episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
so it's not a question of the human forgetting
https://karpathy.ai/lexicap/0012-large.html#00:06:51.300
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and the AI somehow trying to get advantage
https://karpathy.ai/lexicap/0012-large.html#00:06:53.460
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
of better memory.
https://karpathy.ai/lexicap/0012-large.html#00:06:55.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what was that like?
https://karpathy.ai/lexicap/0012-large.html#00:06:56.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I mean, that was an incredible accomplishment.
https://karpathy.ai/lexicap/0012-large.html#00:06:57.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what did it feel like before the event?
https://karpathy.ai/lexicap/0012-large.html#00:06:59.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Did you have doubt, hope?
https://karpathy.ai/lexicap/0012-large.html#00:07:02.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Where was your confidence at?
https://karpathy.ai/lexicap/0012-large.html#00:07:05.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, that's great.
https://karpathy.ai/lexicap/0012-large.html#00:07:08.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So great question.
https://karpathy.ai/lexicap/0012-large.html#00:07:09.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So 18 months earlier, I had organized a similar brains
https://karpathy.ai/lexicap/0012-large.html#00:07:10.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
versus AI competition with a previous AI called Cloudyco
https://karpathy.ai/lexicap/0012-large.html#00:07:14.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and we couldn't beat the humans.
https://karpathy.ai/lexicap/0012-large.html#00:07:17.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So this time around, it was only 18 months later.
https://karpathy.ai/lexicap/0012-large.html#00:07:20.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And I knew that this new AI, Libratus, was way stronger,
https://karpathy.ai/lexicap/0012-large.html#00:07:23.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but it's hard to say how you'll do against the top humans
https://karpathy.ai/lexicap/0012-large.html#00:07:27.820
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
before you try.
https://karpathy.ai/lexicap/0012-large.html#00:07:31.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So I thought we had about a 50, 50 shot.
https://karpathy.ai/lexicap/0012-large.html#00:07:32.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And the international betting sites put us
https://karpathy.ai/lexicap/0012-large.html#00:07:35.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
as a four to one or five to one underdog.
https://karpathy.ai/lexicap/0012-large.html#00:07:38.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So it's kind of interesting that people really believe
https://karpathy.ai/lexicap/0012-large.html#00:07:41.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in people and over AI, not just people.
https://karpathy.ai/lexicap/0012-large.html#00:07:44.700
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
People don't just over believe in themselves,
https://karpathy.ai/lexicap/0012-large.html#00:07:48.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but they have overconfidence in other people as well
https://karpathy.ai/lexicap/0012-large.html#00:07:50.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
compared to the performance of AI.
https://karpathy.ai/lexicap/0012-large.html#00:07:53.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And yeah, so we were a four to one or five to one underdog.
https://karpathy.ai/lexicap/0012-large.html#00:07:55.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And even after three days of beating the humans in a row,
https://karpathy.ai/lexicap/0012-large.html#00:07:59.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we were still 50, 50 on the international betting sites.
https://karpathy.ai/lexicap/0012-large.html#00:08:02.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Do you think there's something special and magical
https://karpathy.ai/lexicap/0012-large.html#00:08:06.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
about poker and the way people think about it,
https://karpathy.ai/lexicap/0012-large.html#00:08:09.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in the sense you have,
https://karpathy.ai/lexicap/0012-large.html#00:08:12.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I mean, even in chess, there's no Hollywood movies.
https://karpathy.ai/lexicap/0012-large.html#00:08:14.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Poker is the star of many movies.
https://karpathy.ai/lexicap/0012-large.html#00:08:17.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And there's this feeling that certain human facial
https://karpathy.ai/lexicap/0012-large.html#00:08:21.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
expressions and body language, eye movement,
https://karpathy.ai/lexicap/0012-large.html#00:08:26.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
all these tells are critical to poker.
https://karpathy.ai/lexicap/0012-large.html#00:08:30.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Like you can look into somebody's soul
https://karpathy.ai/lexicap/0012-large.html#00:08:33.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and understand their betting strategy and so on.
https://karpathy.ai/lexicap/0012-large.html#00:08:35.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's probably why, possibly,
https://karpathy.ai/lexicap/0012-large.html#00:08:37.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
do you think that is why people have a confidence
https://karpathy.ai/lexicap/0012-large.html#00:08:41.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that humans will outperform?
https://karpathy.ai/lexicap/0012-large.html#00:08:43.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Because AI systems cannot, in this construct,
https://karpathy.ai/lexicap/0012-large.html#00:08:45.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
perceive these kinds of tells.
https://karpathy.ai/lexicap/0012-large.html#00:08:48.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
They're only looking at betting patterns
https://karpathy.ai/lexicap/0012-large.html#00:08:51.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and nothing else, betting patterns and statistics.
https://karpathy.ai/lexicap/0012-large.html#00:08:53.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what's more important to you
https://karpathy.ai/lexicap/0012-large.html#00:08:58.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
if you step back on human players, human versus human?
https://karpathy.ai/lexicap/0012-large.html#00:09:02.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
What's the role of these tells,
https://karpathy.ai/lexicap/0012-large.html#00:09:06.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
of these ideas that we romanticize?
https://karpathy.ai/lexicap/0012-large.html#00:09:08.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, so I'll split it into two parts.
https://karpathy.ai/lexicap/0012-large.html#00:09:11.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So one is why do humans trust humans more than AI
https://karpathy.ai/lexicap/0012-large.html#00:09:15.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and have overconfidence in humans?
https://karpathy.ai/lexicap/0012-large.html#00:09:20.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I think that's not really related to the tell question.
https://karpathy.ai/lexicap/0012-large.html#00:09:22.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
It's just that they've seen these top players,
https://karpathy.ai/lexicap/0012-large.html#00:09:25.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
how good they are, and they're really fantastic.
https://karpathy.ai/lexicap/0012-large.html#00:09:28.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So it's just hard to believe that an AI could beat them.
https://karpathy.ai/lexicap/0012-large.html#00:09:31.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So I think that's where that comes from.
https://karpathy.ai/lexicap/0012-large.html#00:09:36.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And that's actually maybe a more general lesson about AI.
https://karpathy.ai/lexicap/0012-large.html#00:09:37.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
That until you've seen it overperform a human,
https://karpathy.ai/lexicap/0012-large.html#00:09:40.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it's hard to believe that it could.
https://karpathy.ai/lexicap/0012-large.html#00:09:43.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But then the tells, a lot of these top players,
https://karpathy.ai/lexicap/0012-large.html#00:09:45.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
they're so good at hiding tells
https://karpathy.ai/lexicap/0012-large.html#00:09:50.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that among the top players,
https://karpathy.ai/lexicap/0012-large.html#00:09:52.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it's actually not really worth it
https://karpathy.ai/lexicap/0012-large.html#00:09:56.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
for them to invest a lot of effort
https://karpathy.ai/lexicap/0012-large.html#00:09:59.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
trying to find tells in each other
https://karpathy.ai/lexicap/0012-large.html#00:10:01.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because they're so good at hiding them.
https://karpathy.ai/lexicap/0012-large.html#00:10:03.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So yes, at the kind of Friday evening game,
https://karpathy.ai/lexicap/0012-large.html#00:10:05.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
tells are gonna be a huge thing.
https://karpathy.ai/lexicap/0012-large.html#00:10:09.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
You can read other people.
https://karpathy.ai/lexicap/0012-large.html#00:10:11.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And if you're a good reader,
https://karpathy.ai/lexicap/0012-large.html#00:10:13.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you'll read them like an open book.
https://karpathy.ai/lexicap/0012-large.html#00:10:14.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But at the top levels of poker now,
https://karpathy.ai/lexicap/0012-large.html#00:10:16.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the tells become a much smaller and smaller aspect
https://karpathy.ai/lexicap/0012-large.html#00:10:18.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
of the game as you go to the top levels.
https://karpathy.ai/lexicap/0012-large.html#00:10:21.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
The amount of strategies, the amount of possible actions
https://karpathy.ai/lexicap/0012-large.html#00:10:24.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
is very large, 10 to the power of 100 plus.
https://karpathy.ai/lexicap/0012-large.html#00:10:28.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So there has to be some, I've read a few of the papers
https://karpathy.ai/lexicap/0012-large.html#00:10:35.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
related, it has to form some abstractions
https://karpathy.ai/lexicap/0012-large.html#00:10:37.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
of various hands and actions.
https://karpathy.ai/lexicap/0012-large.html#00:10:42.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what kind of abstractions are effective
https://karpathy.ai/lexicap/0012-large.html#00:10:44.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
for the game of poker?
https://karpathy.ai/lexicap/0012-large.html#00:10:47.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, so you're exactly right.
https://karpathy.ai/lexicap/0012-large.html#00:10:49.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So when you go from a game tree that's 10 to the 161,
https://karpathy.ai/lexicap/0012-large.html#00:10:50.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
especially in an imperfect information game,
https://karpathy.ai/lexicap/0012-large.html#00:10:55.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it's way too large to solve directly,
https://karpathy.ai/lexicap/0012-large.html#00:10:58.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
even with our fastest equilibrium finding algorithms.
https://karpathy.ai/lexicap/0012-large.html#00:11:00.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So you wanna abstract it first.
https://karpathy.ai/lexicap/0012-large.html#00:11:03.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And abstraction in games is much trickier
https://karpathy.ai/lexicap/0012-large.html#00:11:07.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
than abstraction in MDPs or other single agent settings.
https://karpathy.ai/lexicap/0012-large.html#00:11:10.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Because you have these abstraction pathologies
https://karpathy.ai/lexicap/0012-large.html#00:11:15.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that if I have a finer grained abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:11:17.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the strategy that I can get from that for the real game
https://karpathy.ai/lexicap/0012-large.html#00:11:19.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
might actually be worse than the strategy
https://karpathy.ai/lexicap/0012-large.html#00:11:23.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I can get from the coarse grained abstraction.
https://karpathy.ai/lexicap/0012-large.html#00:11:25.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So you have to be very careful.
https://karpathy.ai/lexicap/0012-large.html#00:11:27.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Now the kinds of abstractions, just to zoom out,
https://karpathy.ai/lexicap/0012-large.html#00:11:28.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we're talking about, there's the hands abstractions
https://karpathy.ai/lexicap/0012-large.html#00:11:31.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and then there's betting strategies.
https://karpathy.ai/lexicap/0012-large.html#00:11:34.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, betting actions, yeah.
https://karpathy.ai/lexicap/0012-large.html#00:11:37.280