episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and where have you seen it done successfully in your work?
https://karpathy.ai/lexicap/0012-large.html#00:34:24.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I'm not sure the goal really is modeling humans.
https://karpathy.ai/lexicap/0012-large.html#00:34:27.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Like for example, if I'm playing a zero sum game,
https://karpathy.ai/lexicap/0012-large.html#00:34:33.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I don't really care that the opponent
https://karpathy.ai/lexicap/0012-large.html#00:34:36.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
is actually following my model of rational behavior,
https://karpathy.ai/lexicap/0012-large.html#00:34:39.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because if they're not, that's even better for me.
https://karpathy.ai/lexicap/0012-large.html#00:34:42.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Right, so see with the opponents in games,
https://karpathy.ai/lexicap/0012-large.html#00:34:46.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the prerequisite is that you formalize
https://karpathy.ai/lexicap/0012-large.html#00:34:51.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the interaction in some way
https://karpathy.ai/lexicap/0012-large.html#00:34:56.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that can be amenable to analysis.
https://karpathy.ai/lexicap/0012-large.html#00:34:57.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And you've done this amazing work with mechanism design,
https://karpathy.ai/lexicap/0012-large.html#00:35:01.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
designing games that have certain outcomes.
https://karpathy.ai/lexicap/0012-large.html#00:35:04.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But, so I'll tell you an example
https://karpathy.ai/lexicap/0012-large.html#00:35:10.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
from my world of autonomous vehicles, right?
https://karpathy.ai/lexicap/0012-large.html#00:35:12.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
We're studying pedestrians,
https://karpathy.ai/lexicap/0012-large.html#00:35:15.460
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and pedestrians and cars negotiate
https://karpathy.ai/lexicap/0012-large.html#00:35:17.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in this nonverbal communication.
https://karpathy.ai/lexicap/0012-large.html#00:35:20.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
There's this weird game dance of tension
https://karpathy.ai/lexicap/0012-large.html#00:35:22.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
where pedestrians are basically saying,
https://karpathy.ai/lexicap/0012-large.html#00:35:25.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I trust that you won't kill me,
https://karpathy.ai/lexicap/0012-large.html#00:35:27.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and so as a jaywalker, I will step onto the road
https://karpathy.ai/lexicap/0012-large.html#00:35:28.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
even though I'm breaking the law, and there's this tension.
https://karpathy.ai/lexicap/0012-large.html#00:35:31.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And the question is, we really don't know
https://karpathy.ai/lexicap/0012-large.html#00:35:34.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
how to model that well in trying to model intent.
https://karpathy.ai/lexicap/0012-large.html#00:35:36.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And so people sometimes bring up ideas
https://karpathy.ai/lexicap/0012-large.html#00:35:40.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
of game theory and so on.
https://karpathy.ai/lexicap/0012-large.html#00:35:43.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Do you think that aspect of human behavior
https://karpathy.ai/lexicap/0012-large.html#00:35:44.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
can use these kinds of imperfect information approaches,
https://karpathy.ai/lexicap/0012-large.html#00:35:49.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
modeling, how do you start to attack a problem like that
https://karpathy.ai/lexicap/0012-large.html#00:35:53.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
when you don't even know how to design the game
https://karpathy.ai/lexicap/0012-large.html#00:35:57.860
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to describe the situation in order to solve it?
https://karpathy.ai/lexicap/0012-large.html#00:36:00.940
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Okay, so I haven't really thought about jaywalking,
https://karpathy.ai/lexicap/0012-large.html#00:36:04.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but one thing that I think could be a good application
https://karpathy.ai/lexicap/0012-large.html#00:36:06.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in autonomous vehicles is the following.
https://karpathy.ai/lexicap/0012-large.html#00:36:10.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So let's say that you have fleets of autonomous cars
https://karpathy.ai/lexicap/0012-large.html#00:36:13.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
operating by different companies.
https://karpathy.ai/lexicap/0012-large.html#00:36:16.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So maybe here's the Waymo fleet and here's the Uber fleet.
https://karpathy.ai/lexicap/0012-large.html#00:36:18.340
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
If you think about the rules of the road,
https://karpathy.ai/lexicap/0012-large.html#00:36:22.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
they define certain legal rules,
https://karpathy.ai/lexicap/0012-large.html#00:36:24.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but that still leaves a huge strategy space open.
https://karpathy.ai/lexicap/0012-large.html#00:36:26.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Like as a simple example, when cars merge,
https://karpathy.ai/lexicap/0012-large.html#00:36:30.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
how humans merge, they slow down and look at each other
https://karpathy.ai/lexicap/0012-large.html#00:36:32.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and try to merge.
https://karpathy.ai/lexicap/0012-large.html#00:36:36.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Wouldn't it be better if these situations
https://karpathy.ai/lexicap/0012-large.html#00:36:39.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
would already be prenegotiated
https://karpathy.ai/lexicap/0012-large.html#00:36:40.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
so we can actually merge at full speed
https://karpathy.ai/lexicap/0012-large.html#00:36:43.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and we know that this is the situation,
https://karpathy.ai/lexicap/0012-large.html#00:36:45.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
this is how we do it, and it's all gonna be faster.
https://karpathy.ai/lexicap/0012-large.html#00:36:47.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But there are way too many situations to negotiate manually.
https://karpathy.ai/lexicap/0012-large.html#00:36:50.540
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So you could use automated negotiation,
https://karpathy.ai/lexicap/0012-large.html#00:36:54.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
this is the idea at least,
https://karpathy.ai/lexicap/0012-large.html#00:36:56.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you could use automated negotiation
https://karpathy.ai/lexicap/0012-large.html#00:36:57.780
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to negotiate all of these situations
https://karpathy.ai/lexicap/0012-large.html#00:36:59.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
or many of them in advance.
https://karpathy.ai/lexicap/0012-large.html#00:37:02.060
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And of course it might be that,
https://karpathy.ai/lexicap/0012-large.html#00:37:04.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
hey, maybe you're not gonna always let me go first.
https://karpathy.ai/lexicap/0012-large.html#00:37:05.460
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Maybe you said, okay, well, in these situations,
https://karpathy.ai/lexicap/0012-large.html#00:37:09.180
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I'll let you go first, but in exchange,
https://karpathy.ai/lexicap/0012-large.html#00:37:11.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you're gonna give me too much,
https://karpathy.ai/lexicap/0012-large.html#00:37:13.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you're gonna let me go first in this situation.
https://karpathy.ai/lexicap/0012-large.html#00:37:14.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So it's this huge combinatorial negotiation.
https://karpathy.ai/lexicap/0012-large.html#00:37:17.260
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And do you think there's room in that example of merging
https://karpathy.ai/lexicap/0012-large.html#00:37:20.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to model this whole situation
https://karpathy.ai/lexicap/0012-large.html#00:37:24.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
as an imperfect information game
https://karpathy.ai/lexicap/0012-large.html#00:37:25.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
or do you really want to consider it to be a perfect?
https://karpathy.ai/lexicap/0012-large.html#00:37:27.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
No, that's a good question, yeah.
https://karpathy.ai/lexicap/0012-large.html#00:37:30.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
That's a good question.
https://karpathy.ai/lexicap/0012-large.html#00:37:32.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Do you pay the price of assuming
https://karpathy.ai/lexicap/0012-large.html#00:37:33.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that you don't know everything?
https://karpathy.ai/lexicap/0012-large.html#00:37:37.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, I don't know.
https://karpathy.ai/lexicap/0012-large.html#00:37:39.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
It's certainly much easier.
https://karpathy.ai/lexicap/0012-large.html#00:37:40.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Games with perfect information are much easier.
https://karpathy.ai/lexicap/0012-large.html#00:37:42.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So if you can't get away with it, you should.
https://karpathy.ai/lexicap/0012-large.html#00:37:45.060
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But if the real situation is of imperfect information,
https://karpathy.ai/lexicap/0012-large.html#00:37:49.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
then you're gonna have to deal with imperfect information.
https://karpathy.ai/lexicap/0012-large.html#00:37:52.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Great, so what lessons have you learned
https://karpathy.ai/lexicap/0012-large.html#00:37:55.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
the Annual Computer Poker Competition?
https://karpathy.ai/lexicap/0012-large.html#00:37:58.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
An incredible accomplishment of AI.
https://karpathy.ai/lexicap/0012-large.html#00:38:00.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
You look at the history of Deep Blue, AlphaGo,
https://karpathy.ai/lexicap/0012-large.html#00:38:03.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
these kind of moments when AI stepped up
https://karpathy.ai/lexicap/0012-large.html#00:38:07.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in an engineering effort and a scientific effort combined
https://karpathy.ai/lexicap/0012-large.html#00:38:10.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to beat the best of human players.
https://karpathy.ai/lexicap/0012-large.html#00:38:13.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So what do you take away from this whole experience?
https://karpathy.ai/lexicap/0012-large.html#00:38:16.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
What have you learned about designing AI systems
https://karpathy.ai/lexicap/0012-large.html#00:38:19.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that play these kinds of games?
https://karpathy.ai/lexicap/0012-large.html#00:38:22.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And what does that mean for AI in general,
https://karpathy.ai/lexicap/0012-large.html#00:38:23.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
for the future of AI development?
https://karpathy.ai/lexicap/0012-large.html#00:38:28.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, so that's a good question.
https://karpathy.ai/lexicap/0012-large.html#00:38:30.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So there's so much to say about it.
https://karpathy.ai/lexicap/0012-large.html#00:38:32.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I do like this type of performance oriented research.
https://karpathy.ai/lexicap/0012-large.html#00:38:35.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Although in my group, we go all the way from like idea
https://karpathy.ai/lexicap/0012-large.html#00:38:39.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to theory, to experiments, to big system building,
https://karpathy.ai/lexicap/0012-large.html#00:38:42.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to commercialization, so we span that spectrum.
https://karpathy.ai/lexicap/0012-large.html#00:38:44.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But I think that in a lot of situations in AI,
https://karpathy.ai/lexicap/0012-large.html#00:38:47.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you really have to build the big systems
https://karpathy.ai/lexicap/0012-large.html#00:38:51.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and evaluate them at scale
https://karpathy.ai/lexicap/0012-large.html#00:38:53.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
before you know what works and doesn't.
https://karpathy.ai/lexicap/0012-large.html#00:38:55.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And we've seen that in the computational
https://karpathy.ai/lexicap/0012-large.html#00:38:57.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
game theory community, that there are a lot of techniques
https://karpathy.ai/lexicap/0012-large.html#00:39:00.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that look good in the small,
https://karpathy.ai/lexicap/0012-large.html#00:39:02.880