episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Baiting actions.
https://karpathy.ai/lexicap/0012-large.html#00:11:38.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So there's information abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:11:39.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
don't talk about general games, information abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:11:41.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which is the abstraction of what chance does.
https://karpathy.ai/lexicap/0012-large.html#00:11:44.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And this would be the cards in the case of poker.
https://karpathy.ai/lexicap/0012-large.html#00:11:47.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And then there's action abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:11:50.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which is abstracting the actions of the actual players,
https://karpathy.ai/lexicap/0012-large.html#00:11:52.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which would be bets in the case of poker.
https://karpathy.ai/lexicap/0012-large.html#00:11:57.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yourself and the other players?
https://karpathy.ai/lexicap/0012-large.html#00:11:59.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yes, yourself and other players.
https://karpathy.ai/lexicap/0012-large.html#00:12:01.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And for information abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:12:03.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we were completely automated.
https://karpathy.ai/lexicap/0012-large.html#00:12:08.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So these are algorithms,
https://karpathy.ai/lexicap/0012-large.html#00:12:11.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but they do what we call potential aware abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:12:13.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
where we don't just look at the value of the hand,
https://karpathy.ai/lexicap/0012-large.html#00:12:16.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but also how it might materialize
https://karpathy.ai/lexicap/0012-large.html#00:12:19.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
into good or bad hands over time.
https://karpathy.ai/lexicap/0012-large.html#00:12:20.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And it's a certain kind of bottom up process
https://karpathy.ai/lexicap/0012-large.html#00:12:22.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
with integer programming there and clustering
https://karpathy.ai/lexicap/0012-large.html#00:12:25.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and various aspects, how do you build this abstraction?
https://karpathy.ai/lexicap/0012-large.html#00:12:27.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And then in the action abstraction,
https://karpathy.ai/lexicap/0012-large.html#00:12:31.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
there it's largely based on how humans and other AIs
https://karpathy.ai/lexicap/0012-large.html#00:12:34.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
have played this game in the past.
https://karpathy.ai/lexicap/0012-large.html#00:12:40.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But in the beginning,
https://karpathy.ai/lexicap/0012-large.html#00:12:42.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we actually used an automated action abstraction technology,
https://karpathy.ai/lexicap/0012-large.html#00:12:43.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which is provably convergent
https://karpathy.ai/lexicap/0012-large.html#00:12:47.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that it finds the optimal combination of bet sizes,
https://karpathy.ai/lexicap/0012-large.html#00:12:51.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but it's not very scalable.
https://karpathy.ai/lexicap/0012-large.html#00:12:54.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So we couldn't use it for the whole game,
https://karpathy.ai/lexicap/0012-large.html#00:12:55.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but we use it for the first couple of betting actions.
https://karpathy.ai/lexicap/0012-large.html#00:12:57.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what's more important, the strength of the hand,
https://karpathy.ai/lexicap/0012-large.html#00:12:59.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
so the information abstraction or the how you play them,
https://karpathy.ai/lexicap/0012-large.html#00:13:03.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the actions, does it, you know,
https://karpathy.ai/lexicap/0012-large.html#00:13:09.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the romanticized notion again,
https://karpathy.ai/lexicap/0012-large.html#00:13:11.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
is that it doesn't matter what hands you have,
https://karpathy.ai/lexicap/0012-large.html#00:13:13.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that the actions, the betting may be the way you win
https://karpathy.ai/lexicap/0012-large.html#00:13:15.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
no matter what hands you have.
https://karpathy.ai/lexicap/0012-large.html#00:13:19.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, so that's why you have to play a lot of hands
https://karpathy.ai/lexicap/0012-large.html#00:13:20.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
so that the role of luck gets smaller.
https://karpathy.ai/lexicap/0012-large.html#00:13:23.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So you could otherwise get lucky and get some good hands
https://karpathy.ai/lexicap/0012-large.html#00:13:26.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and then you're gonna win the match.
https://karpathy.ai/lexicap/0012-large.html#00:13:29.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Even with thousands of hands, you can get lucky
https://karpathy.ai/lexicap/0012-large.html#00:13:31.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because there's so much variance
https://karpathy.ai/lexicap/0012-large.html#00:13:35.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in No Limit Texas Holden because if we both go all in,
https://karpathy.ai/lexicap/0012-large.html#00:13:36.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it's a huge stack of variance, so there are these
https://karpathy.ai/lexicap/0012-large.html#00:13:40.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
massive swings in No Limit Texas Holden.
https://karpathy.ai/lexicap/0012-large.html#00:13:43.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's why you have to play not just thousands,
https://karpathy.ai/lexicap/0012-large.html#00:13:47.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but over 100,000 hands to get statistical significance.
https://karpathy.ai/lexicap/0012-large.html#00:13:50.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So let me ask another way this question.
https://karpathy.ai/lexicap/0012-large.html#00:13:55.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
If you didn't even look at your hands,
https://karpathy.ai/lexicap/0012-large.html#00:13:57.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but they didn't know that, the opponents didn't know that,
https://karpathy.ai/lexicap/0012-large.html#00:14:02.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
how well would you be able to do?
https://karpathy.ai/lexicap/0012-large.html#00:14:04.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Oh, that's a good question.
https://karpathy.ai/lexicap/0012-large.html#00:14:06.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
There's actually, I heard this story
https://karpathy.ai/lexicap/0012-large.html#00:14:07.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that there's this Norwegian female poker player
https://karpathy.ai/lexicap/0012-large.html#00:14:09.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
called Annette Oberstad who's actually won a tournament
https://karpathy.ai/lexicap/0012-large.html#00:14:11.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
by doing exactly that, but that would be extremely rare.
https://karpathy.ai/lexicap/0012-large.html#00:14:15.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So you cannot really play well that way.
https://karpathy.ai/lexicap/0012-large.html#00:14:18.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Okay, so the hands do have some role to play, okay.
https://karpathy.ai/lexicap/0012-large.html#00:14:23.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So Labradus does not use, as far as I understand,
https://karpathy.ai/lexicap/0012-large.html#00:14:27.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
they use learning methods, deep learning.
https://karpathy.ai/lexicap/0012-large.html#00:14:33.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Is there room for learning in,
https://karpathy.ai/lexicap/0012-large.html#00:14:35.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
there's no reason why Labradus doesn't combine
https://karpathy.ai/lexicap/0012-large.html#00:14:40.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
with an AlphaGo type approach for estimating
https://karpathy.ai/lexicap/0012-large.html#00:14:44.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the quality for function estimator.
https://karpathy.ai/lexicap/0012-large.html#00:14:46.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
What are your thoughts on this,
https://karpathy.ai/lexicap/0012-large.html#00:14:49.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
maybe as compared to another algorithm
https://karpathy.ai/lexicap/0012-large.html#00:14:52.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which I'm not that familiar with, DeepStack,
https://karpathy.ai/lexicap/0012-large.html#00:14:54.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the engine that does use deep learning,
https://karpathy.ai/lexicap/0012-large.html#00:14:56.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that it's unclear how well it does,
https://karpathy.ai/lexicap/0012-large.html#00:14:59.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but nevertheless uses deep learning.
https://karpathy.ai/lexicap/0012-large.html#00:15:01.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what are your thoughts about learning methods
https://karpathy.ai/lexicap/0012-large.html#00:15:03.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to aid in the way that Labradus plays in the game of poker?
https://karpathy.ai/lexicap/0012-large.html#00:15:05.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, so as you said,
https://karpathy.ai/lexicap/0012-large.html#00:15:09.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Labradus did not use learning methods
https://karpathy.ai/lexicap/0012-large.html#00:15:10.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and played very well without them.
https://karpathy.ai/lexicap/0012-large.html#00:15:13.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Since then, we have actually, actually here,
https://karpathy.ai/lexicap/0012-large.html#00:15:15.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we have a couple of papers on things
https://karpathy.ai/lexicap/0012-large.html#00:15:17.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that do use learning techniques.
https://karpathy.ai/lexicap/0012-large.html#00:15:20.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Excellent.
https://karpathy.ai/lexicap/0012-large.html#00:15:22.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And deep learning in particular.
https://karpathy.ai/lexicap/0012-large.html#00:15:24.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And sort of the way you're talking about
https://karpathy.ai/lexicap/0012-large.html#00:15:26.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
where it's learning an evaluation function,
https://karpathy.ai/lexicap/0012-large.html#00:15:29.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but in imperfect information games,
https://karpathy.ai/lexicap/0012-large.html#00:15:33.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
unlike let's say in Go or now also in chess and shogi,
https://karpathy.ai/lexicap/0012-large.html#00:15:37.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it's not sufficient to learn an evaluation for a state
https://karpathy.ai/lexicap/0012-large.html#00:15:42.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because the value of an information set
https://karpathy.ai/lexicap/0012-large.html#00:15:47.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
depends not only on the exact state,
https://karpathy.ai/lexicap/0012-large.html#00:15:52.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but it also depends on both players beliefs.
https://karpathy.ai/lexicap/0012-large.html#00:15:55.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Like if I have a bad hand,
https://karpathy.ai/lexicap/0012-large.html#00:15:59.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I'm much better off if the opponent thinks I have a good hand
https://karpathy.ai/lexicap/0012-large.html#00:16:01.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and vice versa.
https://karpathy.ai/lexicap/0012-large.html#00:16:04.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
If I have a good hand,
https://karpathy.ai/lexicap/0012-large.html#00:16:05.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I'm much better off if the opponent believes
https://karpathy.ai/lexicap/0012-large.html#00:16:06.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I have a bad hand.
https://karpathy.ai/lexicap/0012-large.html#00:16:09.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So the value of a state is not just a function of the cards.
https://karpathy.ai/lexicap/0012-large.html#00:16:11.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
It depends on, if you will, the path of play,
https://karpathy.ai/lexicap/0012-large.html#00:16:15.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but only to the extent that it's captured
https://karpathy.ai/lexicap/0012-large.html#00:16:19.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in the belief distributions.
https://karpathy.ai/lexicap/0012-large.html#00:16:22.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's why it's not as simple
https://karpathy.ai/lexicap/0012-large.html#00:16:23.720