Noam Brown: AI vs Humans in Poker and Games of Strategic Negotiation
嘉宾：Noam Brown
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录19 章节
§01Transcript · 00:00–07:48§02Transcript · 07:51–15:31§03Transcript · 15:33–22:37§04Transcript · 22:40–29:55§05Transcript · 29:59–37:37§06Transcript · 37:39–45:32§07Transcript · 45:34–52:59§08Transcript · 53:01–1:00:39§09Transcript · 1:00:41–1:09:02§10Transcript · 1:09:05–1:17:01§11Transcript · 1:17:03–1:24:33§12Transcript · 1:24:35–1:32:34§13Transcript · 1:32:36–1:40:28§14Transcript · 1:40:30–1:48:31§15Transcript · 1:48:35–1:56:34§16Transcript · 1:56:39–2:05:57§17Transcript · 2:06:00–2:14:41§18Transcript · 2:14:44–2:22:50§19Transcript · 2:22:52–2:29:15§01Transcript · 00:00–07:48
A lot of people were saying like,
oh, this whole idea of game theory, it's just nonsense.
And if you really want to make money,
you got to like look into the other person's eyes
and read their soul and figure out what cards they have.
But what happened was where we played our bot
against four top heads up, no limit,
hold them poker players.
And the bot wasn't trying to adapt to them.
It wasn't trying to exploit them.
It wasn't trying to do these mind games.
It was just trying to approximate the Nash equilibrium
and it crushed them.
The following is a conversation with Noah Brown,
research scientist at FAIR,
Facebook AI research group at Meta AI.
He co-created the first AI system
that achieved superhuman level performance
in no limit Texas hold them, both heads up and multiplayer.
And now recently he co-created an AI system
that can strategically out negotiate humans
using natural language
in a popular board game called diplomacy,
which is a war game that emphasizes negotiation.
This is Alex Friedman podcast to support it.
Please check out our sponsors in the description.
And now dear friends, here's Noam Brown.
You've been a lead on three amazing AI projects.
So we've got Libratus that solved
or at least achieved human level performance
on no limit Texas hold them poker
with two players, heads up.
You got Pleribus that solved no limit Texas hold them poker
with six players.
And just now you have Cicero.
These are all names of systems that solved
or achieved human level performance
on the game of diplomacy,
which for people who don't know
is a popular strategy board game.
It was loved by JFK, John F. Kennedy and Henry Kissinger
and many other big famous people in the decades since.
So let's talk about poker and diplomacy today.
First poker, what is the game of no limit Texas hold them?
And how is it different from chess?
- Well, no limit Texas hold them poker
is the most popular variant of poker in the world.
So, you know, you go to a casino,
you play sit down at the poker table.
The game that you're playing is no limit Texas hold them.
If you watch movies about poker,
like casino Royale or rounders,
the game that they're playing
is no limit Texas hold them poker.
Now it's very different from limit hold them
in that you can bet any amount of chips that you want.
And so the stakes escalate really quickly.
You start out with like one or $2 in the pot.
And then by the end of the hand,
you've got like $1,000 in there maybe.
- So the option to increase the number very aggressively
and very quickly is always there.
- Right, the no limit aspect is there's no limits
to how much you can bet.
You know, in limit hold them,
there's like $2 in the pot,
you can only bet like $2.
But if you got $10,000 in front of you,
you're always welcome to put $10,000 into the pot.
- So I've got a chance to hang out with Phil Helmuth
who plays all these different variants of poker.
And correct me if I'm wrong,
but it seems like no limit rewards crazy
versus the other ones rewards
more kind of calculated strategy.
Or no, because you're sort of looking
from an analytic perspective,
is strategy also rewarded in no limit Texas hold them?
- I think both variants reward strategy.
But I think what's different about no limit hold them
is it's much easier to get jumpy.
You know, you go in there thinking you're gonna play
for like $100 or something.
And suddenly there's like, you know, $1,000 in the pot.
A lot of people can't handle that.
- Can you define jumpy?
- When you're playing poker,
you always want to choose the action
that's going to maximize your expected value.
It's kind of like with investing, right?
Like if you're ever in a situation
where the amount of money that's at stake
is going to have a material impact on your life,
then you're gonna play in a more risk averse style.
You know, if somebody makes a huge bet,
you're gonna, if you're playing no limit hold them
and somebody makes a huge bet,
there might come a point where you're like,
this is too much money for me to handle.
Like I can't risk this amount.
And that's what throws a lot of people off.
So that's the big difference I think
between no limit and limit.
- What about on the action side
when you're actually making that big bet?
That's what I mean by crazy.
I was trying to refer to the technical,
the technical term of crazy,
meaning use the big jump in the bet
to completely throw off the other person
in terms of their ability to reason optimally.
- I think that's right.
I think one of the key strategies in poker
is to put the other person into an uncomfortable position.
And if you're doing that, then you're playing poker well.
And there's a lot of opportunities to do that
in no limit hold them.
You know, you can have like $50 in there,
you throw in a thousand dollar bet,
and you know, that's sometimes if you do it right,
it puts the other person in a really tough spot.
Now it's also possible that you make huge mistakes that way.
And so it's really easy to lose a lot of money
in no limit hold them if you don't know what you're doing.
But there's a lot of upside potential too.
- So when you build systems, AI systems
that play these games, we'll talk about poker,
we'll talk about diplomacy.
Are you drawn in in part by the beauty of the game itself,
AI aside, or is it to you primarily a fascinating
problem set for the AI to solve?
- I'm drawn in by the beauty of the game.
When I started playing poker when I was in high school,
and the idea to me that there is a correct,
an objectively correct way of playing poker.
And if you could figure out what that is,
then you're making unlimited money basically.
That's like a really fascinating concept to me.
And so I was fascinated by the strategy of poker,
even when I was like 16 years old.
It wasn't until like much later
that I actually worked on poker AIs.
- So there was a sense that you can solve poker,
like in the way you can solve chess, for example,
or checkers, I believe checkers got solved, right?
- Yeah, checkers is completely solved.
- Optimal strategy. - Optimal strategy.
It's impossible to beat the AI.
- Yeah, and so in that same way,
you could technically solve chess.
- You could solve chess, you could solve poker.
- You could solve poker.
- So this gets into the concept of a Nash equilibrium.
- So it is a Nash equilibrium.
- Okay, so in any finite two-player zero-sum game,
there is an optimal strategy that if you play it,
you are guaranteed to not lose an expectation
no matter what your opponent does.
And this is kind of a radical concept to a lot of people,
but it's true in chess, it's true in poker,
it's true in any finite two-player zero-sum game.
And to give some intuition for this,
you can think of rock, paper, scissors.
In rock, paper, scissors, if you randomly choose
between throwing rock, paper, and scissors
with equal probability,
then no matter what your opponent does,
you are not going to lose an expectation.
You're not going to lose an expectation in the long run.
Now, the same is true for poker.
There exists some strategy,
some really complicated strategy,
that if you play that,
you are guaranteed to not lose money in the long run.
And I should say, this is for two-player poker.
Six-player poker is a different story.
- Yeah, it's a beautiful giant mess.
When you say in expectation,
you're guaranteed not to lose in expectation.
What does in expectation mean?
- Poker's a very high variance game.
So you're going to have hands where you win,
you're going to have hands with your lose.
Even if you're playing the perfect strategy,
you can't guarantee that you're going to win
every single hand.
But if you play for long enough,
then you are guaranteed to at least break even,
and in practice, probably win.
- So that's in expectation,
the size of your stack, generally speaking.
Now, that doesn't include anything
about the fact that you can go broke.
It doesn't include any of those kinds
of normal real-world limitations.
You're talking in a theoretical world.
What about the zero-sum aspect?
How big of a constraint is that?
How big of a constraint is finite?
- So finite's not a huge constraint.
§02Transcript · 07:51–15:31
So I mean, most games that you play are finite in size.
It's also true, actually,
that there exists this perfect strategy
in many infinite games as well.
Technically, the game has to be compact.
There are some edge cases
where you don't have a Nash equilibrium
in a two-player zero-sum game.
So you can think of a game where,
if we're playing a game where whoever names
the bigger number is the winner,
there's no Nash equilibrium to that game.
- 17, okay. - Yeah, exactly.
18, but you beat.
- You win again.
You're good at this.
- I've played a lot of games.
- Okay, so that's, and then the zero-sum aspect.
The zero-sum. - Zero-sum aspect.
So there exists a Nash equilibrium
in non-two-player zero-sum games as well.
And by the way, just to clarify what I mean
by two-player zero-sum, I mean there's two players,
and whatever one player wins, the other player loses.
So if we're playing poker and I win $50,
that means that you're losing $50.
Now, outside of two-player zero-sum games,
there still exists Nash equilibria,
but they're not as meaningful.
Because you can think of a game like Risk.
If everybody else on the board decides
to team up against you and take you out,
there's no perfect strategy you can play
that's gonna guarantee that you win there.
There's just nothing you can do.
So outside of two-player zero-sum games,
there's no guarantee that you're going to win
by playing a Nash equilibrium.
- Have you ever tried to model in
the other aspects of the game,
which is like the pleasure you draw from playing the game?
And then if you're a professional poker player,
if you're exciting, even if you lose,
the money you would get from the attention you get
to the sponsors and all that kind of stuff,
is that, that would be a fun thing to model in.
Or is that make it sort of super complex
to include the human factor in its full complexity?
- I think you bring up a couple of good points there.
So I think a lot of professional poker players,
I mean, they get a huge amount of money,
not from actually playing poker,
but from the sponsorships and having a personality
that people want to tune in and watch.
That's a big way to make a name for yourself in poker.
- I just wonder from an AI perspective,
if you create, and we'll talk about this more,
maybe AI system that also talks trash
and all that kind of stuff,
that that becomes part of the function to maximize.
So it's not just optimal poker play.
Maybe sometimes you want to be chaotic.
Maybe sometimes you want to be suboptimal
and you lose the chaos.
And maybe sometimes you want to be overly aggressive
because the audience loves that.
That'd be fascinating.
- I think what you're getting at here
is that there's a difference between making an AI
that wins a game and an AI that's fun to play with.
- Yeah.
- Yeah.
- Or fun to watch.
So those are all different things,
fun to play with and fun to watch.
- Yeah, and I think, I've heard talks from game designers
and they say people that work on AI
for actual recreational games that people play.
And they say, yeah, there's a big difference
between trying to make an AI that actually wins.
And you look at a game like "Civilization",
the way that the AIs play is not optimal for trying to win.
They're playing a different game.
They're trying to have personalities.
They're trying to be fun and engaging.
And that makes for a better game.
- Yeah.
- And we also talk about NPCs.
We just talked to Todd Howard,
who is the creator of "Fallout" and the "Elder Scrolls" series
and "Starfield", the new game coming out.
And the creator of what I think is the greatest game
of all time, which is "Skyrim" and the NPCs there.
The AI that governs that whole game is very interesting,
but the NPCs also are super interesting.
And considering what language models might do to NPCs
in an open world RPG role-playing game,
it's super exciting.
- Yeah, honestly, I think this is one
of the first applications where we're going to see
real consumer interaction with large language models.
I guess "Elder Scrolls VI" is in development now.
They're probably pretty close to finishing it,
but I would not be surprised at all if "Elder Scrolls VII"
was using large language models for their NPCs.
- No, they're not.
I mean, I'm not saying anything.
I'm not saying anything.
- Okay, this is me speculating, not you.
- No, but they're just releasing the "Starfield" game.
They do one game at a time.
- Yeah.
- And so whatever it is, whenever the date is,
I don't know what the date is, calm down,
but it would be, I don't know, like 2024, '25, '26.
So it's actually very possible
that would include language models.
- I was listening to this talk by a gaming executive
when I was in grad school.
And one of the questions that a person in the audience asked
is why are all these games so focused
on fighting and killing?
And the person responded that it's just so much harder
to make an AI that can talk with you and cooperate with you
than it is to make an AI that can fight you.
And I think once this technology develops further
and you can reach a point where like
not every single line of dialogue has to be scripted,
it unlocks a lot of potential for new kinds of games,
like much more like positive interactions
that are not so focused on fighting.
And I'm really looking forward to that.
- It might not be positive.
It might be just drama.
So you'll be in like a "Call of Duty" game
and instead of doing the shooting,
you'll just be hanging out and like arguing with an AI
about like passive aggressive.
And then you won't be able to sleep that night.
You have to return and continue the argument
that you were emotionally hurt.
I mean, yeah, I think that's actually an exciting world.
Whatever is the drama, the chaos that we love,
the push and pull of human connection,
I think it's possible to do that in the video game world.
And I think you could be messier
and make more mistakes in a video game world,
which is why it would be a nice place.
And also it doesn't have as deep
of a real psychological impact
because inside video games,
it's kind of understood that you're in a not a real world.
So whatever crazy stuff AI does,
we have some flexibility to play.
Just like with a game of diplomacy, it's a game.
This is not real geopolitics, not real war.
It's a game.
So you can have a little bit of fun, a little bit of chaos.
Okay, back to "Nashville Coyote Pair."
How do we find the Nash equilibrium?
- All right, so there's different ways
to find a Nash equilibrium.
So the way that we do it
is with this process called self-play.
Basically, we have this algorithm
that starts by playing totally randomly,
and it learns how to play the game
by playing against itself.
So it will start playing the game totally randomly,
and then if it's playing poker,
it'll eventually get to the end of the game and make $50.
And then it will review all of the decisions
that it made along the way and say,
what would have happened
if I had chosen this other action instead?
If I had raised here instead of called,
what would the other player have done?
And because it's playing against a copy of itself,
it's able to do that counterfactual reasoning.
So it can say, okay, well, if I took this action
and the other person takes this action,
and then I take this action,
and eventually I make $150 instead of 50.
And so it updates the regret value for that action.
Regret is basically like how much does it regret
having not played that action in the past?
And when it encounters that same situation again,
it's going to pick actions that have higher regret
with higher probability.
Now, it'll just keep simulating the games this way.
It'll keep accumulating regrets for different situations.
And in the long run,
if you pick actions that have higher regret
with higher probability in the correct way,
it's proven to converge to a Nash equilibrium.
- Even for super complex games?
Even for imperfect information games?
- It's true for all games.
It's true for chess, it's true for poker.
It's particularly useful for poker.
§03Transcript · 15:33–22:37
- So this is the method
of counterfactual regret minimization?
- This is counterfactual regret minimization.
- That doesn't have to do with self-play,
it has to do with just any,
if you follow this kind of process, self-play or not,
you will be able to arrive at an optimal set of actions.
- So this counterfactual regret minimization
is a kind of self-play.
It's a principled kind of self-play
that's proven to converge to Nash equilibria,
even in imperfect information games.
Now you can have other forms of self-play
and people use other forms of self-play
for perfect information games,
where you have more flexibility,
the algorithm doesn't have to be as theoretically sound
in order to converge to that class of games
because it's a simpler setting.
- Sure, so I kind of, in my brain,
the word self-play has mapped to neural networks,
but we're speaking something bigger
than just neural networks.
It could be anything.
The self-play mechanism is just the mechanism
of a system playing itself.
- Exactly, yeah.
Self-play is not tied specifically to neural nets,
it's a kind of reinforcement learning, basically.
And I would also say this process of trying to reason,
oh, what would the value have been
if I had taken this other action instead?
This is very similar to how humans learn
to play a game like poker.
You probably played poker before
and with your friends, you probably ask,
oh, what do you have called me if I raise there?
That's a person trying to do the same kind of learning
from a counterfactual that the AI is doing.
- Okay, and if you do that at scale,
you're gonna be able to learn an optimal policy.
- Yeah, now where the neural nets come in,
I said, okay, if it's in that situation again,
then it will choose the action that has high regret.
Now, the problem is that poker is such a huge game.
I think No Limit Texas Hold'em,
the version that we were playing,
has 10 to the 161 different decision points,
which is more than the number of atoms
in the universe squared.
- That's heads up?
- That's heads up, yeah.
- 10 to the 161, you said?
- Yeah, I mean, it depends on the number of chips
that you have, the stacks and everything,
but the version that we were playing was 10 to the 161.
- Which I assume would be a somewhat simplified version
anyway, 'cause I bet there's some step function
you had for bets.
- Oh, no, no, no, I'm saying we played the full game.
You can bet whatever amount you want.
Now, the bot maybe was constrained
in what it considered for bet sizes,
but the person on the other side
could bet whatever they wanted.
- Yeah, I mean, 161 plus or minus 10 doesn't matter.
- Yeah, yeah.
And so the way neural nets help out here is,
you know, you don't have to run into the same exact situation
'cause that's never gonna happen again.
The odds of you running into the same exact situation
are pretty slim, but if you run into a similar situation,
then you can generalize from other states
that you've been in that kind of look like that one,
and you can say like, well, these other situations,
I had high regret for this action,
and so maybe I should play that action here as well.
- Which is the more complex game?
Chess or poker or go or poker?
Do you know?
- That is a controversial question.
- Okay.
- I'm gonna-
- It's like somebody's screaming on Reddit right now.
It depends on which subreddit you're on.
Is it chess or is it poker?
- I'm sure like David Silver's
gonna get really angry at me.
- Yeah.
- I'll say, I'm gonna say poker actually,
and I think for a couple of reasons.
- They're not here to defend themselves.
- So first of all,
you have the imperfect information aspect.
And so it's, we can go into that,
but like once you introduce imperfect information,
things get much more complicated.
- So we should say,
maybe you can describe what is seen to the players,
what is not seen in the game of Texas Hold 'Em.
- Yeah, so Texas Hold 'Em,
you get two cards face down that only you see.
And so that's the hidden information of the game.
The other players also all get two cards face down
that only they see.
And so you have to kind of, as you're playing,
reason about like, okay, what do they think I have?
What do they have?
What do they think I think they have?
That kind of stuff.
And that's kind of where bluffing comes into play, right?
Because the fact that you can bluff,
the fact that you can bet with a bad hand and still win,
is because they don't know what your cards are.
- Right.
- And that's the key difference
between a perfect information game like poker,
sorry, like chess and go,
and imperfect information games like poker.
- This is what trash talk looks like.
The implied statement is,
the game I solved is much tougher.
But yeah, so when you're playing,
I'm just gonna do random questions here.
So when you're playing your opponent
under imperfect information,
is there some degree to which you're trying
to estimate the range of hands that they have?
Or is that not part of the algorithm?
So what are the different approaches
to the imperfect information game?
- So the key thing to understand
about why imperfect information makes things difficult,
is that you have to worry not just about
which actions to play,
but the probability that you're gonna play those actions.
So you think about rock, paper, scissors, for example,
rock, paper, scissors is an imperfect information game.
- Right.
- Because you don't know what I'm about to throw.
- I do, but yeah, usually not, yeah.
- Yeah, and so you can't just say like,
I'm just gonna throw a rock every single time,
because the other person's gonna figure that out
and notice a pattern,
and then suddenly you're gonna start losing.
And so you don't just have to figure out
like which action to play,
you have to figure out the probability that you play it.
And really importantly, the value of an action
depends on the probability that you're gonna play it.
So if you're playing rock every single time,
that value is really low.
But if you're never playing rock,
you play rock like 1% of the time,
then suddenly the other person
is probably gonna be throwing scissors.
And when you throw a rock,
the value of that action is gonna be really high.
Now you take that to poker,
what that means is the value of bluffing, for example,
if you're the kind of person that never bluffs
and you have this reputation as somebody that never bluffs,
and suddenly you bluff,
there's a really good chance that that bluff is gonna work
and you're gonna make a lot of money.
On the other hand, if you got a reputation,
like if they seen you play for a long time and they see,
oh, you're the kind of person that's bluffing all the time,
when you bluff, they're not gonna buy it
and they're gonna call you down
and you're gonna lose a lot of money.
And that, finding that balance
of how often you should be bluffing
is the key challenge of a game of poker.
And you contrast that with a game like chess,
it doesn't matter if you're opening with the queen's gambit
10% of the time or 100% of the time,
the value, the expected value is the same.
So that's why we need these algorithms that understand,
not just we have to figure out what actions are good,
but the probabilities,
we need to get the exact probabilities correct.
And that's actually when we created the bot Libratus,
Libratus means balanced
because the algorithm that we designed
was designed to find that right balance
of how often it should play each action.
- The balance of how often in the key sort of branching
is the bluff or not to bluff.
Is that a good crude simplification
of the major decision in poker?
- It's a good simplification,
I think that's like the main tension,
but it's not just how often to bluff or not to bluff,
it's like, how often should you bet in general?
How often should you, what kind of bet should you make?
Should you bet big or should you bet small?
And with which hands?
And so this is where the idea of a range comes from,
§04Transcript · 22:40–29:55
because when you're bluffing with a particular hand
in a particular spot,
you don't want there to be a pattern
for the other person to pick up on.
You don't want them to figure out,
oh, whenever this person is in this spot,
they're always bluffing.
And so you have to reason about,
okay, would I also bet with a good hand in this spot?
You wanna be unpredictable.
So you have to think about what would I do
if I had this different set of cards?
- Is there explicit estimation of like a theory of mind
that the other person has about you,
or is that just a emergent thing that happens?
- The way that the bots handle it,
that are really successful,
they have an explicit theory of mind.
So they're explicitly reasoning about
what's the common knowledge belief?
What do you think I have?
What do I think you have?
What do you think I think you have?
It's explicitly reasoning about that.
- Is there multiple yous there?
So maybe that's jumping ahead to six players,
but is there a stickiness to the person?
So it's an iterative game.
You're playing the same person.
There's a stickiness to that, right?
You're gathering information as you play.
It's not every hand is a new hand.
Is there a continuation in terms of estimating
what kind of player I'm facing here?
- That's a good question.
So you could approach the game that way.
The way that the bots do it,
they don't, and the way that humans approach it also,
expert human players,
the way they approach it is to basically assume
that you know my strategy.
So I'm going to try to pick a strategy
where even if I were to play it for 10,000 hands
and you could figure out exactly what it was,
you still wouldn't be able to beat it.
Basically what that means is I'm trying
to approximate the Nash equilibrium.
I'm trying to be perfectly balanced
because if I'm playing the Nash equilibrium,
even if you know what my strategy is,
like I said, I'm still unbeatable in expectation.
So that's what the bot aims for.
And that's actually what a lot
of expert poker players aim for as well,
to start by playing the Nash equilibrium.
And then maybe if they spot weaknesses
in the way you're playing,
then they can deviate a little bit
to take advantage of that.
- They aim to be unbeatable in expectation.
Okay.
So who's the greatest poker player of all time
and why is it Phil Hellmuth?
So this is for Phil.
So he's known at least in part
for maybe playing suboptimally
and he still wins a lot.
It's a bit chaotic.
So maybe can you speak from an AI perspective
about the genius of his madness
or the madness of his genius?
So playing suboptimally, playing chaotically
as a way to make it hard to pin down
about what your strategy is.
- So, okay.
The thing that I should explain first of all
with like Nash equilibrium,
it doesn't mean that it's predictable.
The whole point of it is that
you're trying to be unpredictable.
Now, I think when somebody like Phil Hellmuth
might be really successful,
is not in being unpredictable,
but in being able to take advantage
of the other player and figure out
where they're being predictable
or guiding the other player into thinking
that you have certain weaknesses
and then understanding how they're going
to change their behavior.
They're going to deviate from a Nash equilibrium style
of play to try to take advantage
of those perceived weaknesses
and then counter exploit them.
So you kind of get into the mind games there.
- So you think about these heads up poker
as a dance between two agents,
I guess, are you playing the cards
or are you playing the player?
- So this gets down to a big argument
in the poker community and the academic community.
For a long time, there was this debate
of like what's called GTO,
game theory optimal poker or exploitative play.
And up until about like 2017,
when we did the Labradors match,
I think actually exploitative play had the advantage.
A lot of people were saying like,
oh, this whole idea of game theory, it's just nonsense.
And if you really want to make money,
you got to like look into the other person's eyes
and read their soul and figure out what cards they have.
But what happened was people started adopting
the game theory optimal strategy
and they were making good money.
And they weren't trying to adapt so much to the other player.
They were just trying to play the Nash equilibrium.
And then what really solidified it, I think,
was the Labradors match,
where we played our bot against four top heads up,
no limit hold 'em poker players.
And the bot wasn't trying to adapt to them.
It wasn't trying to exploit them.
It wasn't trying to do these mind games.
It was just trying to approximate the Nash equilibrium
and it crushed them.
I think, you know, we were playing for $50, $100 blinds.
And over the course of about 120,000 hands,
it made close to $2 million.
- 120,000 hands.
- 120,000 hands.
- Against humans.
- Yeah, and this was fake money to be clear.
So there was real money at stake.
There was $200,000.
- First of all, all money is fake.
But that's a different conversation.
We give it meaning.
It's a phenomena that gets meaning
from our complex psychology as a human civilization.
It's emerging from the collective intelligence
of the human species.
But that's not what you mean.
You mean like there's literally,
you can't buy stuff with it.
Okay, can you actually step back
and take me through that competition?
- Yeah, okay.
So when I was in grad school,
there was this thing called
the annual computer poker competition,
where every year all the different research labs
that were working on AI for poker would get together.
They would make a bot.
They would play them against each other.
And we made a bot that actually won the 2014 competition,
the 2016 competition.
And so we decided we're gonna take this bot, build on it,
and play against real top professional
heads up no limit Texas Hold'em poker players.
So we invited four of the world's best players
in this specialty.
And we challenged them to 120,000 hands of poker
over the course of 20 days.
And we had $200,000 in prize money at stake,
where it would basically be divided among them,
depending on how well they did relative to each other.
So we wanted to have some incentive
for them to play their best.
- Did you have a confidence 2014, 16,
that this is even possible?
How much doubt was there?
- So we did a competition actually in 2015,
where we also played against professional poker players
and the bot lost by a pretty sizable margin actually.
Now there were some big improvements from 2015 to 2017.
And so-
- Can you speak to the improvements?
Is it computational in nature?
Is it the algorithm, the methods?
- It was really an algorithmic approach.
That was the difference.
So 2015, it was much more focused on trying to come up
with a strategy upfront,
like trying to solve the entire game of poker,
like, and then just have a lookup table
where you're saying like, "Oh, I'm in this situation.
"What's the strategy?"
The approach that we took in 2017
was much more search-based.
It was trying to say, "Okay, well, let me in real time
"try to compute a much better strategy
"than what I had pre-computed
"by playing against myself during self-play."
- What is the search space for poker?
What are you searching over?
What's that look like?
There's different actions like raising, calling.
Yeah, what are the actions?
Is it just a search over actions?
§05Transcript · 29:59–37:37
- So in a game like chess, the search is like,
"Okay, I'm in this chess position
"and I can move these different pieces
"and see where things end up."
In poker, what you're searching over
is the actions that you can take for your hand,
the probabilities that you take those actions,
and then also the probabilities that you take other actions
with other hands that you might have.
And that's kind of like hard to wrap your head around.
Like, why are you searching over these other hands
that you might have and trying to figure out
what you would do with those hands?
And the idea is, again, you wanna always be balanced
and unpredictable.
And so if you're a search algorithm that's saying like,
"Oh, I want to raise with this hand."
Well, in order to know whether that's a good action,
like let's say it's a bluff.
Let's say you have a bad hand and you're saying like,
"Oh, I think I should be betting here
"with this really bad hand and bluffing."
Well, that's only a good action
if you're also betting with a strong hand.
Otherwise, it's an obvious bluff.
- So if your action in some sense
maximizes your unpredictability,
so that action could be mapped by your opponent
to a lot of different hands, then that's a good action.
- Basically what you wanna do is put your opponent
into a tough spot.
So you want them to always have some doubt,
like, "Should I call here?
"Should I fold here?"
And if you are raising in the appropriate balance
between bluffs and good hands,
then you're putting them into that tough spot.
And so that's what we're trying to do.
We're always trying to search for a strategy
that would put the opponent into a difficult position.
- Can you give a metric that you're trying to maximize
or minimize?
Does this have to do with the regret thing
that we're talking about in terms of putting your opponent
in a maximally tough spot?
- Yeah, ultimately what you're trying to maximize
is your expected winnings, your expected value,
the amount of money that you're gonna walk away from,
assuming that your opponent was playing optimally
in response.
So you're gonna assume that your opponent
is also playing as well as possible
a Nash equilibrium approach,
because if they're not,
then you're just gonna make more money.
Anything that deviates, by definition,
the Nash equilibrium is the strategy
that does the best in expectation.
And so if you're deviating from that,
then you're just, they're gonna lose money.
And since it's a two player zero sum game,
that means you're gonna make money.
- So there's not an explicit, like objective function
that maximizes the toughness of the spot they're put in.
You're always, this is from like a self play
reinforcement learning perspective.
You're just trying to maximize winnings
and the rest is implicit.
- That's right, yeah.
So what we're actually trying to maximize
is the expected value,
given that the opponent is playing optimally
in response to us.
Now in practice, what that ends up looking like
is it's putting the opponent into difficult situations
where there's no obvious decision to be made.
- So the system doesn't know anything
about the difficulty of the situation.
- Not at all, doesn't care.
- Okay, in my head it was getting excited
whenever I was making the other, the opponent sweat.
Okay, so you're, in 2015, you didn't do as well.
So what's the journey from that to a system
that in your mind could have a chance?
- So 2015, we got, we beat pretty badly
and we actually learned a lot from that competition.
And in particular, what became clear to me
is that the way the humans were approaching the game
was very different from how the bot was approaching the game.
The bot would not be doing search.
It would just be trying to compute,
it would do like months of self play.
It would just be playing against itself for months,
but then when it's actually playing the game,
it would just act instantly.
And the humans, when they're in a tough spot,
they would sit there and think for sometimes
even like five minutes about whether they're gonna call
or fold a hand.
And it became clear to me that that's,
there's a good chance that that's what's missing
from our bot.
So I actually did some initial experiments
to try to figure out how much of a difference
does this actually make?
And the difference was huge.
- As a signal to the human player,
how long you took to think?
- No, no, no, I'm not saying
that there were any timing tells.
I was saying when the human,
like the bot would always act instantly.
It wouldn't try to come up with a better strategy
in real time over what it had pre-computed during training.
Whereas the human, like they have all this intuition
about how to play, but they're also in real time
leveraging their ability to think,
just to search, to plan,
and coming up with an even better strategy
than what their intuition would say.
- So you're saying that there's,
you're doing, that's what you mean by
you're doing search also.
You have an intuition and search on top of that,
looking for a better solution.
- Yeah, that's what I mean by search.
That instead of acting instantly,
a neural net usually gives you a response
in like a hundred milliseconds or something.
It depends on the size of the net.
But if you can leverage extra computational resources,
you can possibly get a much better outcome.
And we did some experiments
in small scale versions of poker.
And what we found was that
if you do a little bit of search, even just a little bit,
it was the equivalent of making your,
you know, your pre-computed strategy.
Like you can kind of think of it as your neural net,
a thousand times bigger.
It was just a little bit of search.
And it just like blew away all of the research
that we had been working on
and trying to like scale up this like pre-computed solution.
It was dwarfed by the benefit that we got from search.
- Can you just linger on what you mean by search here?
You're searching over a space of actions
for your hand and for other hands.
How are you selecting the other hands to search over?
- So yeah. - Randomly?
- No, it's all the other hands that you could have.
So when you're playing no limit Texas hold 'em,
you've got two face down cards.
And so that's 52 choose two, 1,326 different combinations.
Now that's actually a little bit lower
because there's face up cards in the middle.
And so you can eliminate those as well.
But you're looking at like around a thousand
different possible hands that you can have.
And so when we're doing, when the bot's doing search,
it's thinking explicitly,
there are these thousand different hands that I could have.
There are these thousand different hands that you could have.
Let me try to figure out what would be a better strategy
than what I've pre-computed for these hands and your hands.
- Okay, so that search, how do you fuse that
with what the neural net is telling you
or what the train system is telling you?
- Yeah, so you kind of like,
where the train system comes in is the value at the end.
So there's, you only look so far ahead.
You look like maybe one round ahead.
So if you're on the flop,
you're looking to the start of the turn.
And at that point you can use the pre-computed solution
to figure out what's the value here of this strategy.
- Is it of a single action, essentially in that spot?
You're getting a value or is it the value
of the entire series of actions?
- Well, it's kind of both
because you're trying to maximize the value
for the hand that you have,
but in the process, in order to maximize the value
of the hand that you have,
you have to figure out what would I be doing
with all these other hands as well.
- Okay, but are you in the search
or was going to the end of the game?
- In Libratus, we did.
So we only use search starting on the turn.
And then we searched all the way to the end of the game.
- The turn, the river.
Can we take it through the terminology?
- Yeah, there's four rounds of poker.
So there's the pre-flop, the flop, the turn and the river.
And so we would start doing search halfway through the game.
Now the first half of the game, that was all pre-computed.
It would just act instantly.
And then when it got to the halfway point,
then it would always search to the end of the game.
§06Transcript · 37:39–45:32
Now we later improved this
so it wouldn't have to search all the way
to the end of the game.
It would actually search just a few moves ahead.
But that came later and that drastically reduced
the amount of computational resources that we needed.
- But the moves, 'cause you can keep betting
on top of each other.
That's what you mean by moves.
So like that's where you don't just get one bet
per turn of poker.
You can have multiple arbitrary number of bets, right?
- Right, I'm trying to think like,
I'm gonna bet and then what are you gonna do in response?
Are you gonna raise me or are you gonna call?
And then if you raise, what should I do?
So it's reasoning about that whole process
up until the end of the game in the case of Libratus.
- So for Libratus, what's the most number of re-raises
have you ever seen?
- You probably cap out at like five or something
because at that point you're basically all in.
- I mean, is there like interesting patterns like that
that you've seen that the game does?
Like you'll have like alpha zero doing way more sacrifices
than humans usually do.
Is there something like Libratus was constantly re-raising
or something like that that you've noticed?
- There was something really interesting
that we observed with Libratus.
So humans, when they're playing poker,
they usually size their bets relative
to the size of the pot.
So if the pot has $100 in there,
maybe you bet like $75 or somewhere around there,
somewhere between like 50 and $100.
And with Libratus, we gave it the option
to basically bet whatever it wanted.
It was actually really easy for us to say like,
oh, if you want, you can bet like 10 times the pot.
And we didn't think it would actually do that.
It was just like, why not give it the option?
And then during the competition,
it actually started doing this.
And by the way, this was like a very last minute decision
on our part to add this option.
And so we did not think the bot would do this.
And I was actually kind of worried
when it did start to do this, like, oh, is this a problem?
Like humans don't do this.
Like is it screwing up?
But it would put the humans into really difficult spots
when it would do that.
Because you could imagine like you have the second best hand
that's possible given the board.
And you're thinking like,
oh, you're in a really great spot here.
And suddenly the bot bets $20,000 into a $1,000 pot.
And it's basically saying like,
I have the best hand or I'm bluffing.
And you having the second best hand,
like now you get a really tough choice to make.
And so the humans would sometimes think like five
or 10 minutes about like, what do you do?
Should I call? Should I fold?
And when I saw the humans like really struggling
with that decision, like that's when I realized like,
oh, actually this is maybe a good thing to do after all.
- And of course the system is a no that it's making,
again, like we said, that it's putting them in a tough spot.
It's just, that's part of the optimal,
the game theory optimal.
- Right, from the bot's perspective,
it's just doing the thing
that's going to make it the most money.
And the fact that it's putting the humans
in a difficult spot, like that's just, you know,
a side effect of that.
And this was, I think the one thing,
I mean, there were a few things
that the humans walked away from,
but this was the number one thing
that the humans walked away from the competition saying like,
we need to start doing this.
And now these overbets, what are called overbets,
have become really common in high-level poker play.
- Have you ever talked to like somebody like Daniel
DeGranio about this?
He seems to be a student of the game.
- I did actually have a conversation
with Daniel DeGranio once, yeah.
I was visiting the Isle of Man
to talk to poker stars about AI.
And Daniel DeGranio was there
when we had dinner together with some other people.
And yeah, he was really interested in it.
He mentioned that he was like, you know,
excited about like learning from these AIs.
- So he wasn't scared, he was excited.
- He was excited.
And he honestly, he wanted to play against the bot.
He thought he had a decent chance of beating it.
I think, you know, this was like several years ago
when I think it was like not as clear to everybody
that, you know, the AIs were taking over.
I think now people recognize that like
if you're playing against a bot,
there's like no chance that you have in a game like poker.
- So consistently the bots will win.
The bots have heads up and in other variants too.
So multi, six player Texas hold 'em,
no limit Texas hold 'em, the bots win?
- Yeah, that's the case.
So I think there is some debate about like,
is it true for every single variant of poker?
I think for every single variant of poker,
if somebody really put in the effort,
they can make an AI that would beat all humans at it.
We've focused on the most popular variants.
So heads up, no limit Texas hold 'em.
And then we followed that up with six player poker as well,
where we managed to make a bot
that beat expert human players.
And I think even there now,
it's pretty clear that humans don't stand a chance.
- See, I would love to hook up an AI system
that looks at EEG, like how,
like actually tries to optimize the toughness of the spot
it puts a human in.
And I would love to see how different is that
from the game theory optimal.
So you try to maximize the heart rate of the human player,
like the freaking out over a long period of time.
I wonder if there's going to be different strategies
that emerge that are close in terms of effectiveness.
'Cause something tells me you could still be
achieve superhuman level performance
by just making people sweat.
- I feel like that there's a good chance
that that is the case, yeah.
If you're able to see like,
that it's like a decent proxy for score, right?
And this is actually like the common poker wisdom
where they're teaching players, before there were bots,
and they were trying to teach people how to play poker.
They would say like, the key to the game
is to put your opponent into difficult spots.
It's a good estimate for
if you're making the right decision.
- So what else can you say about
the fundamental role of search in poker?
And maybe if you can also relate it to chess and go
in these games,
what's the role of search to solving these games?
- Yeah, I think a lot of people under,
this is true for the general public
and I think it's true for the AI community.
A lot of people underestimate the importance of search
for these kinds of game AI results.
An example of this is TD Gammon that came out in 1992.
This was the first real instance of a neural net
being used in a game AI.
It's a landmark achievement.
It was actually the inspiration for AlphaZero
and it used search.
It used two-ply search to figure out its next move.
You got Deep Blue.
There, it was very heavily focused on search,
looking many, many moves ahead farther than any human could.
And that was key for why it won.
And then even with something like AlphaGo,
I mean, AlphaGo is commonly hailed
as a landmark achievement for neural nets, and it is,
but there's also this huge component of search,
Monte Carlo Tree Search to AlphaGo,
that was key, absolutely essential
for the AI to be able to beat top humans.
I think a good example of this is you look at
the latest versions of AlphaGo,
like it was called AlphaZero,
and there's this metric called Elo rating
where you can compare different humans
and you can compare bots to humans.
Now, a top human player is around 3,600 Elo,
maybe a little bit higher now.
AlphaZero, the strongest version, is around 5,200 Elo.
But if you take out the search that's being done
at test time, and by the way, what I mean by search
is the planning ahead, the thinking of like,
oh, if I move my, if I place this stone here
and then he does this,
and then you look like five moves ahead
and you see like what the board state looks like,
that's what I mean by search.
If you take out the search that's done during the game,
the Elo rating drops to around 3,000.
So even today, what, seven years after AlphaGo,
if you take out the Monte Carlo Tree Search
that's being done at when playing against the human,
the bots are not superhuman.
§07Transcript · 45:34–52:59
Nobody has made a raw neural net that is superhuman in Go.
- That's worth lingering on, that's quite profound.
So without search, that just means
looking at the next move and saying,
this is the best move.
So having a function that estimates accurately
what the best move is without search.
- Yeah, and all these bots, they have the,
what's called a policy network, where it will tell you,
this is what the neural net thinks is the next best move.
And it's kind of like the intuition that a human has.
You know, the human looks at the board
and any Go or chess master will be able to tell you like,
oh, instantly, here's what I think the right move is.
And the bot is able to do the same thing.
But just like how a human, grandmaster,
can make a better decision if they have more time to think,
when you add on this Monte Carlo Tree Search,
the bot is able to make a better decision.
- Yeah, I mean, of course a human is doing something
like search in their brain, but it's not,
I hesitate to draw a hard line,
but it's not like a Monte Carlo Tree Search.
It's more like sequential language model generation.
So it's like a different, it's a,
the neural network is doing the searching.
I wonder what the human brain is doing in terms of searching
'cause you're doing that like computation.
A human is computing.
They have intuition, they have gut,
they have a really strong ability to estimate,
you know, amongst the top players,
of what is good and not position
without calculating all the details.
But they're still doing search in their head,
but it's a different kind of search.
Have you ever thought about like,
what is the difference between the human,
the search that the human is performing
versus what computers are doing?
- I have thought a lot about that,
and I think it's a really important question.
So the AI in Alpha and Alphas in AlphaGo
or any of these Go AIs,
they're all doing Monte Carlo Tree Search,
which is a particular kind of search.
And it's actually a symbolic tabular search.
It uses the neural net to guide its search,
but it isn't actually like full on neural net.
Now, that kind of search is very successful
in these kinds of like perfect information board games
like chess and Go.
But if you take it to a game like poker, for example,
it doesn't work.
It can't understand the concept of hidden information.
It doesn't understand the balance that you have to strike
between like the amount that you're raising
versus the amount that you're calling.
And in every one of these games,
you see a different kind of search.
And the human brain is able to plan
for all these different games in a very general way.
Now, I think that's one thing
that we're missing from AI today.
And I think it's a really important missing piece.
The ability to plan and reason more generally
across a wide variety of different settings.
- In a way where the general reasoning
makes you better at each one of the games, not worse.
- Yeah, so you can kind of think of it
as like neural nets today,
they'll give you like Transformers, for example,
are super general, but they'll give you,
it'll output an answer in like 100 milliseconds.
And if you tell it like,
"Oh, you've got five minutes to give me a decision,
feel free to take more time to make a better decision."
It's not gonna know what to do with that.
But a human, if you're playing a game like chess,
they're gonna give you a very different answer
depending on if you say,
"Oh, you've got 100 milliseconds
or you've got five minutes."
- Yeah, I mean, people have started using
Transformers language models in an iterative way
that does improve the answer
or like showing the work kind of idea.
- Yeah, they got this thing called
chain of thought reasoning.
And that's, I think-
- Super promising, right?
- Yeah, and I think it's a good step in the right direction.
I would kind of like say it's similar
to Monte Carlo rollouts in a game like chess.
There's a kind of search that you can do
where you're saying like,
"I'm gonna roll out my intuition and see like,
without really thinking,
what are the better decisions I can make
farther down the path?
What would I do if I just acted according to intuition
for the next 10 moves?"
And that gets you an improvement,
but I think that there's much richer kinds of planning
that we could do.
- So when Labradors actually beat the poker players,
what did that feel like?
What was that?
I mean, actually on that day,
what were you feeling like?
Were you nervous?
I mean, poker was one of the games that you thought
like is not gonna be solvable
'cause it's the human factor.
So at least in the narratives,
we'll tell ourselves the human factor
is so fundamental to the game of poker.
- Yeah, the Labradors competition
was super stressful for me.
Also, I mean, I was working on this
like basically continuously for a year
leading up to the competition.
I mean, for me, it became like very clear,
like, okay, this is the search technique,
this is the approach that we need.
And then I spent a year working on this
pretty much like nonstop.
- Can we actually get into details?
Like what programming languages is it written in?
What's some interesting implementation details
that are like fun/painful?
- Yeah, so one of the interesting things about Labradors
is that we had no idea what the bar was
to actually beat top humans.
We could play against like our prior bots
and that kind of gives us some sense of like,
are we making progress?
Are we going in the right direction?
But we had no idea like what the bar actually was.
And so we threw a huge amount of resources
at trying to make the strongest bot possible.
So we use C++, it was parallelized.
We were using, I think like a thousand CPUs,
maybe more actually.
And today that sounds like nothing,
but for a grad student back in 2016,
that was a huge amount of resources.
- Well, it's still a lot for even any grad student today.
It's still tough to get,
or even to allow yourself to think in that,
in terms of scale at CMU, at MIT, anything like that.
- Yeah, and talking about terabytes of memory.
So it was a very parallelized,
and it had to be very fast too,
because the more games that you could simulate,
the stronger the bot would be.
- So is there some like John Carmack style,
like efficiencies you had to come up with,
like an efficient way to represent the hand,
all that kind of stuff?
- There were all sorts of optimizations that I had to make
to try to get this thing to run as fast as possible.
They were like, how do you minimize the latency?
How do you like, you know, package things together
so that like you minimize the amount of communication
between the different nodes?
How do you like optimize the algorithms
so that you can, you know,
try to squeeze out more and more
from the game that you're actually playing?
All these kinds of different decisions
that I, you know, had to make.
- Just a fun question.
What IDE did you use?
What for C++ at the time?
- I think I used Visual Studio actually.
- Okay. - Yeah.
- Is that still carried through to today?
- VS Code is what I use today.
It seems like it's pretty popular.
- It's the community, basically converged on.
Okay, cool.
So you got this super optimized C++ system,
and then you show up to the day of competition.
- Yeah.
- Humans versus machine.
How did it feel throughout the day?
- Super stressful.
I mean, I thought going into it
that we had like a 50/50 chance.
Because basically I thought if they play
in a totally normal style, I think we'll squeak out a win.
But there's always a chance
that they can find some weakness in the bot.
And if they do, and we're playing like for 20 days,
120,000 hands of poker.
They have a lot of time to find weaknesses in the system.
And if they do, we're gonna get crushed.
And that's actually what happened
in the previous competition.
§08Transcript · 53:01–1:00:39
The humans, they started out,
it wasn't like they were winning from the start.
But then they found these weaknesses
that they could take advantage of.
And for the next 10 days,
they were just crushing the bot, stealing money from it.
- What were the weaknesses they found?
Like maybe over betting was effective,
that kind of stuff.
So certain betting strategies worked?
- What they found is, yeah, over betting,
like betting certain amounts,
the bot would have a lot of trouble
dealing with those sizes.
And then also, when the bot got
into really difficult all-in situations,
it wasn't able to, because it wasn't doing search,
it had to clump different hands together
and it would treat them identically.
And so it wouldn't be able to distinguish,
you know, like having a king high flush
versus an ace high flush.
And in some situations that really matters a lot.
And so they could put the bot into those situations
and then the bot would just bleed money.
- Clever humans.
Okay, so I didn't realize it was over 20 days.
So what were the humans like over those 20 days?
And what was the bot like?
- So we had set up the competition, you know,
like I said, there was $200,000 in prize money
and they would get paid a fraction of that
depending on how well they did relative to each other.
So I was kind of hoping that they wouldn't work together
to try to find weaknesses in the bot,
but they enter the competition
with their like number one objective being to beat the bot.
And they didn't care about like individual glory.
They were like, we're all gonna work as a team
to try to take down this bot.
And so they immediately started comparing notes.
What they would do is they would coordinate
looking at different parts of the strategy
to try to, you know, find out weaknesses.
And then at the end of the day,
we actually sent them a log of all the hands
that were played and what cards the bot had
on each of those hands.
- Oh wow. - Yeah.
- That's gutsy.
- Yeah, it was honestly,
and I'm not sure why we did that in retrospect,
but I mean, I'm glad we did it
'cause we ended up winning anyway,
but that if you've ever played poker before,
like that is golden information.
I mean, to know, usually when you play poker,
you see about a third of the hands to show down
and to just hand them all the cards
that the bot had on every single hand,
that was just a gold mine for them.
And so then they would review the hands
and try to see like, okay,
could they find patterns in the bot, weaknesses?
And could they, then they would coordinate
and study together and try to figure out,
okay, now this person's gonna explore
this part of the strategy for weaknesses.
This person's gonna explore this part
of the strategy for weaknesses.
- It's a kind of psychological warfare,
showing them the hands.
- Yeah.
- I mean, I'm sure you didn't think of it that way,
but like doing that means you're confident
in the bot's ability to win.
- Well, that's one way of putting it.
I wasn't super confident.
- Yeah.
- So, going in, like I said,
I think I had like 50/50 odds on us winning.
When we actually, when we announced the competition,
the poker community decided to gamble on who would win.
And their initial odds against us were like four to one.
They were really convinced that the humans
were gonna pull out a win.
The bot ended up winning for three days straight.
And even then after three days,
the betting odds were still just 50/50.
And then at that point, it started to look like
the humans were coming back.
They started to like, you know,
but poker is a very high variance game.
And I think what happened is like,
they thought that they spotted some weaknesses
that weren't actually there.
And then around day eight,
it was just very clear
that they were getting absolutely crushed.
And from that point, I mean, for a while there,
I was super stressed out thinking like,
oh my God, the humans are coming back
and they've found weaknesses
and now we're just gonna lose the whole thing.
But no, it ended up going in the other direction
and the bot ended up like crushing them in the long run.
- How did it feel at the end?
Like as a human being,
as a person who loves,
appreciates the beauty of the game of poker
and as a person who appreciates the beauty of AI,
is there, did you feel a certain kind of way about it?
- I felt a lot of things, man.
I mean, at that point in my life,
I had spent five years working on this project
and it was a huge sense of accomplishment.
I mean, to spend five years working on something
and finally see it succeed.
Yeah, I wouldn't trade that for anything in the world.
- Yeah, because that's a real benchmark.
It's not like getting some percent accuracy on a data set.
This is like real, this is real world.
It's just a game, but it's also a game
that means a lot to a lot of people.
And this is humans doing their best to beat the machine.
So this is a real benchmark, unlike anything else.
- Yeah, and I mean, this is what I had been dreaming about
since I was like 16 playing poker,
you know, with my friends in high school.
The idea that you could find a strategy
to approximate the Nash equilibrium,
be able to beat all the poker players in the world with it.
So to actually see that come to fruition and be realized,
that was, it's kind of magical.
- Yeah, especially money is on the line too.
It's different than chess.
And that aspect, like people get,
that's why you want to look at betting markets
if you want to actually understand what people really think.
And in the same sense, poker, it's really high stakes
'cause it's money.
And to solve that game, that's an amazing accomplishment.
So the leap from that to multi-way six player poker,
what's, how difficult is that jump?
And what are some interesting differences
between heads up poker and multi-way poker?
- Yeah, so I mentioned, you know,
Nash equilibrium in two player zero-sum games.
If you play that strategy,
you are guaranteed to not lose an expectation
no matter what your opponent does.
Now, once you go to six player poker,
you're no longer playing a two player zero-sum game.
And so there was a lot of debate
among the academic community and among the poker community
about how well these techniques would extend
beyond just two player heads up poker.
Now, what I had come to realize is that
the techniques actually I thought
really would extend to six player poker
because even though in theory,
they don't give you these guarantees
outside of two player zero-sum games,
in practice, it still gives you a really strong strategy.
Now, there were a lot of complications
that would come up with six player poker
besides like the game theoretic aspect.
I mean, for one, the game is just exponentially larger.
So the main thing that allowed us
to go from two player to six player
was the idea of depth limited search.
So I said before, like, you know, we would do search,
we would plan out, the bot would plan out
like what it's going to do next
and for the next several moves.
And in Libratus, that search was done
extending all the way to the end of the game.
So it would have to start from the turn onwards,
like looking maybe 10 moves ahead,
it would have to figure out
what it was doing for all those moves.
Now, when you get to six player poker,
it can't do that exhaustive search anymore
'cause the game is just way too large.
But by only having to look a few moves ahead
and then stopping there and substituting a value estimate
of like how good is that strategy at that point,
then we're able to do a much more scalable form of search.
- Is there something cool,
looking at the paper right now,
is there something cool in the paper in terms of graphics?
A game tree traversal via Monte Carlo.
- I think if you go down a bit.
- Figure one, an example of equilibrium selection problem.
Ooh, so yeah.
What do we know about equilibria
when there's multiple players?
- So when you go outside of two players, you're a sum.
So a Nash equilibrium is a set of strategies,
like one strategy for each player,
§09Transcript · 1:00:41–1:09:02
where no player has an incentive
to switch to a different strategy.
And so you can kind of think of it as like,
imagine you have a game where there's a ring.
That's actually the visual here.
You got a ring and the object of the game
is to be as far away from the other players as possible.
There's a Nash equilibrium is for all the players
to be spaced equally apart around this ring.
But there's infinitely many different Nash equilibria.
There's infinitely many ways
to space four dots along a ring.
And if every single player independently
computes a Nash equilibrium,
then there's no guarantee that the joint strategy
that they're all playing is going to be a Nash equilibrium.
They're just gonna be like random dots
scattered along this ring,
rather than four coordinated dots
being equally spaced apart.
- Is it possible to sort of optimally
do this kind of selection,
to do the selection of the equilibria you're chasing?
So is there like a meta problem to be solved here?
- So the meta problem is in some sense,
how do you understand the Nash equilibria
that the other players are going to play?
And even if you do that, again,
there's no guarantee that you're going to win.
So, if you're playing risk, like I said,
and all the other players decide to team up against you,
you're gonna lose.
Nash equilibrium doesn't help you there.
And so there is this big debate about
whether Nash equilibrium and all these techniques
that compute it are even useful
once you go outside of two player zero-sum games.
Now, I think for many games,
there is a valid criticism here.
And I think when we talk about,
when we go to something like diplomacy,
we run into this issue that the approach
of trying to approximate a Nash equilibrium
doesn't really work anymore.
But it turns out that in six player poker,
because six player poker is such an adversarial game,
where none of the players
really try to work with each other,
the techniques that were used in two player poker
to try to approximate an equilibrium,
those still end up working in practice
in six player poker as well.
- There's some deep way in which six player poker
is just a bunch of heads up poker, like games in one.
It's like embedded in it.
So the competitiveness is more fundamental to poker
than the cooperation.
- Right, yeah.
Poker is just such an adversarial game.
There's no real cooperation.
In fact, you're not even allowed to cooperate in poker.
It's considered collusion.
It's against the rules.
And so for that reason,
the techniques end up working really well.
And I think that's true more broadly
in extremely adversarial games in general.
- But that's sort of in practice
versus being able to prove something.
- That's right.
Nobody has a proof that that's the case.
And it could be that six player poker
belongs to some class of games
where approximating an equilibrium through self-play
provably works well.
And there are other classes of games
beyond just two players, zero sum,
where this is proven to work well.
So there are these kinds of games called potential games,
which I won't go into.
It's kind of like a complicated concept,
but there are classes of games
where this approach to approximating an equilibrium
is proven to work well.
Now, six player poker is not known to belong
to one of those classes,
but it is possible that there is some classic games
where it either provably performs well
or provably performs not that badly.
- So what are some interesting things about Pleribus
that was able to achieve human level performance
on this or superhuman level performance
on the six player version of poker?
- Personally, I think the most interesting thing
about Pleribus is that it was so much cheaper than Libratus.
I mean, Libratus, if you had to put a price tag
on the computational resources that went into it,
I would say the final training run took about $100,000.
You go to Pleribus, the final training run
would cost like less than $150 on AWS.
- Is this normalized to computational inflation?
So meaning, does this just have to do with the fact
that Pleribus was trained like a year later?
- No, no, no, it's not.
I mean, first of all, like, yeah,
computing resources are getting cheaper every day,
but you're not gonna see a thousand fold decrease
in the computational resources over two years
or even anywhere close to that.
The real improvement was algorithmic improvements
and in particular, the ability to do depth limited search.
- So does depth limited search also work for Libratus?
- Yeah, yes.
So where this depth limited search came from is,
you know, I developed this technique
and ran it on two player poker first
and that reduced the computational resources needed
to make an AI that was superhuman
from, you know, $100,000 for Libratus
to something you could train on your laptop.
- What do you learn from that, from that discovery?
- What I would take away from that
is that algorithmic improvements really do matter.
- How would you describe the more general case
of limited depth search?
So it's basically constraining the scale, temporal,
or in some other way of the computation you're doing,
in some clever way.
So like with, like how else can you significantly
constrain computation, right?
- Well, I think the idea is that we want to be able
to leverage search as much as possible.
And the way that we were doing it in Libratus
required us to search all the way to the end of the game.
Now, if you're playing a game like chess,
the idea that you're gonna search always
to the end of the game is kind of unimaginable, right?
Like there's just so many situations
where you just won't be able to use search in that case
or the cost would be, you know, prohibitive.
And this technique allowed us to leverage search
and without having to pay such a huge
computational cost for it,
and be able to apply it more broadly.
- So to what degree did you use neural nets
for Libratus and Pleribus?
And more generally, what role do neural nets have to play
in superhuman level performance in poker?
- So we actually did not use neural nets at all
for Libratus or Pleribus.
And a lot of people found this surprising back in 2017.
I think they find it surprising today
that we were able to do this without using any neural nets.
And I think the reason for that,
I mean, I think neural nets are incredibly powerful
and the techniques that are used today,
even for poker AIs, do rely quite heavily on neural nets.
But it wasn't the main challenge for poker.
Like I think what neural nets are really good for,
if you're in a situation where finding features
for a value function is really difficult,
then neural nets are really powerful.
And this was the problem in Go, right?
Like the problem in Go was that,
or the final problem in Go at least,
was that nobody had a good way of looking at a board
and figuring out who was winning or losing,
describing through a simple algorithm
who was winning or losing.
And so there neural nets were super helpful
because you could just feed in a ton
of different board positions into this neural net,
and it would be able to predict then
who was winning or losing.
But in poker, the features weren't the challenge.
The challenge was how do you design a scalable algorithm
that would allow you to find this balanced strategy
that would understand that you have to bluff
with the right probability?
- So can that be somehow incorporated
into the value function?
The complexity of poker that you've described?
- Yeah, so the way the value functions work in poker,
like the latest and greatest poker AIs,
they do use neural nets for the value function.
The way it's done is very different
from how it's done in a game like chess or Go,
because in poker, you have to reason about beliefs.
And so the value of a state depends on the beliefs
that players have about what the different cards are.
Like if you have pocket aces,
then whether that's a really, really good hand
or just an okay hand depends on whether you know
I have pocket aces.
Rather, if you know that I have pocket aces,
then if I bet, you're gonna fold immediately.
But if you think that I have a really bad hand,
then I could bet with pocket aces and make a ton of money.
So the value function in poker these days
takes the beliefs as an input,
§10Transcript · 1:09:05–1:17:01
which is very different from how chess and Go AIs work.
- So as a person who appreciates the game,
who do you think is the greatest poker player of all time?
- That's a tough question.
- Can AI help answer that question?
Can you actually analyze the quality of play?
So the chess engines can give estimates
of the quality of play.
I wonder if there's a,
is there an Elo rating type of system for poker?
I suppose you could, but there's just not enough.
You would have to play a lot of games, right?
A very large number of games,
like more than you would in chess.
The deterministic game makes it easier to estimate Elo.
I think.
- I think it is much harder to estimate
something like Elo rating in poker.
I think it's doable.
The problem is that the game is very high variance.
So you could play,
you could be profitable in poker for a year
and you could actually be a bad player
just because the variance is so high.
I mean, you've got top professional poker players
that would lose for a year
just because they're on a really bad streak.
- So for Elo, you have to have a nice clean way of saying
if player A played player B and A beats B,
that says something, that's a signal.
In poker, that's a very noisy signal.
- It's a very noisy signal.
Now there is a signal there.
And so you could do this calculation.
It would just be much harder.
But the same way that AIs have now taken over chess
and all the top professional chess players train with AIs,
the same is true for poker.
The game has become a very computational,
people train with AIs to try to find out
where they're making mistakes,
try to learn from the AIs to improve their strategy.
So now, yeah, so the game has been revolutionized
in the past five years by the development of AI
in this sport.
- The skill with which you avoided the question
of the greatest of all time was impressive.
- So my feeling is that it's a difficult question
because just like in chess,
where you can't really compare Magnus Carlsen today
to Garry Kasparov, because the game has evolved so much.
The poker players today are so far beyond the skills
of people that were playing even 10 or 20 years ago.
So you look at the kinds of all-stars that were on ESPN
at the height of the poker boom,
pretty much all those players are actually not that good
at the game today, at least the strategy aspect.
I mean, they might still be good at reading the player
at the other side of the table and trying to figure out
are they bluffing or not?
But in terms of the actual computational strategy
of the game, a lot of them have really struggled
to keep up with that development.
Now, so for that reason, I'll give an answer
and I'm gonna say Daniel Legranio,
who you actually had on the podcast recently,
I saw it was a great episode.
- I love this so much.
(laughing)
And Phil's gonna hate this so much.
- And I'm gonna give him credit
because he is one of the few old school,
really strong players that have kept up
with the development of AI.
- So he is trying to, he's constantly studying
the game theory optimal way of playing.
- Exactly, yeah.
And I think a lot of the old school poker players
have just kind of given up on that aspect
and I gotta give Daniel Legranio credit
for keeping up with all the developments
that are happening in the sport.
- Yeah, it's fascinating to watch.
It's fascinating to watch where it's headed.
Yeah, so there you go, some love for Daniel.
Quick pause, bath and break?
- Yeah, let's do it.
- Let's go from poker to diplomacy.
What is at a high level the game of diplomacy?
- Yeah, so I talked a lot about two players,
zero sum games.
And what's interesting about diplomacy
is that it's very different from these adversarial games
like chess, go, poker, even Starcraft and Dota.
Diplomacy has a much bigger cooperative element to it.
It's a seven player game.
It was actually created in the fifties
and it takes place before World War I.
It's like a map of Europe with seven great powers
and they're all trying to form alliances with each other.
There's a lot of negotiation going on.
And so the whole focus of the game
is on forming alliances with the other players
to take on the other players.
- England, Germany, Russia, Turkey,
Austria, Hungary, Italy, and France.
- That's right, yeah.
So the way the game works is on each turn,
you spend about five to 15 minutes
talking to the other players in privates
and you make all sorts of deals with them.
You say like, "Hey, let's work together.
Let's team up against this other player."
Because the only way that you can make progress
is by working with somebody else against the others.
And then after that negotiation period is done,
all the players simultaneously submit their moves
and they're all executed at the same time.
And so you can tell people like,
"Hey, I'm gonna support you this turn,"
but then you don't follow through with it.
And they're only gonna figure that out
once they see the moves being read off.
- How much of it is natural language,
like written actual text?
How much is like,
you're actually saying phrases that are structured?
- So there's different ways to play the game.
You can play it in person,
and in that case, it's all natural language.
Free form communication.
There's no constraints on the kinds of deals
that you can make, the kinds of things that you can discuss.
You can also play it online.
So you can send long emails back and forth.
You can play it live online or over voice chat.
But the focus, the important thing to understand
is that this is unstructured communication.
You can say whatever you want.
You can make any sorts of deals that you want
and everything is done privately.
So it's not like you're all around the board together
having a conversation.
You're grabbing somebody going off into a corner
and conspiring behind everybody else's back
about what you're planning.
- And there's no limit in theory to the conversation
you can have directly with one person.
- That's right.
You can make all sorts of,
you can talk about anything.
You can say like,
"Hey, let's have a long-term alliance against this guy."
You can say like,
"Hey, can you support me this turn?
And in return, I'll do this other thing for you next turn."
Or, you know, yeah,
just you can talk about like what you talked about
with somebody else
and gossip about like what they're planning.
The way that I would describe the game
is that it's kind of like a mix between Risk,
poker, and the TV show "Survivor."
There's like this big element of like trying to,
yeah, there's a big social element.
And the best way that I would describe the game
is that it's really a game about people
rather than the pieces.
- So Risk, because it is a map,
it's kind of war game-like.
Poker, because there's a game theory component
that's very kind of strategic.
So you could convert it
into an artificial intelligence problem.
And then Survivor, because of the social component,
strong social component.
I saw that somebody said online
that the internet version of the game
has this quality of that it's easier
to almost to do like role-playing.
As opposed to being yourself,
you can actually like be the,
like really imagine yourself as the leader of France
or Russia and so on.
Like really pretend to be that person.
It's actually fun to really lean into being that leader.
- Yeah, so some players do go this route
where they just like kind of view it as a strategy game,
but also a role-playing game where they can like act out,
like, what would I be like if I was, you know,
a leader of France in 1900?
- I'll forfeit right away.
No, I'm just kidding.
And they sometimes use like the old-timey language
to like, or how they imagined the elites
would talk at that time.
Anyway, so what are the different turns of the game?
Like what are the rounds?
- Yeah, so on every turn,
you got like a bunch of different units
§11Transcript · 1:17:03–1:24:33
that you start out with.
So you start out controlling like just a few units
and the object of the game is to gain control
of a majority of the map.
If you're able to do that, then you've won the game.
But like I said, the only way that you're able to do that
is by working with other players.
So on every turn, you can issue a move order.
So for each of your units,
you can move them to an adjacent territory,
or you can keep them where they are,
or you can support a move or a hold of a different unit.
So--
- What are the territories?
Well, how is the map divided up?
- It's kind of like risk where the map is divided up
into like 50 different territories.
Now you can enter a territory
if you're moving into that territory with more supports
than the person that's in there
or the person that's trying to move in there.
So if you're moving in and there's somebody already there,
then if neither of you have support, it's a one versus one
and you'll bounce back and neither of you will make progress.
If you have a unit that's supporting
that move into the territory, then it's a two versus one
and you'll kick them out
and they'll have to retreat somewhere.
- What does support mean?
- Support is like, it's an action
that you can issue in the game.
So you can say this unit, you write down,
this unit is supporting this other unit into this territory.
- Are these units from opposing forces?
- They could be, they could be.
And this is where the interesting aspect
of the game comes in
because you can support your own units into territory,
but you can also support other people's units
into territories.
And so that's what the negotiations really revolve around.
- But you don't have to do the thing you say
you're going to do, right?
- Yeah, and so this is--
- So you can say, I'm gonna support you,
but then backstab the person.
- Yeah, that's absolutely right.
- And that tension is core to the game?
- That tension is absolutely core to the game.
The fact that you can make all sorts of promises,
but you have to reason about the fact that like,
hey, they might not trust you
if you say you're gonna do something,
or they might be lying to you
when they say that they're gonna support you.
- So maybe just to jump back,
what's the history of the game in general?
Is it true that Henry Kissinger loved the game
and JFK and all those?
I've heard like a bunch of different people that,
or is that just one of those things
that the cool kids say they do,
but they don't actually play?
- So the game was created in the '50s.
- Yeah.
- And from what I understand, it was JFK's,
it was played in like the JFK White House,
Henry Kissinger's favorite game.
I don't know if it's true,
but that's definitely what I've heard.
- It's interesting that they went with World War I
when it was created after World War II.
- So the story that I've heard for the creation of the game
is it was created by somebody that had looked at
the history of the 20th century,
and they saw World War I as a failure of diplomacy.
So-
- Yeah.
- They saw the fact that this war broke out
as like the diplomats of all these countries
really failed to prevent a war.
And he wanted to create a game
that would basically teach people about diplomacy.
And it's really fascinating that in his ideal version
of the game of diplomacy, nobody actually wins the game.
Because the whole point is that if somebody is about to win,
then the other players should be able to work together
to stop that person from winning.
And so the ideal version of the game
is just one where nobody actually wins.
And it's kind of has a nice like wholesome take home message
then that war is ultimately futile.
And-
- And that optimal, that futile optimal
could be achieved through great diplomacy.
- Yeah.
- So is there some asymmetry in terms of
which is more powerful, Russia versus Germany
versus France and so on?
- So I think the general consensus is that France
is the strongest power in the game.
But the beautiful thing about diplomacy
is that it's self-balancing, right?
So the fact that France has an inherited advantage
from the beginning means that the other players
are less likely to work with it.
- I saw that Russia has four units
versus four of something
that the others have three of something.
- That's true, yeah.
So Russia starts off with four units
while all the other players start with three.
But Russia is also in a much more vulnerable position
because they have to like,
they have a lot more neighbors as well.
- Got it.
Larger territory, more, yeah, right.
More border to defend.
Okay, what else is important to know about the rules?
So how many rounds are there?
Like, is this iterative game?
Is it finite?
Do you just keep going indefinitely?
- Usually the game lasts, I would say about 15 or 20 turns.
There's in theory, no limit.
It could last longer, but at some point,
I mean, if you're playing a house game with friends,
at some point you just get tired and you all agree like,
okay, we're gonna end the game here and call it a draw.
If you're playing online, there's usually like set limits
on when the game will actually end.
- And what's the end, what's the termination condition?
Like, does one country have to conquer everything else?
- So if somebody is able to actually gain control
of a majority of the map, then they've won the game.
And that is a solo victory as it's called.
Now that pretty rarely happens,
especially with strong players, because like I said,
the game is designed to incentivize the other players
to put a stop to that and all work together
to stop the superpower.
Usually what ends up happening is that, you know,
all the players agree to a draw and then the score,
the win is divided among the remaining players.
There's a lot of different scoring systems.
The one that we used in our research basically gives
a score relative to how much control you have of the map.
So the more that you control, the higher your score.
- What's the history of using this game
as a benchmark for AI research?
Do people use it?
- Yeah, so people have been working on AI for diplomacy
since about the '80s.
There was some really exciting research back then,
but the approach that was taken was very different
from what we see today.
I mean, the research in the '80s was a very rule-based
approach, kind of a heuristic approach.
It was very in line with the kind of research
that was being done in the '80s.
You know, basically trying to encode human knowledge
into the strategy of the AI.
- Sure.
- And, you know, it's understandable.
I mean, the game is so incredibly different
and so much more complicated than the kinds of games
that people were working on like chess and go and poker
that it was honestly even hard to like start
making any progress in diplomacy.
- Can you just formulate what is the problem
from an AI perspective and why is it hard?
Why is it a challenging game to solve?
- So there's a lot of aspects in diplomacy
that make it a huge challenge.
First of all, you have the natural language components.
And I think this really is what makes it
arguably the most difficult game
among like the major benchmarks.
The fact that you have to,
it's not about moving pieces on the board.
Your action space is basically all the different sentences
that you could communicate to somebody else in this game.
And-
- Is there, can we just like linger on that?
So is part of it like the ambiguity in the language?
If it was like very strict,
if you narrowed the set of possible sentences
you could do it,
would that simplify the game significantly?
- The real difficulty is the breadth of things
that you can talk about.
You can have natural language in other games
like Settlers of Catan, for example,
like you could have a natural language,
Settlers of Catan AI.
But the things that you're gonna talk about
are basically like, am I trading you two sheep for a wood
or three sheep for a wood?
Whereas in a game like Diplomacy,
the breadth of conversations that you're going to have
§12Transcript · 1:24:35–1:32:34
are like, am I going to support you?
Are you gonna support me in return?
Which units are gonna do what?
What did this other person promise you?
They're lying because they told this other person
that they're gonna do this instead.
If you help me out this turn,
then in the future I'll do these things
that will help you out.
The depth and breadth of these conversations
is really complicated.
And it's all being done in natural language.
Now you could approach it,
and we actually considered doing this,
like having a simplified language
to make this complexity smaller.
But ultimately we thought the most impactful way
of doing this research would be to address
the natural language component head on
and just try to go for the full game upfront.
- Just looking at sample games
and what the conversations look like.
Greetings England, this should prove to be a fun game
since all the private press
is going to be made public at the end.
At the least it will be interesting to see
if the press changes because of that.
Anyway, good, okay.
So there's like a--
- Yeah, that's just kind of like the generic greetings
at the beginning of the game.
I think that the meat comes a little bit later
when you're starting to talk about
specific strategy and stuff.
- I agree there are a lot of advantages
to the two of us keeping in touch
and our nations make strong natural allies
in the middle game.
So that kind of stuff, making friends, making enemies.
- Yeah, or like if you look at the next line,
so the person saying like,
I've heard bits about a Lepanto and an octopus opening
and basically telling Austria like,
hey, just a heads up, you know,
I've heard these whispers about like
what might be going on behind your back.
- Yeah, so there's all kinds of complexities
in the language of that, right?
Like to interpret what the heck that means.
It's hard for us humans, but for AI it's even harder
'cause you have to understand like at every level
the semantics of that.
- Right, I mean, there's the complexity in understanding
when somebody is saying this to me, what does that mean?
And then there's also the complexity of like,
should I be telling this person this?
Like I've overheard these whispers,
should I be telling this person that like,
hey, you might be getting attacked by this other power?
- Okay, so how are we supposed to think about?
Okay, so that's the natural language.
How do you even begin trying to solve this game?
It seems like the Turing test on steroids.
- Yeah, and I mean, there's the natural language aspect.
And then even besides the natural language aspect,
you also have the cooperative elements of the game.
And I think this is actually something
that I find really interesting.
If you look at all the previous game AI breakthroughs,
they've all happened in these purely adversarial games
where you don't actually need to understand
how humans play the game.
It's all just AI versus AI, right?
Like you look at checkers, chess, go, poker,
Starcraft, Dota 2, like in some of those cases,
they leveraged human data, but they never needed to.
They were always just trying to have a scalable algorithm
that then they could throw a lot of computational resources
at a lot of memory at, and then eventually it would converge
to an approximation of a Nash equilibrium.
This perfect strategy that in a two player zero-sum game
guarantees that they're going to be able
to not lose to any opponent.
- So you can't leverage self-play to solve this game.
- You can leverage self-play,
but it's no longer sufficient to beat humans.
- So how do you integrate the human into the loop of this?
- So what you have to do is incorporate human data.
And to kind of give you some intuition
for why this is the case,
like imagine you're playing a negotiation game,
like diplomacy, but you're training completely from scratch
without any human data.
The AI is not going to suddenly like figure out
how to communicate in English.
It's going to figure out some weird robot language
that only it will understand.
And then when you stick that in a game
with six other humans,
they're going to think this person's talking gibberish
and they're just going to ally with each other
and team up against the bot,
or not even team up against the bot,
but just not work with the bot.
And so in order to be able to play this game with humans,
it has to understand the human way of playing the game,
not this machine way of playing the game.
- Yeah, yeah, that's fascinating.
So, right.
That's a nuanced thing to understand
'cause a chess playing program
doesn't need to play like a human to beat a human.
- Exactly.
- But here you have to play like a human
in order to beat them.
- Or at least you have to understand
how humans play the game
so that you can understand how to work with them.
If they have certain expectations
about what does it mean to be a good ally?
What does it mean to have like a reciprocal relationship
where we're working together?
You have to abide by those conventions.
And if you don't,
they're just going to work with somebody else instead.
- Do you think of this as a clean,
in some deep sense of the spirit of the Turing test
as formulated by Alan Turing?
Is it, in some sense,
this is what the Turing test actually looks like?
- So, because of open-ended natural language conversation
it seems like very difficult to evaluate.
Like here at a high stakes
where humans are trying to win a game,
that seems like how you actually perform the Turing test.
- I think it's different from the Turing test.
Like the way that the Turing test is formulated,
it's about trying to distinguish a human from a machine
and seeing, oh, could the machine successfully pass
as a human in this adversarial setting
where the player is trying to figure out
whether it's a machine or a human.
Whereas in diplomacy, it's not about trying to figure out
whether this player is a human or a machine.
It's ultimately about whether I can work with this player
regardless of whether they are a human or a machine.
And can the machine do that better than a human can?
- Yeah, I'm going to think about that,
but that just feels like the implied requirement for that
is for the machine to be human-like.
- I think that's true,
that if you're going to play in this human game,
you have to somehow adapt to the human surroundings
and the human play style.
- And to win, you have to adapt.
So you can't, if you're the outsider,
if you're not human-like,
I feel like that's a losing strategy.
- I think that's correct, yeah.
- Yeah, so, okay.
What are the complexities here?
What was your approach to it?
- Before I get to that,
one thing I should explain
why we decided to work on Diplomacy.
So basically what happened is in 2019,
I was wrapping up the work on six-player poker on Pluribus
and was trying to think about what to work on next.
And I had been seeing all these other breakthroughs
happening in AI.
I mean, 2019, you have StarCraft,
you have AlphaStar beating humans in StarCraft,
you've got the Dota 2 stuff happening at OpenAI,
you have GPT-2 or GPT-3 coming,
I think it was GPT-2 at the time.
And it became clear that AI was progressing
really, really rapidly.
And people were throwing out these other games
about what should be the next challenge for multi-agent AI.
And I just felt like we had to aim bigger.
If you look at a game like chess or a game like Go,
they took decades for researchers
to ultimately reach superhuman performance at.
I mean, chess took 40 years of AI research
Go took another 20 years.
And we thought that diplomacy
would be this incredibly difficult challenge
that could easily take a decade
to make an AI that could play competently.
But we felt like that was a goal worth aiming for.
And so honestly, I was kind of reluctant
to work on it at first
because I thought it was like too far
out of the realm of possibility.
But I was talking to a coworker of mine, Adam Lear,
and he was basically saying like,
"Why not aim for it?
"We'll learn some interesting things along the way
"and maybe it'll be possible."
And so we decided to go for it.
§13Transcript · 1:32:36–1:40:28
And I think it was the right choice
considering just how much progress there was in AI
and that progress has continued in the years since.
- So winning in diplomacy, what does that really look like?
It means talking to six other players,
six other entities, agents,
and convincing them of stuff
that you want them to be convinced of.
Like what exactly, I'm trying to get like,
to deeply understand what the problem is.
- Ultimately, the problem is simple to quantify, right?
Like you're going to play this game with humans
and you want your score on average
to be as high as possible.
You know, if you can say like,
"I am winning more than any human alive,"
then you're a champion diplomacy player.
Now, ultimately we didn't reach that.
We got to human level performance.
We actually, so we played about 40 games
with real humans online.
The bot came in second out of all players
that played five or more games.
And so not like number one, but way, way higher than-
- What was the expertise level?
Are they beginners?
Are they intermediate players, advanced players?
Do you have a sense?
- That's a great question.
And so I think this kind of goes into
how do you measure the performance in diplomacy?
And I would argue that when you're measuring performance
in a game like this,
you don't actually want to measure it
in games with all expert players.
It's kind of like if you're developing a self-driving car,
you don't want to measure that car on the road
with a bunch of expert stunt drivers.
You want to put it on a road of like an actual American city
and see is this car crashing less often
than an expert driver would.
So that's the metric that we've used.
We're saying like, we're going to stick this game,
we're going to stick this bot in games
with a wide variety of skill levels.
And then are we doing better than a strong
or expert human player would in the same situation?
- That's quite brilliant.
'Cause I played a lot of sports in my life,
like I did tennis, judo, whatever.
And it's somehow almost easier
to go against experts almost always.
I don't know.
I think they're more predictable in the quality of play.
The space of strategies you're operating under
is narrower against experts.
It's more fun.
It's really frustrating to go against beginners.
Also 'cause beginners talk trash to you
when they somehow do beat you.
So that's a human thing that they add
is not to be worried about that.
But yeah, the variance in strategies is greater,
especially with natural language.
It's just all over the place then.
- Yeah.
And honestly, when you look at
what makes a good human diplomacy player,
obviously they're able to handle themselves
in games with other expert humans,
but where they really shine
is when they're playing with these weak players.
And they know how to take advantage
of the fact that they're a weak player,
that they won't be able to pull off a stab as well,
or that they have certain tendencies
and they can take them under their wing
and persuade them to do things
that might not even be in their interest.
The really good diplomacy players
are able to take advantage of the fact
that there are some weak players in the game.
- Okay, so if you have to incorporate human play data,
how do you do that?
How do you do that in order to train
an AI system to play diplomacy?
- Yeah, so that's really the crux of the problem.
How do we leverage the benefits of self-play
that have been so successful
in all these other previous games,
while keeping the strategy as human compatible as possible?
And so what we did is we first trained a language model,
and then we made that language model controllable
on a set of intents, what we call intents,
which are basically like an action that we want to play
and an action that we would like the other player to play.
And so this gives us a way to generate dialogue
that's not just trying to imitate the human style,
whatever a human would say in this situation,
but to actually give it an intent,
a purpose in its communication.
We can talk about a specific move
or we can make a specific request.
And the determination of what that move is
that we're discussing comes from a strategic reasoning model
that uses reinforcement learning and planning.
- So the computing the intents for all the players,
how is that done?
Just as a starting point,
is that with reinforcement learning
or is that just optimal,
determining what the optimal is for intents?
- It's a combination of reinforcement learning and planning.
Actually very similar to how we approached poker
and how people have approached chess and Go as well.
We're using self-play and search to try to figure out
what is an optimal move for us
and what is a desirable move
that we would like this other player to play.
Now, the difference between the way that we approached
reinforcement learning and search in this game
versus those previous games
is that we have to keep it human compatible.
We have to understand how the other person
is likely to play rather than just assuming
that they're gonna play like a machine.
- And how language gets them to play
in a way that maximize the chance of following the intent
you want them to follow.
Okay, how do you do that?
How do you connect language to intent?
- So the way that RL and planning is done
is actually not using language.
So we're coming up with this plan for the action
that we're gonna play and the other person's gonna play
and then we feed that action into the dialogue model
that will then send a message according to those plans.
- So the language model there is mapping action to...
- To message. - To message.
One word at a time.
- Basically one message at a time.
So we'll feed into the dialogue model,
like here are the actions that you should be discussing.
Here's the message,
here's like the content of the message
that we would like you to send
and then it will actually generate a message
that corresponds to that.
- Okay, does this actually work?
- It works surprisingly well.
- Okay, how...
(laughs)
Oh man, the number of ways it probably goes horribly,
I would have imagined it goes horribly wrong.
So how the heck is it effective at all?
- I mean, there are a lot of ways that this could fail.
So for example, I mean, you could have a situation
where you're basically like,
we don't tell the language model,
like here are the pieces of our action
or the other person's action
that you should be communicating.
And so like, let's say you're about to attack somebody,
you probably don't wanna tell them
that you're going to attack them,
but there's nothing in the language,
like the language model is not very smart
at the end of the day.
So it doesn't really have a way of knowing like,
well, what should I be talking about?
Should I tell this person
that I'm about to attack them or not?
So we have to like develop a lot of other techniques
that deal with that.
Like one of the things we do for example,
is we try to calculate if I'm going to send this message,
what would I expect the other person to do in response?
So if it's a message like,
hey, I'm gonna attack you this turn,
they're probably gonna, you know, attack us
or defend against that attack.
And so we have a way of recognizing like,
hey, sending this message
is a negative expected value action
and we should not send this message.
- So you have for particular kinds of messages,
you have like an extra function
that does the, estimates the value of that message.
- Yeah, so we have these kinds of filters that like-
- So it's a filter.
So there's a good, is that filter in your network
or is it rule-based?
- That's a neural network.
So we're, well, it's a combination.
It's a neural network, but it's also using planning.
It's trying to compute like,
what is the policy that the other players are going to play
given that this message has been sent?
And then is that better than not sending the message?
- I feel like that's how my brain works too.
§14Transcript · 1:40:30–1:48:31
Like there's a language model that generates random crap
and then there's these other neural nets
that are essentially filters.
At least that's what I tweet.
I'll usually, my process of tweeting,
I'll think of something and it's hilarious to me.
And then about five seconds later,
the filter network comes in and says,
no, no, that's not funny at all.
I mean, there's something interesting
to that kind of process.
So you have a set of actions that you want,
you have an intent that you want to achieve,
an intent that you want your opponent to achieve,
then you generate messages.
And then you evaluate if those messages
will achieve the goal you want.
- Yeah, and we're filtering for several things.
We're filtering like, is this a sensible message?
So sometimes language models will send,
will generate messages that are just like totally nonsense.
And we try to filter those out.
We also try to filter out messages that are basically lies.
So diplomacy has this reputation as a game
that's really about deception and lying,
but we try to actually minimize the amount
that the bot would lie.
This was actually mostly a-
- Or are you?
No, I'm just kidding.
All right, go ahead.
(laughing)
- I mean, like part of the reason for this
is that we actually found that lying
would make the bot perform worse in the long run.
It would end up with a lower score.
Because once the bot lies,
people would never trust it again.
And trust is a huge aspect of the game of diplomacy.
- I'm taking notes here,
'cause I think this applies to life lessons too.
- Oh, I think it's a really, yeah, really strong-
- So like lying is a dangerous thing to do.
Like you want to avoid obvious lying.
- Yeah, I mean, I think when people play diplomacy
for the first time,
they approach it as a game of deception and lying.
And they, ultimately, if you talk to top diplomacy players,
what they'll tell you is that diplomacy
is a game about trust
and being able to build trust in an environment
that encourages people to not trust anyone.
So that's the ultimate tension in diplomacy.
How can this AI reason
about whether you are being honest in your communication?
And how can the AI persuade you that it is being honest
when it is telling you that,
"Hey, I'm actually going to support you this turn."
- Is there some sense,
I don't know if you step back and think,
that this process will indirectly
help us study human psychology?
So like if trust is the ultimate goal,
wouldn't that help us understand
what are the fundamental aspects of forming trust
between humans and between humans and AI?
I mean, that's a really, really important question
that's much bigger than strategy games.
It's how can,
that's fundamental to the human-robot interaction problem.
How do we form trust between intelligent entities?
- So one of the things I'm really excited about
with diplomacy,
there's never really been a good domain
to investigate these kinds of questions.
And diplomacy gives us a domain
where trust is really at the center of it.
And it's not just like you've hired
a bunch of mechanical Turkers that are being paid
and trying to get through the task as quickly as possible.
You have these people that are really invested
in the outcome of the game,
and they're really trying to do the best that they can.
And so I'm really excited that we're able to,
we actually have put together this,
we're open sourcing all of our models,
we're open sourcing all of the code,
and we're making the data that we've used
available to researchers
so that they can investigate these kinds of questions.
- So the data of the different,
the human and the AI play of diplomacy,
and the models that you use
for the generation of the messages and the filtering.
- Yeah, not just even the data of the AI
playing with the humans,
but all the training data that we had,
that we use to train the AI
to understand how humans play the game.
We're setting up a system
where researchers will be able to apply
to be able to gain access to that data
and be able to use it in their own research.
- We should say, what is the name of the system?
- We're calling the bot Cicero.
- Cicero.
And what's the name, like you're open sourcing,
what's the name of the repository and the project?
Is it also just called Cicero the big project?
Or is it still coming up with a name?
- The data set comes from this website, webdiplomacy.net,
is this site that's been online for like 20 years now.
And it's one of the main sites
that people use to play diplomacy on it.
We've got like 50,000 games of diplomacy
with natural language communication,
over 10 million messages.
So it's a pretty massive data set that people can use to,
we're hoping that the academic community
and the research community is able to use it
for all sorts of interesting research questions.
- So do you, from having studied this game,
is this a sufficiently rich problem space
to explore this kind of human AI interaction?
- Yeah, absolutely.
And I think it's maybe the best data set
that I can think of out there
to investigate these kinds of questions of negotiation,
trust, persuasion.
I wouldn't say it's the best data set in the world
for human AI interaction, that's a very broad field.
But I think that it's definitely up there as like,
if you're really interested in language models
interacting with humans in a setting
where their incentives are not fully aligned,
this seems like an ideal data set for investigating that.
- So you have a paper with some impressive results
and just an impressive paper that taken this problem on.
What's the most exciting thing to you
in terms of the results from the paper?
- Well, I think there's a few--
- Ideas or results?
- Yeah, I think there's a few aspects of the results
and that I think are really exciting.
So first of all, the fact that we were able to achieve
such strong performance,
I was surprised by and pleasantly surprised by.
So we played 40 games of diplomacy with real humans
and the bot placed second out of all players
that have played five or more games.
So it's about 80 players total,
19 of whom played five or more games
and the bot was ranked second out of those players.
And the bot was really good in two dimensions.
One, being able to establish strong connections
with the other players on the board,
being able to like persuade them to work with it,
being able to coordinate with them
about like how it's going to work with them.
And then also the raw tactical
and strategic aspects of the game,
being able to understand
what the other players are likely to do,
being able to model their behavior
and respond appropriately to that,
the bot also really excelled at.
- What are some interesting things that the bot said?
By the way, are you allowed to swear in the,
like are there rules to what you're allowed to say
and not in diplomacy?
- You can say whatever you want.
I think the site will get very angry at you
if you start like threatening somebody.
And we actually--
- Like if you threaten somebody,
you're supposed to do it politely.
- Yeah, politely, you know, like keep it in character.
We actually had a researcher watching the bot 24/7,
well, whenever we play a game,
we had a bot watching it to make sure
that it wouldn't go off the rails
and start like threatening somebody or something like that.
- I would just love it if the bot started like mocking,
mocking everybody,
like some weird quirky strategies would emerge.
That have you seen anything interesting that you,
huh, that's a weird,
that's a weird behavior,
either the filter or the language model
that was weird to you.
- That was, yeah, there were definitely like
things that the bot would do
that were not in line with like
how humans would approach the game.
And that, in a good way, the humans actually,
you know, we've talked to some expert diplomacy players
about these results and their takeaway is that,
well, maybe humans are approaching this the wrong way.
And this is actually like the right way to play the game.
- So what's required to win?
§15Transcript · 1:48:35–1:56:34
Like what does it mean to mess up
or to exploit the suboptimal behavior of a player?
Like is there optimally rational behavior
and irrational behavior that you need to estimate,
that kind of stuff?
Like what stands out to you?
Like, is there a crack that you can exploit?
Is there like a weakness that you can exploit in the game
that everybody's looking for?
- Well, I think you're asking kind of two questions there.
So one, like modeling the irrationality
and the suboptimality of humans.
You can't, in diplomacy,
you can't treat all the other players like they're machines.
And if you do that,
you're going to end up playing really poorly.
And so we actually ran this experiment.
So we trained a bot in a two player,
zero-sum version of diplomacy,
the same way that you might approach a game
like chess or poker.
And the bot was superhuman.
It would crush any competitor.
And then we took that same training approach
and we trained a bot for the full seven player version
of the game through self-play without any human data.
And we stuck it in a game with six humans
and it got destroyed.
Even in the version of the game
where there's no explicit natural language communication,
it still got destroyed
because it just wouldn't be able to understand
how the other players were approaching the game
and be able to work with that.
- Can you just linger on that,
meaning like there's an individual,
there's an individual personality to each player
and then you're supposed to remember that.
But what do you mean it's not able to understand the players?
- Well, it would, for example,
expect the human to support it in a certain way
when the human would simply like,
think like, no, I'm not supposed to support you here.
It's kind of like, you know,
if you develop a self-driving car
and it's trained completely from scratch
with other self-driving cars,
it might learn to drive on the left side of the road.
And that's a totally reasonable thing to do
if you're with these other self-driving cars
that are also driving on the left side of the road.
But if you put it in an American city, it's gonna crash.
- But I guess the intuition I'm trying to build up
is why does it then crush a human player heads up
versus multiple?
- This is an aspect of two players zero sum
versus games that involve cooperation.
So in a two player zero sum game,
you can do self-play from scratch
and you will arrive at the Nash equilibrium
where you don't have to worry about the other player
playing in a very human suboptimal style.
That's just gonna be,
the only way that deviating from a Nash equilibrium
would change things is if it helped you.
- So what's the dynamic of cooperation
that's effective in diplomacy?
Do you always have to have one friend in the game?
- You always want to maximize your friends
and minimize your enemies.
- Got it.
And boy, and the lying comes into play there.
So the more friends you have, the better.
- Yeah, I mean, I guess you have to attack somebody
or else you're not gonna make progress.
- Right, so that's the tension, but this is too real.
This is too real.
This is too close to geopolitics
of actual military conflict in the world.
Okay, that's fascinating.
So that cooperation element
is what makes the game really, really hard.
- Yeah, and to give you an example
of how this suboptimality and irrationality comes into play,
there's a really common situation in the game of diplomacy
that where one player starts to win
and they're at the point where they're controlling
about half the map.
And the remaining players
who have all been fighting each other the whole game
all have to work together now
to stop this other player from winning
or else everybody's gonna lose.
And it's kind of like "Game of Thrones."
I don't know if you've seen the show
where you got the others coming from the north
and all the people have to work out their differences
and stop them from taking over.
And the bot will do this.
The bot will work with the other players
to stop the superpower from winning.
But if it's trained from scratch
or it doesn't really have a good grounding
in how humans approach it,
it will also at the same time attack the other players
with its extra units.
So all the units that are not necessary
to stop the superpower from winning,
it will use those to grab as many centers as possible
from the other players.
And in totally rational play,
the other players should just live with that.
They have to understand like,
"Hey, a score of one is better than a score of zero.
"So, okay, he's grabbed my centers,
"but I'll just deal with it."
But humans don't act that way, right?
The human gets really angry at the bot
and ends up throwing the game
because I'm gonna screw you over
because you did something that's not fair to me.
- Got it.
And are you supposed to model that?
Is the bot supposed to model that kind of human frustration?
- Yeah, exactly.
So that is something that seems almost impossible to model
purely from scratch without any human data.
It's a very cultural thing.
And so you need human data to be able to understand that,
"Hey, that's how humans behave."
And you have to work around that.
It might be suboptimal, it might be irrational,
but that's an aspect of humanity that you have to deal with.
- So how difficult is it to train on human data
given that human data is very limited
versus what a purely self-play mechanism can generate?
- That's actually one of the major challenges
that we faced in the research,
that we had a good amount of human data.
We had about 50,000 games.
What we try to do is leverage as much self-play as possible
while still leveraging the human data.
So what we do is we do self-play,
very similar to how it's been done in poker and Go,
but we try to regularize the self-play
towards the human data.
Basically, the way to think about it is
we penalize the bot for choosing actions
that are very unlikely under the human data set.
- How do you know?
Is there some kind of function that says,
"This is human-like and not?"
- Yeah, so we train a bot through supervised learning
to model the human play as much as possible.
So we basically train a neural net on those 50,000 games,
and that gives us an approximate,
that gives us a policy that resembles to some extent
how humans actually play the game.
Now, this isn't a perfect model of human play
because we don't have unlimited data.
We don't have unlimited neural net capacity,
but it gives us some approximation.
- Is there some data on the internet
that's useful besides just diplomacy?
So on the language side of things, is there some,
can you go to like Reddit?
And so sort of background model formulation
that's useful for the game of diplomacy.
- Yeah, absolutely.
And so for the language model,
which is kind of like a separate question,
we didn't use the language model during self-play training,
but we pre-trained the language model
on tons of internet data as much as possible.
And then we fine-tuned it specifically
on the diplomacy games.
So we are able to leverage the wider data set
in order to fill in some of the gaps
in how communication happens more broadly
besides just specifically in these diplomacy games.
- Okay, cool.
What are some interesting things that came to life
from this work to you?
Like what are some insights about games
where natural language is involved
and cooperation, deep cooperation is involved?
- Well, I think there's a few insights.
So first of all, the fact that you can't rely purely
or even largely on self-play,
that you really have to have an understanding
of how humans approach the game.
I think that that's one of the major conclusions
that I'm drawing from this work.
And that is, I think, applicable more broadly
to a lot of different games.
So we've actually already taken the approaches
that we've used in diplomacy and tried them
on a cooperative card game called Hanabi.
And we've had a lot of success in that game as well.
On the language side, I think the fact
§16Transcript · 1:56:39–2:05:57
that we were able to control the language model
through this intense approach was very effective.
And it allowed us, instead of just imitating
how humans would communicate, we're able to go beyond that
and able to feed into its superhuman strategies
that it can then generate messages corresponding to.
- Is there something you could say about detecting
whether a person or AI is lying or not?
- The bot doesn't explicitly try to calculate
whether somebody is lying or not.
But what it will do is try to predict
what actions they're going to take,
given the communications, given the messages
that they've sent to us.
So given our conversation,
what do I think you're going to do?
And implicitly, there is a calculation
about whether you're lying to me in that.
Based on your messages, if I think you're going
to attack me this turn, even though your messages say
that you're not, then essentially the bot
is predicting that you're lying.
But it doesn't view it as lying the same way
that we would view it as lying.
- But you could probably reformulate with all the same data
and make a classifier lying or not.
- Yeah, I think you could do that.
That was not something that we were focused on,
but I think that it is possible that,
if you came up with some measurements of like,
what does it mean to tell a lie?
Because there's a spectrum, right?
Like if you're withholding some information, is that a lie?
If you're mostly telling the truth,
but you forgot to mention this one action out of 10,
is that a lie?
It's hard to draw the line, but if you're willing to do that
and then you could possibly use it to--
- This feels like an argument inside a relationship now.
What constitutes a lie?
Depends what you mean by the definition of the word is.
Okay, still it's fascinating because trust and lying
is all intermixed into this and it's language models
that are becoming more and more sophisticated.
It's just a fascinating space to explore.
What do you see as the future of this work
that is inspired by the breakthrough performance
that you're getting here with diplomacy?
- I think there's a few different directions
to take this work.
I think really what it's showing us is the potential
that language models have.
I mean, I think a lot of people didn't think
that this kind of result was possible even today,
despite all the progress that's been made in language models.
And so it shows us how we can leverage the power
of things like self-play on top of language models
to get increasingly better performance.
And the ceiling is really much higher
than what we have right now.
- Is this transferable somehow to chatbots
for the more general task of dialogue?
So there is a kind of negotiation here,
a dance between entities that are trying to cooperate
and at the same time, a little bit adversarial,
which I think maps somewhat to the general,
the entire process of Reddit or like internet communication.
You're cooperating, you're adversarial,
you're having debates, you're having a camaraderie,
all that kind of stuff.
- I think one of the things that's really useful
about diplomacy is that we have a well-defined
value function.
There is a well-defined score that the bot
is trying to optimize.
And in a setting like a general chatbot setting,
it would need that kind of objective
in order to fully leverage the techniques
that we've developed.
- What about like what we talked about earlier
with NPCs inside video games?
Like how can it be used to create
for Elder Scrolls VI more compelling NPCs
that you could talk to instead of committing
all kinds of violence with a sword and fighting dragons,
just sitting in a tavern and drink all day
and talk to the chatbot?
- The way that we've approached AI in diplomacy
is you condition the language on an intent.
Now that intent in diplomacy is an action,
but it doesn't have to be.
And you can imagine, you could have NPCs
in video games or the metaverse or whatever,
where there's some intent or there's some objective
that they're trying to maximize,
and you can specify what that is.
And then the language can correspond to that intent.
Now, I'm not saying that this is happening imminently,
but I'm saying that this is like a future application
potentially of this direction of research.
- So what's the more general formulation of this?
Making self-play be able to scale the way self-play does
and still maintain human-like behavior.
- The way that we've approached self-play in diplomacy
is we're trying to come up with good intents
to condition the language model on.
And the space of intents is actions
that can be played in the game.
Now, there is the potential to have a broader set of intents,
things like long-term cooperation or long-term objectives
or gossip about what another player was saying.
These are things that we're currently not conditioning
the language model on, and so we're not able to control it
to say like, "Oh, you should be talking
"about this thing right now."
But it's quite possible that you could expand
the scope of intents to be able to allow it
to talk about those things.
Now, in the process of doing that,
the self-play would become much more complicated.
And so that is a potential for future work.
- Okay, the increase in the number of intents.
I still am not quite clear how you keep the self-play
integrated into the human world.
- Yeah.
- I'm a little bit loose on understanding how you do that.
- So we train in neural nets to imitate the human data
as closely as possible,
and that's what we call the anchor policy.
And now when we're doing self-play,
the problem with the anchor policy
is that it's not a perfect approximation
of how humans actually play.
Because we don't have infinite data,
because we don't have unlimited neural network capacity,
it's actually a relatively suboptimal approximation
of how humans actually play.
And we can improve that approximation
by adding planning and RL.
And so what we do is we get a better approximation,
a better model of human play by,
during the self-play process,
we say you can deviate from this human anchor policy
if there is an action that has, you know,
particularly high expected value.
But it would have to be a really high expected value
in order to deviate from this human-like policy.
So you basically say,
try to maximize your expected value
while at the same time,
stay as close as possible to the human policy.
And there is a parameter that controls
the relative weighting of those competing objectives.
- So the question I have
is how sophisticated can the anchor policy get?
So I have a policy that approximates human behavior, right?
- Yeah.
- So as you increase the number of intents,
as you generalize the space in which this is applicable,
and given that the human data is limited,
try to anticipate a policy that works
for a much larger number of cases.
Like how difficult is the process
of forming a damn good anchor policy?
- Well, it really comes down
to how much human data you have.
So it's all about scaling the human data.
- I think the more human data you have, the better.
And I think that that's going to be the major bottleneck
in scaling to more complicated domains.
But that said, there might be the potential,
just like in the language model,
where we leveraged tons of data on the internet
and then specialized it for diplomacy.
There is the future potential
that you can leverage huge amounts of data across the board
and then specialize it in the data set
that you have for diplomacy.
And that way you're essentially augmenting
the amount of data that you have.
- To what degree does this apply
to the general, the real world diplomacy, the geopolitics?
You know, there's a game theory has a history
of being applied to understand
and to give us hope about nuclear weapons, for example.
The mutually assured destruction
is a game theoretic concept that you can formulate.
Some people say it's oversimplified,
but nevertheless, here we are
and we somehow haven't blown ourselves up.
Do you see a future where this kind of system
can be used to help us make decisions,
geopolitical decisions in the world?
- Well, like I said, the original motivation
for the game of diplomacy was the failures of World War I,
the diplomatic failures that led to war.
And the real take-home message of diplomacy is that,
you know, if people approach diplomacy,
the right way, then war is ultimately unsuccessful.
The way that I see it, war is
§17Transcript · 2:06:00–2:14:41
an inherently negative sum game, right?
There's always a better outcome than war
for all the parties involved.
And my hope is that, you know, as AI progresses,
then maybe this technology could be used
to help people make better decisions across the board
and, you know, hopefully avoid
negative sum outcomes like war.
- Yeah, I mean, I just came back from Ukraine.
I'm going back there.
On deep personal levels, think a lot about
how peace can be achieved.
And I'm a big believer in conversation,
leaders getting together and having conversations
and trying to understand each other.
Yeah, it's fascinating to think
whether each one of those leaders
can run a simulation ahead of time.
Like if I'm an asshole,
(chuckles)
what are the possible consequences?
If I'm nice, what are the possible consequences?
My guess is that if the president of the United States
got together with Vladimir Zelensky and Vladimir Putin,
that there would be significant benefits
to the president of the United States not having the ego
of kind of playing down, of giving away a lot of chips
for the future success of a world.
So giving a lot of power to the two presidents
of the competing nations to achieve peace.
That's my guess,
but it'd be nice to run a bunch of simulations.
But then you have to have human data, right?
You really, 'cause it's like the game of diplomacy
is fundamentally different than geopolitics.
You need data.
You need like, I guess that's the question I have.
Like how transferable is this to,
like I don't know, any kind of negotiation, right?
Like to any kind of, some local, I don't know,
a bunch of lawyers like arguing,
like a divorce, like divorce lawyers.
Like how transferable is this
to all kinds of human negotiation?
- Well, I feel like this isn't a question
that's unique to diplomacy.
I mean, I think you look at RL breakthroughs,
reinforcement learning breakthroughs
in previous games as well,
like AI for StarCraft, AI for Atari.
You haven't really seen it deployed in the real world
because you have these problems of,
it's really hard to collect a lot of data
and you don't have a well-defined action space.
You don't have a well-defined reward function.
These are all things that you really need
for reinforcement learning
and planning to be really successful today.
Now, there are some domains where you do have that.
Code generation is one example.
Theorem proving mathematics, that's another example
where you have a well-defined action space.
You have a well-defined reward function.
And those are the kinds of domains
where I can see RL in the short term
being incredibly powerful.
But yeah, I think that those are the barriers
to deploying this at scale in the real world.
But the hope is that in the long run,
we'll be able to get there.
- Yeah, but see, diplomacy feels like closer
to the real world than does StarCraft.
Like 'cause it's natural language, right?
You're operating in the space of intents
and in the space of natural language,
that feels very close to the real world.
And it also feels like you could get data on that
from the internet.
- Yeah, and that's why I do think that diplomacy
is taking a big step closer to the real world
than anything that's came before
in terms of game AI breakthroughs.
The fact that we're communicating in natural language,
we're leveraging the fact that we have this
like general data set of dialogue and communication
from a breadth of the internet.
That is a big step in that direction.
We're not 100% there, but we're getting closer at least.
- So if we actually return back to poker and chess,
are some of the ideas that you're learning here
with diplomacy, could you construct AI systems
that play like humans?
Like make for a fun opponent in a game of chess?
- Yeah, absolutely.
We've already started looking into this direction a bit.
So we tried to use the techniques that we've developed
for diplomacy to make chess and go AIs.
And what we found is that it led to much more human-like
strong chess and go players.
The way that AIs like Stockfish today play
is in a very inhuman style.
It's very strong, but it's very different
from how humans play.
And so we can take the techniques
that we've developed for diplomacy.
We do something similar in chess and go,
and we end up with a bot that's both strong and human-like.
To elaborate on this a bit,
like one way to approach making a human-like
AI for chess is to collect a bunch of human games,
like a bunch of human grandmaster games,
and just to supervise learning on those games.
But the problem is that if you do that,
what you end up with is an AI that's substantially weaker
than the human grandmasters that you trained on.
Because the neural net is not able to approximate
the nuance of the strategy.
This goes back to the planning thing that I mentioned,
the search thing that I talked about before,
that these human grandmasters, when they're playing,
they're using search and they're using planning.
And the neural net alone,
unless you have a massive neural net
that's like a thousand times bigger
than what we have right now,
it's not able to approximate those details very effectively.
And on the other hand,
you can leverage search and planning very heavily,
but then what you end up with is an AI
that plays in a very different style
from how humans play the game.
Now, if you strike this intermediate balance
by setting the regularization parameters correctly
and say, you can do planning,
but try to keep it close to the human policy,
then you end up with an AI that plays
in both a very human-like style and a very strong style.
And you can actually even tune it
to have a certain ELO rating.
So you can say, play in the style of like a 2800 ELO human.
- I wonder if you could do specific type of humans
or categories of humans, not just skill, but style.
- Yeah, I think so.
And so this is where the research gets interesting.
Like, one of the things that I was thinking about is,
and this is actually already being done,
there's a researcher at the University of Toronto
that's working on this,
is to make an AI that plays
in the style of a particular player.
Like Magnus Carlsen, for example,
you can make an AI that plays like Magnus Carlsen.
And then where I think this gets interesting is like,
maybe you're up against Magnus Carlsen
in the world championship or something,
you can play against this Magnus Carlsen bot
to prepare against the real Magnus Carlsen.
And you can try to explore strategies
that he might struggle with and try to figure out like,
how do you beat this player in particular?
On the other hand, you can also have Magnus Carlsen
working with this bot to try to figure out where he's weak
and where he needs to improve his strategy.
And so I can envision this future
where data on specific chess and Go players
becomes extremely valuable because you can use that data
to create specific models
of how these particular players play.
- So increasingly human-like behavior in bots, however,
as you've mentioned, makes cheating,
cheat detection much harder.
- It does, yeah.
The way that cheat detection works in a game like poker
and a game like chess and Go, from what I understand,
is trying to see like, is this person making moves
that are very common among chess AIs or AIs in general?
But very uncommon among top human players.
And if you have the development of these AIs
that play in a very strong style,
but also a very human-like style,
then that poses serious challenges for cheat detection.
- And it makes you now ask yourself a hard question
about what is the role of AI systems
as they become more and more integrated in our society?
And this kind of human AI integration
has some deep ethical issues that we should be aware of.
And also it's a kind of cybersecurity challenge, right?
For it to make, you know,
one of the assumptions we have when we play games
is that there's a trust that it's only humans involved.
And the better AI systems we create,
which makes it super exciting,
human-like AI systems with different styles of humans
is really exciting,
but then we have to have the defenses
better and better and better
if we're to trust that we can enjoy
human versus human game in a deeply fair way.
It's fascinating.
It's just, it's humbling.
§18Transcript · 2:14:44–2:22:50
- Yeah, I think there's a lot of like negative potential
for this kind of technology,
but you know, at the same time,
there's a lot of upside for it as well.
So, you know, for example, right now,
it's really hard to learn how to get better
in games like chess and poker and Go
because the way that the AI plays
is so foreign and incomprehensible.
But if you have these AIs that are playing,
you know, you can say like,
oh, I'm a 2000 Elo human, how do I get to 2200?
Now you can have an AI that plays in the style
of a 2200 Elo human, and that will help you get better.
Or, you know, you mentioned this problem of like,
how do you know that you're actually playing with humans
when you're playing like online and in video games?
Well, now we have the potential of populating
these like virtual worlds with agents,
like AI agents that are actually fun to play with,
and you don't have to always be playing with other humans
to, you know, have a fun time.
So yeah, a lot of upside potential too.
And I think, you know, with any sort of tool,
there's the potential for a lot of greatness
and a lot of downsides as well.
- So in the paper that I got a chance to look at,
there's a section on ethical considerations.
What's in that section?
What are some ethical considerations here?
Is it some of the stuff we already talked about?
- There's some things that we've already talked about.
I think specific to diplomacy, you know,
there's also the challenge that the game is,
you know, there is a deception aspect to the game.
And so, you know, developing language models
that are capable of deception is I think a dicey issue
and something that, you know,
makes research on diplomacy particularly challenging.
And, you know, so those kinds of issues of like,
should we even be developing AIs
that are capable of lying to people?
That's something that we have to, you know,
think carefully about.
- That's so cool.
I mean, you have to do that kind of stuff
in order to figure out where the ethical lines are.
But I can see in the future it being illegal
to have a consumer product that lies.
- Yeah, yeah.
- Like your personal assistant AI system
is always have to tell the truth.
But if I ask it, do I look,
did I get fatter over the past month?
I sure as hell want that AI system to lie to me.
So there's a trade off between lying and being nice.
We have to somehow find,
what is the ethics in that?
And we're back to discussions inside relationships.
Anyway, what were you saying?
- Oh, yeah, I was saying like, yeah,
that's kind of going to the question of like,
what is a lie?
You know, is a white lie a bad lie?
Is it an ethical lie?
You know, those kinds of questions.
- Boy, we return time and time again
to deep human questions as we design AI systems.
That's exactly what they do.
They put a mirror to humanity
to help us understand ourselves.
- There's also the issue of like, you know,
in these diplomacy experiments,
in order to do a fair comparison,
you know, what we found is that
there's an inherent anti-AI bias in these kinds of games.
So we actually played a tournament
in a non-language version of the game,
where, you know, we told the participants like,
hey, in every single game, there's going to be an AI.
And what we found is that the humans
would spend basically the entire game,
like trying to figure out who the bot was.
And then as soon as they thought they figured it out,
they would all team up and try to kill it.
And, you know, overcoming that inherent anti-AI bias
is a challenge.
- On the flip side, I think when robots become the enemy,
that's when we get to heal our human divisions,
and then we can become one.
As long as we have one enemy,
it's that Reagan thing when the aliens show up,
that's when we put our side,
our divisions will become one human species.
- Right, we might have our differences,
but we're at least all human.
- At least we all hate the robots.
No, no, no, no.
I think there will be actually in the future
something like a civil rights movement for robots.
I think that's the fascinating thing about AI systems,
and that is they ask, they force us to ask about
ethical questions about what is sentience,
what is, how do we feel about systems
that are capable of suffering,
or capable of displaying suffering?
And how do we design products that show emotion and not?
How do we feel about that?
Lying is another topic.
Are we going to allow bots to lie and not?
And where's the balance between being nice
and telling the truth?
I mean, these are all fascinating human questions,
and it's so exciting to be in a century
where we can create systems that
take these philosophical questions
that have been asked for centuries,
and now we can engineer them inside systems,
where you really have to answer them,
because you'll have transformational impact on human society
depending on what you design inside those systems.
It's fascinating.
And like you said, I feel like diplomacy
is a step towards the direction of the real world,
applying these RL methods towards the real world.
From all the breakthrough performances
in Go and Chess and StarCraft and Dota,
this feels like the real world.
Especially now my mind's been on war,
and military conflict.
This feels like it can give us some deep insights
about human behavior at the large geopolitical scale.
What do you think is the breakthrough
or the directions of work that will take us
towards solving intelligence, towards creating AGI systems?
You've been a part of creating,
by the way, we should say a part of great teams
that do this, of creating systems
that achieve breakthrough performances
on before thought unsolvable problems,
like poker, multiplayer poker, diplomacy.
We're taking steps towards that direction.
What do you think it takes to go all the way
to create superhuman level intelligence?
- You know, there's a lot of people
trying to figure that out right now.
And I should say, the amount of progress that's been made,
especially in the past few years is truly phenomenal.
I mean, you look at where AI was 10 years ago,
and the idea that you can have AIs
that can generate language and generate images
the way they're doing today,
and able to play a game like diplomacy
was just unthinkable, even five years ago,
let alone 10 years ago.
Now, there are aspects of AI that I think are still lacking.
I think there's general agreement
that one of the major issues with AI today
is that it's very data inefficient.
It requires a huge number of samples of training examples
to be able to train.
You know, you look at an AI that plays Go,
and it needs millions of games of Go
to learn how to play the game well.
Whereas a human can pick it up in like, you know,
I don't know, how many games does a human Go player,
Go grandmaster play in their lifetime?
Probably, you know, in the thousands
or tens of thousands, I guess.
So that's one issue.
- Data efficiency.
- Overcoming this challenge of data efficiency.
And this is particularly important
if we want to deploy AI systems in real world settings
where they're interacting with humans,
because, you know, for example, with robotics,
it's really hard to generate a huge number of samples.
It's a different story when you're working in these,
you know, totally virtual games
where you can play a million games and it's no big deal.
- I was planning on just launching like a thousand
of these robots in Austin.
I don't think it's illegal for legged robots
to roam the streets and just collect data.
- That's not a crazy idea.
- Of course, the worst that could happen.
- Yeah, I mean, that's one way
to overcome the data efficiency problem.
It's like scale it, yeah.
- Like I actually tried to see if there's a law
against robots, like legged robots just operating
in the streets of a major city and there isn't,
I couldn't find any.
So I'll take it all the way to the Supreme Court.
Robot rights.
Okay, anyway, sorry, you were saying.
So what are the ideas for getting,
becoming more data efficient?
- I mean, that's the trillion dollar question in AI today.
I mean, if you can figure out how to make AI systems
§19Transcript · 2:22:52–2:29:15
more data efficient, then that's a huge breakthrough.
So nobody really knows right now.
- It could be just a gigantic background model,
language model, and then you do,
the training becomes like prompting that model
to essentially do a kind of querying,
a search into the space of the things it's learned
to customize that to whatever problem you're trying to solve.
So maybe if you form a large enough language model,
you can go quite a long way.
- I think there's some truth to that.
I mean, you look at the way humans approach
a game like poker, they're not coming at it from scratch.
They're coming at it with a huge amount
of background knowledge about how humans work,
how the world works, the idea of money.
So they're able to leverage that kind of information
to pick up the game faster.
So it's not really a fair comparison
to then compare it to an AI
that's like learning from scratch.
And maybe one of the ways that we address
this sample complexity problem is by allowing AIs
to leverage that general knowledge
across a ton of different domains.
- So, like I said, you did a lot of incredible work
in the space of research and actually building systems.
What advice would you give to, let's start with beginners.
What advice would you give to beginners
interested in machine learning?
Just they're at the very start of their journey,
they're in high school and college,
thinking like this seems like a fascinating world.
What advice would you give them?
- I would say that there are a lot of people
working on similar aspects of machine learning
and to not be afraid to try something a bit different.
My own path in AI is pretty atypical
for a machine learning researcher today.
I mean, I started out working on game theory
and then shifting more towards reinforcement learning
as time went on.
And that actually had a lot of benefits, I think,
because it allowed me to look at these problems
in a very different way
from the way a lot of machine learning researchers view it.
And that comes with drawbacks in some respects.
Like I think there's definitely aspects of machine learning
where I'm weaker than most of the researchers out there,
but I think that diversity of perspective,
when I'm working with my teammates,
there's something that I'm bringing to the table
and there's something that they're bringing to the table.
And that kind of collaboration
becomes very fruitful for that reason.
- So there could be problems like poker,
like you've chosen diplomacy,
there could be problems like that still out there
that you can just tackle,
even if it seems extremely difficult.
- I think that there's a lot of challenges left.
And I think having a diversity of viewpoints
and backgrounds is really helpful
for working together to figure out
how to tackle those kinds of challenges.
- So as a beginner, so that,
I would say that's more for like a grad student
where they already built up a base,
like a complete beginner, what's a good journey.
So for you that was doing some more
in the math side of things, doing game theory,
all that, so it's basically build up a foundation
in something, so programming, mathematics,
it could even be physics, but build that foundation.
- Yeah, I would say build a strong foundation
in math and computer science and statistics
and these kinds of areas,
but don't be afraid to try something that's different
and learn something that's different
from the thing that everybody else is doing
to get into machine learning.
There's value in having a different background
than everybody else.
Yeah, so, but certainly having a strong math background,
especially in things like linear algebra
and statistics and probability
are incredibly helpful today for learning about
and understanding machine learning.
- Do you think one day we'll be able to,
since you're taking steps from poker to diplomacy,
one day we'll be able to figure out
how to live life optimally?
- Well, what is it, like in poker and diplomacy,
you need a value function, you need to have a reward system.
And so what does it mean to live a life that's optimal?
- So, okay, so then you can exactly like lay down
a reward function being like, I wanna be rich
or I want to be in a happy relationship.
And then you'll say, well, do X.
- There's a lot of talk today about in AI safety circles
about like misspecification of reward function.
So you say like, okay, my objective is to be rich
and maybe the AI tells you like, okay,
well, if you wanna maximize the probability
that you're rich, go rob a bank.
- Sure.
- And so you wanna, is that really what you want?
Is your objective really to be rich at all costs
or is it more nuanced than that?
- So the unintended consequences, yeah.
Yeah, so maybe life is more about defining
the reward function that minimizes
the unintended consequences than it is about
the actual policy that gets you to the reward function.
Maybe life is just about constantly updating
the reward function.
I think one of the challenges in life
is figuring out exactly what that reward function is.
Sometimes it's pretty hard to specify.
The same way that trying to handcraft the optimal policy
in a game like chess is really difficult.
It's not so clear cut what the reward function is for life.
- I think one day AI will figure it out.
And I wonder what that would be.
Until then, I just really appreciate
the kind of work you're doing.
And it's really fascinating taking a leap
into a more and more real-world-like problem space
and just achieving incredible results
by applying reinforcement learning.
Now, since I saw your work on poker,
you've been a constant inspiration.
It's an honor to get to finally talk to you
and this is really fun.
- Thanks for having me.
- Thanks for listening to this conversation with Noah Brown.
To support this podcast, please check out our sponsors
in the description.
And now, let me leave you with some words from Sun Tzu
in "The Art of War."
"The whole secret lies in confusing the enemy
so that he cannot fathom our real intent."
Thank you for listening and hope to see you next time.
(upbeat music)
(upbeat music)