasure. That was a powerful weapon. It’s like the magic
sword of Merlin. Now, when you look at this
analysis with computer, you understand it
was a broken knife. But also, when you look
at young chess players, under the umbrella of
Kasparov Foundation, I have been involved
in working with them. And I’m talking about kids
of international masters, grandmaster level. It’s such a difference in the
way they approach the game, The way they look at the pieces. It happened time and
again where you’re reaching a certain position
analyzing the game, and they say, bad move. I made a mistake here. I said, fine. Why? Oh. And then it’s a long line. So the machine show– I said, I understand. I could see the screen. But why you think
this move is wrong? And they don’t
understand the question. Because the machine said so. Because it’s on the screen. So somehow their
mind’s being hijacked by the power of the machine. And one of the reasons Magnus
Carlsen was so successful and still a dominant force
in the world of chess– and I remember after working
with him in 2009, 2010, for more than a year, he
never looked at the machine as an ultimate source of wisdom. For him, it was more
like a big calculator to verify his own understanding
and evaluation of the position. This is a big challenge. But I believe it’s
not only chess. It’s elsewhere. Many people just are
staring at the computers. Eyes are just being caught
by the screen expecting to find a solution there. DEMIS HASSABIS: Instead of
thinking for themselves. GARRY KASPAROV: Exactly. So that’s why I always
bring, as a piece of wisdom, the classical phrase from
Pablo Picasso that computers are useless, because they
can only give you answers. DEMIS HASSABIS: Yeah. GARRY KASPAROV: But everything
begins with a question. DEMIS HASSABIS: Since you’re
talking about Magnus Carlsen, you say that interestingly,
although he’s grown up in the
computer chess era, he’s one of the most human– I think you called it– or
intuitive players around. Right? So it’s kind of interesting. GARRY KASPAROV: He’s consistent. Yes, it’s human. Because at the end
of the day, 20 years after my match with
Deep Blue, more people playing chess than ever before. And chess is still very popular,
because at the end of the day, it’s a fight between
two individuals. And what has changed is
not just the game itself, but the way people
are watching it. 20 years ago, or 30
years ago, 40 years ago, the world championship
match was kind of an event of absolute quality. Even Karpov and Kasparov
played the game, and one made a terrible blunder. It could take time in the
precedent of the grandmasters to– DEMIS HASSABIS: To find out. GARRY KASPAROV:
–whisper it, mistake. And it’s something that
you should worship. Today, when I’m watching
the games by Magnus Carlsen, Caruana, and you have
thousands of amateurs from all over the
world watching it. Because they’re screaming,
ah, mistake, mistake! Because the machine shows
immediately– it says, evaluation, drop. So some kind of respect
has disappeared. DEMIS HASSABIS: Yes, from that. It’s a real shame. GARRY KASPAROV: But
also, it added interest. Because people can follow. They have access
to their computer. And they don’t have
to be strong players to understand what is happening. DEMIS HASSABIS: One of
the interesting things you said, actually, about
the chess computers– and I wonder if it’s going
to happen with Go as well. In the countries that are not
traditionally good at chess or Go, because they have
access to these machines, maybe kids in those countries
can now get very strong, right? Like, Magnus in Norway. Or I don’t know whether that’s– GARRY KASPAROV: I’m not sure
Magnus’ rise, meteoric rise was due to computers. Maybe it’s because
in this environment you don’t have to spend
so much time learning from other players. So the process of maturing
for the chess player is much shorter. You have grandmasters
at 14, 15 today that know much more than Bobby
Fischer knew 40 years ago just because they’ve
played better games. They could travel around. They could watch the games. So chess is a perfect
match for internet, because you can
follow the games. You can learn. You can analyze. So there are many things you can
do that dramatically increase the pace of learning
and getting to the top. DEMIS HASSABIS: So
you invented, I think, the concept of
advanced chess, right? Man and computer. GARRY KASPAROV: Human. DEMIS HASSABIS:
Human and computer. GARRY KASPAROV: Sure. DEMIS HASSABIS: Human and
computer versus computer. Have you tried that recently
with the latest chess engines? Is that still true now? GARRY KASPAROV: Yes. While licking my wounds after
Deep Blue match, so I thought, how about bringing it together
just out of curiosity? Because I said, wait
a second, if I just can play with who is a machine,
just against another player, maybe we can play perfect chess. Now, the interesting
thing is when we played this match
with Veselin Topalov, another top player, in
1998, I can tell you the quality was not very high. Because it was a
limited amount of time. And it was so new for us,
how to use the machine. And eventually, I realized–
and we had many events, so-called freestyle
events on the internet, that proved that– it sounds quite ironic– you don’t need a
very strong player to get the best result of
human plus machine combination. It could sound
like a heresy now. But I would say that you
don’t want a strong player. You need a good operator,
a decent player, but someone who will follow
the machine as you guide the machine, but not
to use the machine to back up his or her own ideas. Because, instinctively, if
I team up with a computer, I’ll try to make my own moves. You don’t have to. All you need is just to maximize
the effect of machine’s play. Because machines are so
strong now all you have to do is just to guide them. Sometimes you can feel, no,
just a slight correction. Move here, move there. So it’s something that
requires very different kind of qualities. It’s more about interface. So you don’t need a great
knowledge of the game. It may help. But on the other side,
it may preclude you from sort of using
the machine’s power. Because you’ll try to
play your own game, which could be detrimental. DEMIS HASSABIS:
So it’s something that I think you touched
on in a few places, it’s become known as
Kasparov’s law now, right? Something like it’s where
the process is actually more important. Do you want to
explain what that is? GARRY KASPAROV: And
again, I relied on results of the freestyle tournaments. And what’s happened is
there, as predicted, a human plus machine beat
supercomputer quite handily. But the most unexpected
story was that, as I described, a relatively
weak human or group of humans plus machine or machines plus– DEMIS HASSABIS:
But a good process. GARRY KASPAROV: –a
better process, of course they beat a supercomputer. More remarkably, they beat
a strong human plus machine plus inferior process. So that’s led me
to the conclusion that it’s all about interface. There’s so many ways
of empowering machines with our creativity. So not our creativity
with machine’s brute force of calculation. Actually, you do it
other way around. And then the result is
it could be phenomenal. DEMIS HASSABIS: So it
strikes me in the whole book, you’re very optimistic about
technology in general in terms of what it might be able to do. Is this kind of
process, is that a kind of blueprint for an
advanced chess of how you see things going forwards
in other areas of life with machines and
humans working together in a complementary way? GARRY KASPAROV: I
believe the future is a self-fulfilling prophecy. And I cannot stand all this
doom and gloom predictions. It was quite amazing
when you just look at the change of the
trend in science fiction from ’50s and ’60s where
it was all about optimism, us teaming up with
computers, robots, cyborgs, flying not just to other planets
but to other star systems. And then it changed to
a very dystopian vision of “The Terminator,”
and “The Matrix.” By the way, speaking about
“The Terminator,” recently I just got an idea of just
having a lecture in Dallas, Texas earlier this month. I looked at “The Terminator,”
I said, you know what, guys? I can tell you that’s
another proof of what you call Kasparov’s law. Because we all watched the
first one, human versus machine. But if you follow the
number two or number three, that was exactly what I
said, human plus machine plus a better process
beats the supercomputer. DEMIS HASSABIS:
With big machine. It’s very true. GARRY KASPAROV: Yeah. And I think what we
learned from chess is that there are
many ways of us sort of getting something
new, something positive, out of this cooperation. And by the way, these things
are going to happen anyway. So what’s the point of
trying to slow down? It’s a natural cycle. We have technology
replacing certain elements of human activities. For centuries,
technology was there. Machines have been
replacing blue collar jobs. Now, the difference
is now machines are threatening
people with college degree, political influence,
and Twitter accounts. That’s why we hear all
the stories about it, but that’s actually normal. I think that’s called progress. And if machine’s taking
over certain menial parts of our elements or
aspects of convention, that’s not the end of the world. There’s still many things
that humans can do. All we need is just to
look for new challenges and for new frontiers. DEMIS HASSABIS: So
we’ve just come back from China for the AlphaGo
match against Ke Jie. And one thing that
happens in Go, which is slightly different
than chess, is in Go, there’s a tradition
of players thinking about how far off
from optimal play are they, theoretical
God play or optimal play. So how far do you think
even the top chess computers are from optimal chess? I mean, what do you think
the top Elo rating would be possible to play chess at is? Do you have an idea? GARRY KASPAROV: No. I don’t have an idea. Because as we briefly
discussed at lunch, when you look at the
endgame databases, now we have all seven pieces. That’s 100 terabytes
or whatever. And so every position is being
calculated to the very end. And in many cases, you
just have a position that says made in 492 moves. And I bet you that in
the first 450 moves, you will not see the difference. So I could see probably
420 moves, yes? Now, I don’t know what it
tells about the game we play. Because the average
human game is 50 moves. Now, when you look at
average machine games, it’s maybe 80, 90 moves. It doesn’t mean that the
game should be too long. What we know, the game of
chess is ultimate endgame of 32 pieces. So that’s why I don’t see
any chance in any future that machine will
play E2, E4, and will announce made in 16,755 moves. It’s not going to happen. The number of legal moves in
the game of chess, 10 power 45, that’s enough to feel safe. But it’s not about
solving the game, it’s about winning the game. And I think there’s
still some improvement. Machines could get
better and better. I mean, basically,
the sky is the limit. And today, I still think Magnus,
who was White in his good day, would probably secure a
draw against the machine. But winning against
a computer today, it’s virtually impossible. The level of precision
that is required, the level of vigilance,
just it’s impossible. So we’re not used to
play with such attention. So machines will get better. And by the way, we see an
improvement all the time. I remember this,
as well, by writing my books, my great
predecessors, and then my matches against Karpov,
and then my best games. And some of the
games, the same games, analyzed two, three
years later was a just new version of the same engine. And just I could see that. Some of the moves that I
treated as great in, say, 2009, in 2012, was more
powerful computer. I had my doubts. DEMIS HASSABIS: Yeah, very cool. So, look, I’ve got so
many more questions. But I know I should
give some time– the time is moving forward. So I don’t know. I should let the audience
ask any questions. If you put your hand
up high so I can see? GARRY KASPAROV: Wow. Total silence. DEMIS HASSABIS: We
covered everything? Well, let me ask
another– oh, one second, we have a question down there. AUDIENCE: So now, I
wanted to ask you, I was too young when the
Deep Blue match happened, so I don’t have any
personal memories of it. GARRY KASPAROV:
You’re too young, yes. I can see. [LAUGHTER] AUDIENCE: But when I
read stories of it, it kind of struck me
that the match seemed like it had great publicity. But people really wanted to see
whether, well, to put it blunt, whether you would lose
against the machine. And I found this really– my question is,
basically, do you feel like this was the case? Or do you feel it actually
felt like a normal chess match, where people
would see who would win? Or rather, do you feel like
you had support from your side as well? GARRY KASPAROV: Oh, yeah. I had plenty of support. I can tell you that. Most of the people
who wanted me to lose, they were actually in
the world of chess. Because I was the world
champion for 12 years, and that was the first
event I ever lost. So, naturally, a lot of people
wanted me to lose one day. And since I was
unbeatable in human chess, they had hopes in a machine. [APPLAUSE] But the atmosphere
there was phenomenal. And it was a reflection of the
famous cover of “Newsweek,” “The Brain’s Last Stand.” And I remember when I won
game one of the match, it was a big celebration. I can remember on CNN,
the two presenters, they talked about
it and said, it’s a Russian playing
an American machine, but I’m rooting for a Russian. [LAUGHTER] DEMIS HASSABIS: Any
other questions? All right. AUDIENCE: Hello. [INAUDIBLE] from
the AlphaGo team. Humans seem to be more
efficient in playing both chess and Go probably in that they
evaluate much fewer variations and positions than
computers do, I mean, by many orders of
magnitude probably. Can you give us an
intuition of how this difference can come about? What are humans actually
so good at in chess that they can do
this so efficiently, and that they only need to
examine so few variations as compared to computers? GARRY KASPAROV: Now, we can
talk about general rules. But also you should
remember that there are different playing styles. Because the way Karpov or myself
will look at the same positions can be very different. Because I will maybe
look for an opportunity to break through just
to sharpen the game, to create complication. And Karpov will be
looking for sort of a long-term
strategic advantage that could manifest in the endgame. Those are the differences. What brings us together
is that, as you just said, we didn’t have to analyze
millions of lines. We couldn’t. So we could look for
one or two options. How do we know that those
two options are the best? I don’t know. I just simply know what it is. But, again, we had
another subtle difference. I will probably try to go as
deep as I can calculating. Karpov will try to
look for an option where he doesn’t have
to calculate at all, so relying on his understanding. Because there are many patterns. You can recognize patterns. And then bringing
patterns together, you can have a
picture, big picture. That’s what humans
are unique at. And that’s why, for instance,
if you team up with a computer, sometimes if you
start calculating– I wouldn’t go there. And then it would be
quite interesting to check whether the machine’s
calculation proves it. But in many cases, I
think I’ll be right, especially at a time when
a machine goes very deep and then reaches its horizon. And then you also should
look at the position and say, it smells. There’s something wrong. I don’t know exactly what is
wrong, but something is wrong. There are also situations
where you have to calculate, where you sacrifice something. You sacrifice material. And it’s take it or break it. So you cannot afford to
use your common sense, because you have material down. So my game with Veselin
Topalov, another one I played in 1999, that’s
my longest combination. So I cannot tell you that
I saw every line there. It would not be true. But the combination,
the final position that I saw like a
lightning, just very quickly what will happen at the end,
included the 30 ply lengths, 15 full moves. Ironically, because I saw
this, the final position, later the machine proved
that I could win earlier. And Topalov missed
the chance to– not to escape, but to have the
endgame that he could probably defend. Otherwise, what you described
is just another proof of the Moravec’s paradox. That’s the– DEMIS HASSABIS: I was
going to bring that up. You should explain what
this is, because you talk about that a lot in– GARRY KASPAROV: Yeah, exactly. DEMIS HASSABIS: –your book. GARRY KASPAROV:
Machines are very good at what humans are not so
good and the other way around. It’s interesting, chess was– probably, because
Go and shogi, they were just played elsewhere. But for the Western
science, chess was an ultimate test for
artificial intelligence. And that was another
result of the 1997 match, the expectations of the
founding fathers of computer science like Alan Turing,
Claude Shannon, Norbert Weiner, that machine
beating strong chess player and, of course, the
world champion, would be it. This is the moment for AI. I have to say they were wrong. So it was just this
people was as intelligent as your alarm clock. They tend to go along, but– DEMIS HASSABIS: I
actually have a theory about the Moravec’s paradox
in explanation for that. If you have hand-built systems,
like Deep Blue was, then as the programmers, you have to
understand clearly enough what you’re trying to
codify explicitly, so you can codify it in
the rules or heuristics like Deep Blue was. And the problem is is
that for many things that we take for granted
as humans, like vision or riding a bike, all these
things we do implicitly. We don’t explicitly
understand well enough how we do those things. So we can’t codify it. But that’s why I think
that learning systems, the kinds of things
like AlphaGo, might end up being
more powerful. Because they could
learn from experience how to do those
things like humans do. GARRY KASPAROV: It’s
one of the rules that are learned from my experience
is that anything that we do and we know how we do,
machines will do better. Because we can communicate it. So it’s, one way or
another, codified. So the big question is now
whether machines can ever do– DEMIS HASSABIS: The
intuitive, implicit things. GARRY KASPAROV:
–things that we do without knowing how we do them. DEMIS HASSABIS: Yeah. No, exactly. That’s the big question, right? I mean, I think you
say in your book– up till now, anybody who
attempted learning systems, including your great
teacher Botvinnik, fell short against especially
the hand-coded systems. GARRY KASPAROV: Oh, yeah,
50 years ago, 40 years ago. Because in the beginning,
there was a big debate. And I think Turing– people know, by the way, that
he wrote the first program. In 1952, there was
a chess program. And the trick was that
there was no computer. It’s the only game that
the Turing program– DEMIS HASSABIS: He
wrote it by hand. GARRY KASPAROV: Yes, exactly. He put it on a piece of paper
and calculated the moves. And when I spoke
at the centenary, I asked my friends from
Germany, they actually reconstructed it and
put it in a computer. DEMIS HASSABIS: How cool. GARRY KASPAROV: You can
actually play a Turing machine. Pretty weak, but it’s from 1952. And that’s interesting. They believed that the way
to make machines play chess– it’s not brute force
but understanding. But this concept
failed very quickly, because brute force kept coming
and was like an avalanche. They couldn’t stand a chance. So that’s why the old
attempts, including one of my great teachers
Mikhail Botvinnik, to come up with this parallel
concept of learning failed. And by the end of the ’60s,
early ’70s, the story was over. Now, it seems that we’re
just like in seasons. We go back to this notion. And maybe it will
prove to be superior. DEMIS HASSABIS: Well,
hopefully, AlphaGo will make Botvinnik happy then. You know, it’s sort of a
learning system, right? So it turns out that
Go needed to happen. What do you think
the difference is between Go and chess
that required Go to have to have this other approach? They couldn’t do the
handcrafted approach. GARRY KASPAROV: It’s
a tough question, because I have near absolute
knowledge of the game of chess and almost zero knowledge
of the game of Go. So from what I
know, it’s that Go doesn’t have the same
tactical configuration. So it’s all about strategy. It’s a long-term. And that’s why it’s far
more difficult for machines to learn how to do that. But also, if machines could
do it at a certain level, that could be deadly for humans. Because they suddenly
become superior. And, again, I’m not sure, since
I’m speaking to a great expert. I still think that relatively,
if you compare the strengths, I think the chess playing
engines are relatively stronger than AlphaGo. It’s in absolute ratings. Just because the mistakes
made by human players in Go, they are deadlier. They could offer more
openings for the machine. So in chess, the human
game is always unstable. So it’s not as steady
as the machine. But I think in Go, the
depths of the mistake could be far more
significant than in chess. DEMIS HASSABIS: I guess we’ll
have to put that to the test by teaching AlphaGo chess
to play chess, right? And then we’ll see. GARRY KASPAROV: Ah, see,
that will be interesting. DEMIS HASSABIS: Yeah. GARRY KASPAROV: See how soon
AlphaGo can crush the strongest chess engines. Again, it’s what I said,
it’s not about AlphaGo. But it’s about the nature
of the game Go and chess. But you know– DEMIS HASSABIS: Would
you be pretty surprised if a learning system could beat
the hand-crafted system ever in chess? GARRY KASPAROV: Look, that will
be another level of experiment. Because the current
systems, it’s not primitive brute
force anymore now. That’s why I said, today– and by the way, the
moment I say it, people just look
at me in disbelief from nonprofessional audiences. I say, the free chess
app on your mobile phone is stronger than Deep Blue. They say, ah, no, no, no. They say, you are sore loser. Yes, I’m sore loser. [LAUGHTER] It doesn’t change. This is the fact. DEMIS HASSABIS: No, very good. So I don’t know if
there’s a last question from the audience. We have time for one or two. Yes, from the lady at the front. GARRY KASPAROV: Wow. DEMIS HASSABIS: Yes,
at the front here. AUDIENCE: If you could, would
you play Deep Blue again? GARRY KASPAROV: Oh, [MUMBLES]. Yeah. There are a couple of problems. One, I’m retired, and I don’t
play professional chess. Two, Deep Blue is dead. Yeah. I wanted to play in 1998. And I wish I had a chance. But that’s it. You know, that’s old history. It’s spilled milk, water
on the bridge, you name it. But I played other computers. And as long as I was
an active chess player, I never ducked a challenge. And that’s why this book
begins with the story of me playing 32 chess
computers in 1985, a simultaneous exhibition. I’m not sure, but anyone still
owns antique chess machines? Anybody here? DEMIS HASSABIS:
I still have one. GARRY KASPAROV:
You still have one? DEMIS HASSABIS:
Yeah, I might have one called Kasparov’s version. Yes. GARRY KASPAROV: So I played. It’s 32 machines. And there were
four manufacturers, eight machines each. And I won all the games. Most amazing thing that
nobody was surprised. And I can tell you the progress. It’s from that match in 1985– I played in June, just
a few months before I won the title beating Karpov– to my match with Deep
Blue, just 12 years. It tells you that
something’s happening. But I couldn’t help
but reminding people about this match in 1985. Because I say that
was the golden age. Yeah. Machines were weak. My hair was strong. [LAUGHTER] DEMIS HASSABIS: Yes, we’re
entering another very exciting, I think, interesting era. Look, let’s all thank Garry
for an amazing discussion. GARRY KASPAROV: Thank you. [APPLAUSE] DEMIS HASSABIS:
Thank you, Garry. GARRY KASPAROV: Thank you. DEMIS HASSABIS:
Thank you very much.

10 thoughts on “Garry Kasparov: “Deep Thinking” | Talks at Google

  1. Is kasparov genius oh yeah! but what he did for you or me he didnt do anything for the world he just crushed his oppenent and liked it and lived for it so i think we cannot learn anything from this guy .

  2. A stuttering mad man…..yes my friend deep blue is terminator……. human plus machine plus a processor ….. wow !!!!!!

  3. Can some one let me know when he starts to talking about the deep blue game wishing the next 38 minutes?

Leave a Reply

Your email address will not be published. Required fields are marked *