episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
do they make sense, are they explainable,
https://karpathy.ai/lexicap/0012-large.html#00:52:21.800
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
or do they suffer from the same problems
https://karpathy.ai/lexicap/0012-large.html#00:52:23.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
as neural networks do?
https://karpathy.ai/lexicap/0012-large.html#00:52:25.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's a good question.
https://karpathy.ai/lexicap/0012-large.html#00:52:27.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I would say a little bit yes and no.
https://karpathy.ai/lexicap/0012-large.html#00:52:28.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And what I mean by that is that
https://karpathy.ai/lexicap/0012-large.html#00:52:31.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
these game theoretic strategies,
https://karpathy.ai/lexicap/0012-large.html#00:52:34.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
let's say, Nash equilibrium,
https://karpathy.ai/lexicap/0012-large.html#00:52:36.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
it has provable properties.
https://karpathy.ai/lexicap/0012-large.html#00:52:38.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So it's unlike, let's say, deep learning
https://karpathy.ai/lexicap/0012-large.html#00:52:40.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
where you kind of cross your fingers,
https://karpathy.ai/lexicap/0012-large.html#00:52:42.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
hopefully it'll work.
https://karpathy.ai/lexicap/0012-large.html#00:52:44.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And then after the fact, when you have the weights,
https://karpathy.ai/lexicap/0012-large.html#00:52:45.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
you're still crossing your fingers,
https://karpathy.ai/lexicap/0012-large.html#00:52:47.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and I hope it will work.
https://karpathy.ai/lexicap/0012-large.html#00:52:48.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Here, you know that the solution quality is there.
https://karpathy.ai/lexicap/0012-large.html#00:52:51.160
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
There's provable solution quality guarantees.
https://karpathy.ai/lexicap/0012-large.html#00:52:55.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Now, that doesn't necessarily mean
https://karpathy.ai/lexicap/0012-large.html#00:52:58.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that the strategies are human understandable.
https://karpathy.ai/lexicap/0012-large.html#00:53:00.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
That's a whole other problem.
https://karpathy.ai/lexicap/0012-large.html#00:53:03.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So I think that deep learning and computational game theory
https://karpathy.ai/lexicap/0012-large.html#00:53:04.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
are in the same boat in that sense,
https://karpathy.ai/lexicap/0012-large.html#00:53:08.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that both are difficult to understand.
https://karpathy.ai/lexicap/0012-large.html#00:53:10.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But at least the game theoretic techniques,
https://karpathy.ai/lexicap/0012-large.html#00:53:13.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
they have these guarantees of solution quality.
https://karpathy.ai/lexicap/0012-large.html#00:53:15.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So do you see business operations, strategic operations,
https://karpathy.ai/lexicap/0012-large.html#00:53:19.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
or even military in the future being
https://karpathy.ai/lexicap/0012-large.html#00:53:22.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
at least the strong candidates
https://karpathy.ai/lexicap/0012-large.html#00:53:26.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
being proposed by automated systems?
https://karpathy.ai/lexicap/0012-large.html#00:53:28.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Do you see that?
https://karpathy.ai/lexicap/0012-large.html#00:53:32.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, I do, I do.
https://karpathy.ai/lexicap/0012-large.html#00:53:34.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
But that's more of a belief than a substantiated fact.
https://karpathy.ai/lexicap/0012-large.html#00:53:35.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Depending on where you land in optimism or pessimism,
https://karpathy.ai/lexicap/0012-large.html#00:53:39.640
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
that's a really, to me, that's an exciting future,
https://karpathy.ai/lexicap/0012-large.html#00:53:42.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
especially if there's provable things
https://karpathy.ai/lexicap/0012-large.html#00:53:45.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in terms of optimality.
https://karpathy.ai/lexicap/0012-large.html#00:53:48.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So looking into the future,
https://karpathy.ai/lexicap/0012-large.html#00:53:50.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
there's a few folks worried about the,
https://karpathy.ai/lexicap/0012-large.html#00:53:54.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
especially you look at the game of poker,
https://karpathy.ai/lexicap/0012-large.html#00:53:58.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
which is probably one of the last benchmarks
https://karpathy.ai/lexicap/0012-large.html#00:54:01.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in terms of games being solved.
https://karpathy.ai/lexicap/0012-large.html#00:54:03.360
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
They worry about the future
https://karpathy.ai/lexicap/0012-large.html#00:54:05.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and the existential threats of artificial intelligence.
https://karpathy.ai/lexicap/0012-large.html#00:54:07.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So the negative impact in whatever form on society.
https://karpathy.ai/lexicap/0012-large.html#00:54:10.520
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Is that something that concerns you as much,
https://karpathy.ai/lexicap/0012-large.html#00:54:13.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
or are you more optimistic about the positive impacts of AI?
https://karpathy.ai/lexicap/0012-large.html#00:54:17.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Oh, I am much more optimistic about the positive impacts.
https://karpathy.ai/lexicap/0012-large.html#00:54:21.600
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So just in my own work, what we've done so far,
https://karpathy.ai/lexicap/0012-large.html#00:54:24.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
we run the nationwide kidney exchange.
https://karpathy.ai/lexicap/0012-large.html#00:54:27.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Hundreds of people are walking around alive today,
https://karpathy.ai/lexicap/0012-large.html#00:54:29.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
who would it be?
https://karpathy.ai/lexicap/0012-large.html#00:54:32.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And it's increased employment.
https://karpathy.ai/lexicap/0012-large.html#00:54:34.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
You have a lot of people now running kidney exchanges
https://karpathy.ai/lexicap/0012-large.html#00:54:36.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and at the transplant centers,
https://karpathy.ai/lexicap/0012-large.html#00:54:39.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
interacting with the kidney exchange.
https://karpathy.ai/lexicap/0012-large.html#00:54:42.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
You have extra surgeons, nurses, anesthesiologists,
https://karpathy.ai/lexicap/0012-large.html#00:54:45.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
hospitals, all of that.
https://karpathy.ai/lexicap/0012-large.html#00:54:49.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So employment is increasing from that
https://karpathy.ai/lexicap/0012-large.html#00:54:51.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
and the world is becoming a better place.
https://karpathy.ai/lexicap/0012-large.html#00:54:53.560
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Another example is combinatorial sourcing auctions.
https://karpathy.ai/lexicap/0012-large.html#00:54:55.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
We did 800 large scale combinatorial sourcing auctions
https://karpathy.ai/lexicap/0012-large.html#00:54:59.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
from 2001 to 2010 in a previous startup of mine
https://karpathy.ai/lexicap/0012-large.html#00:55:04.040
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
called CombineNet.
https://karpathy.ai/lexicap/0012-large.html#00:55:08.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And we increased the supply chain efficiency
https://karpathy.ai/lexicap/0012-large.html#00:55:09.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
on that $60 billion of spend by 12.6%.
https://karpathy.ai/lexicap/0012-large.html#00:55:13.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's over $6 billion of efficiency improvement
https://karpathy.ai/lexicap/0012-large.html#00:55:18.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
in the world.
https://karpathy.ai/lexicap/0012-large.html#00:55:21.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And this is not like shifting value
https://karpathy.ai/lexicap/0012-large.html#00:55:22.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
from somebody to somebody else,
https://karpathy.ai/lexicap/0012-large.html#00:55:23.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
just efficiency improvement, like in trucking,
https://karpathy.ai/lexicap/0012-large.html#00:55:25.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
less empty driving, so there's less waste,
https://karpathy.ai/lexicap/0012-large.html#00:55:28.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
less carbon footprint and so on.
https://karpathy.ai/lexicap/0012-large.html#00:55:31.120
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So a huge positive impact in the near term,
https://karpathy.ai/lexicap/0012-large.html#00:55:33.440
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
but sort of to stay in it for a little longer,
https://karpathy.ai/lexicap/0012-large.html#00:55:36.720
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because I think game theory has a role to play here.
https://karpathy.ai/lexicap/0012-large.html#00:55:40.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Oh, let me actually come back on that as one thing.
https://karpathy.ai/lexicap/0012-large.html#00:55:43.080
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I think AI is also going to make the world much safer.
https://karpathy.ai/lexicap/0012-large.html#00:55:45.320
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So that's another aspect that often gets overlooked.
https://karpathy.ai/lexicap/0012-large.html#00:55:49.400
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Well, let me ask this question.
https://karpathy.ai/lexicap/0012-large.html#00:55:53.760
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Maybe you can speak to the safer.
https://karpathy.ai/lexicap/0012-large.html#00:55:54.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So I talked to Max Tegmark and Stuart Russell,
https://karpathy.ai/lexicap/0012-large.html#00:55:56.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
who are very concerned about existential threats of AI.
https://karpathy.ai/lexicap/0012-large.html#00:55:59.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And often the concern is about value misalignment.
https://karpathy.ai/lexicap/0012-large.html#00:56:02.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So AI systems basically working,
https://karpathy.ai/lexicap/0012-large.html#00:56:06.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
operating towards goals that are not the same
https://karpathy.ai/lexicap/0012-large.html#00:56:11.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
as human civilization, human beings.
https://karpathy.ai/lexicap/0012-large.html#00:56:14.680
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
So it seems like game theory has a role to play there
https://karpathy.ai/lexicap/0012-large.html#00:56:17.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
to make sure the values are aligned with human beings.
https://karpathy.ai/lexicap/0012-large.html#00:56:24.200
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I don't know if that's how you think about it.
https://karpathy.ai/lexicap/0012-large.html#00:56:27.880
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
If not, how do you think AI might help with this problem?
https://karpathy.ai/lexicap/0012-large.html#00:56:29.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
How do you think AI might make the world safer?
https://karpathy.ai/lexicap/0012-large.html#00:56:35.280
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Yeah, I think this value misalignment
https://karpathy.ai/lexicap/0012-large.html#00:56:39.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
is a fairly theoretical worry.
https://karpathy.ai/lexicap/0012-large.html#00:56:43.000
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
And I haven't really seen it in,
https://karpathy.ai/lexicap/0012-large.html#00:56:46.480
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
because I do a lot of real applications.
https://karpathy.ai/lexicap/0012-large.html#00:56:49.960
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
I don't see it anywhere.
https://karpathy.ai/lexicap/0012-large.html#00:56:51.840
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
The closest I've seen it
https://karpathy.ai/lexicap/0012-large.html#00:56:53.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
was the following type of mental exercise really,
https://karpathy.ai/lexicap/0012-large.html#00:56:55.240
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
where I had this argument in the late eighties
https://karpathy.ai/lexicap/0012-large.html#00:56:57.920
Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
when we were building
https://karpathy.ai/lexicap/0012-large.html#00:57:00.720