WEBVTT

00:00.000 --> 00:03.440
 The following is a conversation with Thomas Sanholm.

00:03.440 --> 00:06.880
 He's a professor at CMU and co creator of Labratus,

00:06.880 --> 00:09.880
 which is the first AI system to beat top human players

00:09.880 --> 00:13.000
 in the game of Heads Up No Limit Texas Holdem.

00:13.000 --> 00:15.600
 He has published over 450 papers

00:15.600 --> 00:17.320
 on game theory and machine learning,

00:17.320 --> 00:21.120
 including a best paper in 2017 at NIPS,

00:21.120 --> 00:23.560
 now renamed to Newrips,

00:23.560 --> 00:27.040
 which is where I caught up with him for this conversation.

00:27.040 --> 00:30.680
 His research and companies have had wide reaching impact

00:30.680 --> 00:32.160
 in the real world,

00:32.160 --> 00:34.400
 especially because he and his group

00:34.400 --> 00:36.640
 not only propose new ideas,

00:36.640 --> 00:40.440
 but also build systems to prove that these ideas work

00:40.440 --> 00:42.120
 in the real world.

00:42.120 --> 00:44.640
 This conversation is part of the MIT course

00:44.640 --> 00:46.440
 on artificial general intelligence

00:46.440 --> 00:49.040
 and the artificial intelligence podcast.

00:49.040 --> 00:52.400
 If you enjoy it, subscribe on YouTube, iTunes,

00:52.400 --> 00:54.320
 or simply connect with me on Twitter

00:54.320 --> 00:58.080
 at Lex Friedman, spelled F R I D.

00:58.080 --> 01:02.120
 And now here's my conversation with Thomas Sanholm.

01:03.080 --> 01:06.120
 Can you describe at the high level

01:06.120 --> 01:09.320
 the game of poker, Texas Holdem, Heads Up Texas Holdem

01:09.320 --> 01:13.280
 for people who might not be familiar with this card game?

01:13.280 --> 01:14.440
 Yeah, happy to.

01:14.440 --> 01:16.520
 So Heads Up No Limit Texas Holdem

01:16.520 --> 01:18.840
 has really emerged in the AI community

01:18.840 --> 01:21.360
 as a main benchmark for testing these

01:21.360 --> 01:23.560
 application independent algorithms

01:23.560 --> 01:26.440
 for imperfect information game solving.

01:26.440 --> 01:30.960
 And this is a game that's actually played by humans.

01:30.960 --> 01:33.960
 You don't see that much on TV or casinos

01:33.960 --> 01:36.160
 because well, for various reasons,

01:36.160 --> 01:40.240
 but you do see it in some expert level casinos

01:40.240 --> 01:43.080
 and you see it in the best poker movies of all time.

01:43.080 --> 01:45.720
 It's actually an event in the World Series of Poker,

01:45.720 --> 01:48.200
 but mostly it's played online

01:48.200 --> 01:50.880
 and typically for pretty big sums of money.

01:50.880 --> 01:54.560
 And this is a game that usually only experts play.

01:54.560 --> 01:58.720
 So if you go to your home game on a Friday night,

01:58.720 --> 02:01.280
 it probably is not gonna be Heads Up No Limit Texas Holdem.

02:01.280 --> 02:04.640
 It might be No Limit Texas Holdem in some cases,

02:04.640 --> 02:08.720
 but typically for a big group and it's not as competitive.

02:08.720 --> 02:10.520
 While Heads Up means it's two players.

02:10.520 --> 02:13.360
 So it's really like me against you.

02:13.360 --> 02:14.680
 Am I better or are you better?

02:14.680 --> 02:17.520
 Much like chess or go in that sense,

02:17.520 --> 02:19.520
 but an imperfect information game,

02:19.520 --> 02:21.520
 which makes it much harder because I have to deal

02:21.520 --> 02:25.560
 with issues of you knowing things that I don't know

02:25.560 --> 02:27.200
 and I know things that you don't know

02:27.200 --> 02:29.720
 instead of pieces being nicely laid on the board

02:29.720 --> 02:31.120
 for both of us to see.

02:31.120 --> 02:34.840
 So in Texas Holdem, there's two cards

02:34.840 --> 02:37.440
 that you only see that belong to you.

02:37.440 --> 02:38.520
 Yeah. And there is,

02:38.520 --> 02:40.400
 they gradually lay out some cards

02:40.400 --> 02:44.080
 that add up overall to five cards that everybody can see.

02:44.080 --> 02:45.720
 Yeah. So the imperfect nature

02:45.720 --> 02:47.560
 of the information is the two cards

02:47.560 --> 02:48.400
 that you're holding in your hand.

02:48.400 --> 02:49.380
 Up front, yeah.

02:49.380 --> 02:51.840
 So as you said, you first get two cards

02:51.840 --> 02:55.200
 in private each and then there's a betting round.

02:55.200 --> 02:58.320
 Then you get three cards in public on the table.

02:58.320 --> 02:59.240
 Then there's a betting round.

02:59.240 --> 03:01.680
 Then you get the fourth card in public on the table.

03:01.680 --> 03:02.580
 There's a betting round.

03:02.580 --> 03:04.920
 Then you get the 5th card on the table.

03:04.920 --> 03:05.760
 There's a betting round.

03:05.760 --> 03:07.480
 So there's a total of four betting rounds

03:07.480 --> 03:11.140
 and four tranches of information revelation if you will.

03:11.140 --> 03:14.120
 The only the first tranche is private

03:14.120 --> 03:16.520
 and then it's public from there.

03:16.520 --> 03:21.520
 And this is probably by far the most popular game in AI

03:24.040 --> 03:26.380
 and just the general public

03:26.380 --> 03:28.400
 in terms of imperfect information.

03:28.400 --> 03:32.520
 So that's probably the most popular spectator game

03:32.520 --> 03:33.400
 to watch, right?

03:33.400 --> 03:37.260
 So, which is why it's a super exciting game to tackle.

03:37.260 --> 03:40.480
 So it's on the order of chess, I would say,

03:40.480 --> 03:43.680
 in terms of popularity, in terms of AI setting it

03:43.680 --> 03:46.360
 as the bar of what is intelligence.

03:46.360 --> 03:50.400
 So in 2017, Labratus, how do you pronounce it?

03:50.400 --> 03:51.220
 Labratus.

03:51.220 --> 03:52.060
 Labratus.

03:52.060 --> 03:52.900
 Labratus beats.

03:52.900 --> 03:54.080
 A little Latin there.

03:54.080 --> 03:55.520
 A little bit of Latin.

03:55.520 --> 04:00.520
 Labratus beats a few, four expert human players.

04:01.040 --> 04:03.080
 Can you describe that event?

04:03.080 --> 04:04.060
 What you learned from it?

04:04.060 --> 04:04.900
 What was it like?

04:04.900 --> 04:06.860
 What was the process in general

04:06.860 --> 04:09.960
 for people who have not read the papers and the study?

04:09.960 --> 04:12.920
 Yeah, so the event was that we invited

04:12.920 --> 04:14.840
 four of the top 10 players,

04:14.840 --> 04:17.080
 with these specialist players in Heads Up No Limit,

04:17.080 --> 04:19.080
 Texas Holden, which is very important

04:19.080 --> 04:21.400
 because this game is actually quite different

04:21.400 --> 04:23.900
 than the multiplayer version.

04:23.900 --> 04:25.680
 We brought them in to Pittsburgh

04:25.680 --> 04:28.920
 to play at the Reverse Casino for 20 days.

04:28.920 --> 04:31.840
 We wanted to get 120,000 hands in

04:31.840 --> 04:36.160
 because we wanted to get statistical significance.

04:36.160 --> 04:39.040
 So it's a lot of hands for humans to play,

04:39.040 --> 04:42.840
 even for these top pros who play fairly quickly normally.

04:42.840 --> 04:46.400
 So we couldn't just have one of them play so many hands.

04:46.400 --> 04:50.400
 20 days, they were playing basically morning to evening.

04:50.400 --> 04:55.400
 And I raised 200,000 as a little incentive for them to play.

04:55.660 --> 05:00.060
 And the setting was so that they didn't all get 50,000.

05:01.080 --> 05:02.640
 We actually paid them out

05:02.640 --> 05:05.480
 based on how they did against the AI each.

05:05.480 --> 05:09.440
 So they had an incentive to play as hard as they could,

05:09.440 --> 05:11.160
 whether they're way ahead or way behind

05:11.160 --> 05:13.760
 or right at the mark of beating the AI.

05:13.760 --> 05:16.000
 And you don't make any money, unfortunately.

05:16.000 --> 05:17.920
 Right, no, we can't make any money.

05:17.920 --> 05:20.320
 So originally, a couple of years earlier,

05:20.320 --> 05:24.080
 I actually explored whether we could actually play for money

05:24.080 --> 05:28.000
 because that would be, of course, interesting as well,

05:28.000 --> 05:29.520
 to play against the top people for money.

05:29.520 --> 05:33.040
 But the Pennsylvania Gaming Board said no, so we couldn't.

05:33.040 --> 05:35.520
 So this is much like an exhibit,

05:36.400 --> 05:39.760
 like for a musician or a boxer or something like that.

05:39.760 --> 05:41.600
 Nevertheless, they were keeping track of the money

05:41.600 --> 05:46.600
 and brought us close to $2 million, I think.

05:48.200 --> 05:51.840
 So if it was for real money, if you were able to earn money,

05:51.840 --> 05:55.360
 that was a quite impressive and inspiring achievement.

05:55.360 --> 05:59.280
 Just a few details, what were the players looking at?

05:59.280 --> 06:00.460
 Were they behind a computer?

06:00.460 --> 06:02.080
 What was the interface like?

06:02.080 --> 06:05.240
 Yes, they were playing much like they normally do.

06:05.240 --> 06:07.200
 These top players, when they play this game,

06:07.200 --> 06:08.680
 they play mostly online.

06:08.680 --> 06:11.640
 So they're used to playing through a UI.

06:11.640 --> 06:13.280
 And they did the same thing here.

06:13.280 --> 06:14.520
 So there was this layout.

06:14.520 --> 06:17.920
 You could imagine there's a table on a screen.

06:17.920 --> 06:20.080
 There's the human sitting there,

06:20.080 --> 06:21.720
 and then there's the AI sitting there.

06:21.720 --> 06:24.560
 And the screen shows everything that's happening.

06:24.560 --> 06:27.480
 The cards coming out and shows the bets being made.

06:27.480 --> 06:29.940
 And we also had the betting history for the human.

06:29.940 --> 06:33.320
 So if the human forgot what had happened in the hand so far,

06:33.320 --> 06:37.240
 they could actually reference back and so forth.

06:37.240 --> 06:39.480
 Is there a reason they were given access

06:39.480 --> 06:41.200
 to the betting history for?

06:41.200 --> 06:45.860
 Well, we just, it didn't really matter.

06:45.860 --> 06:47.360
 They wouldn't have forgotten anyway.

06:47.360 --> 06:48.800
 These are top quality people.

06:48.800 --> 06:51.300
 But we just wanted to put out there

06:51.300 --> 06:53.460
 so it's not a question of the human forgetting

06:53.460 --> 06:55.320
 and the AI somehow trying to get advantage

06:55.320 --> 06:56.760
 of better memory.

06:56.760 --> 06:57.640
 So what was that like?

06:57.640 --> 06:59.720
 I mean, that was an incredible accomplishment.

06:59.720 --> 07:02.760
 So what did it feel like before the event?

07:02.760 --> 07:05.640
 Did you have doubt, hope?

07:05.640 --> 07:08.160
 Where was your confidence at?

07:08.160 --> 07:09.240
 Yeah, that's great.

07:09.240 --> 07:10.160
 So great question.

07:10.160 --> 07:14.200
 So 18 months earlier, I had organized a similar brains

07:14.200 --> 07:17.840
 versus AI competition with a previous AI called Cloudyco

07:17.840 --> 07:20.560
 and we couldn't beat the humans.

07:20.560 --> 07:23.800
 So this time around, it was only 18 months later.

07:23.800 --> 07:27.820
 And I knew that this new AI, Libratus, was way stronger,

07:27.820 --> 07:31.360
 but it's hard to say how you'll do against the top humans

07:31.360 --> 07:32.440
 before you try.

07:32.440 --> 07:35.160
 So I thought we had about a 50, 50 shot.

07:35.160 --> 07:38.880
 And the international betting sites put us

07:38.880 --> 07:41.800
 as a four to one or five to one underdog.

07:41.800 --> 07:44.700
 So it's kind of interesting that people really believe

07:44.700 --> 07:48.440
 in people and over AI, not just people.

07:48.440 --> 07:50.720
 People don't just over believe in themselves,

07:50.720 --> 07:53.280
 but they have overconfidence in other people as well

07:53.280 --> 07:55.440
 compared to the performance of AI.

07:55.440 --> 07:59.120
 And yeah, so we were a four to one or five to one underdog.

07:59.120 --> 08:02.880
 And even after three days of beating the humans in a row,

08:02.880 --> 08:06.520
 we were still 50, 50 on the international betting sites.

08:06.520 --> 08:09.040
 Do you think there's something special and magical

08:09.040 --> 08:12.160
 about poker and the way people think about it,

08:12.160 --> 08:14.600
 in the sense you have,

08:14.600 --> 08:17.320
 I mean, even in chess, there's no Hollywood movies.

08:17.320 --> 08:21.200
 Poker is the star of many movies.

08:21.200 --> 08:26.200
 And there's this feeling that certain human facial

08:26.640 --> 08:30.760
 expressions and body language, eye movement,

08:30.760 --> 08:33.360
 all these tells are critical to poker.

08:33.360 --> 08:35.000
 Like you can look into somebody's soul

08:35.000 --> 08:37.880
 and understand their betting strategy and so on.

08:37.880 --> 08:41.520
 So that's probably why, possibly,

08:41.520 --> 08:43.640
 do you think that is why people have a confidence

08:43.640 --> 08:45.640
 that humans will outperform?

08:45.640 --> 08:48.920
 Because AI systems cannot, in this construct,

08:48.920 --> 08:51.040
 perceive these kinds of tells.

08:51.040 --> 08:53.200
 They're only looking at betting patterns

08:53.200 --> 08:58.200
 and nothing else, betting patterns and statistics.

08:58.200 --> 09:02.200
 So what's more important to you

09:02.200 --> 09:06.120
 if you step back on human players, human versus human?

09:06.120 --> 09:08.600
 What's the role of these tells,

09:08.600 --> 09:11.880
 of these ideas that we romanticize?

09:11.880 --> 09:15.480
 Yeah, so I'll split it into two parts.

09:15.480 --> 09:20.480
 So one is why do humans trust humans more than AI

09:20.480 --> 09:22.600
 and have overconfidence in humans?

09:22.600 --> 09:25.920
 I think that's not really related to the tell question.

09:25.920 --> 09:28.600
 It's just that they've seen these top players,

09:28.600 --> 09:31.040
 how good they are, and they're really fantastic.

09:31.040 --> 09:36.040
 So it's just hard to believe that an AI could beat them.

09:36.040 --> 09:37.680
 So I think that's where that comes from.

09:37.680 --> 09:40.600
 And that's actually maybe a more general lesson about AI.

09:40.600 --> 09:43.200
 That until you've seen it overperform a human,

09:43.200 --> 09:45.080
 it's hard to believe that it could.

09:45.080 --> 09:50.080
 But then the tells, a lot of these top players,

09:50.560 --> 09:52.760
 they're so good at hiding tells

09:52.760 --> 09:56.240
 that among the top players,

09:56.240 --> 09:59.480
 it's actually not really worth it

09:59.480 --> 10:01.200
 for them to invest a lot of effort

10:01.200 --> 10:03.160
 trying to find tells in each other

10:03.160 --> 10:05.640
 because they're so good at hiding them.

10:05.640 --> 10:09.840
 So yes, at the kind of Friday evening game,

10:09.840 --> 10:11.800
 tells are gonna be a huge thing.

10:11.800 --> 10:13.160
 You can read other people.

10:13.160 --> 10:14.120
 And if you're a good reader,

10:14.120 --> 10:16.440
 you'll read them like an open book.

10:16.440 --> 10:18.280
 But at the top levels of poker now,

10:18.280 --> 10:21.960
 the tells become a much smaller and smaller aspect

10:21.960 --> 10:24.480
 of the game as you go to the top levels.

10:24.480 --> 10:28.120
 The amount of strategies, the amount of possible actions

10:28.120 --> 10:33.120
 is very large, 10 to the power of 100 plus.

10:35.400 --> 10:37.880
 So there has to be some, I've read a few of the papers

10:37.880 --> 10:42.080
 related, it has to form some abstractions

10:42.080 --> 10:44.040
 of various hands and actions.

10:44.040 --> 10:47.560
 So what kind of abstractions are effective

10:47.560 --> 10:49.200
 for the game of poker?

10:49.200 --> 10:50.880
 Yeah, so you're exactly right.

10:50.880 --> 10:55.360
 So when you go from a game tree that's 10 to the 161,

10:55.360 --> 10:58.000
 especially in an imperfect information game,

10:58.000 --> 11:00.200
 it's way too large to solve directly,

11:00.200 --> 11:03.280
 even with our fastest equilibrium finding algorithms.

11:03.280 --> 11:07.200
 So you wanna abstract it first.

11:07.200 --> 11:10.920
 And abstraction in games is much trickier

11:10.920 --> 11:15.440
 than abstraction in MDPs or other single agent settings.

11:15.440 --> 11:17.760
 Because you have these abstraction pathologies

11:17.760 --> 11:19.880
 that if I have a finer grained abstraction,

11:19.880 --> 11:23.240
 the strategy that I can get from that for the real game

11:23.240 --> 11:25.240
 might actually be worse than the strategy

11:25.240 --> 11:27.160
 I can get from the coarse grained abstraction.

11:27.160 --> 11:28.760
 So you have to be very careful.

11:28.760 --> 11:31.080
 Now the kinds of abstractions, just to zoom out,

11:31.080 --> 11:34.480
 we're talking about, there's the hands abstractions

11:34.480 --> 11:37.280
 and then there's betting strategies.

11:37.280 --> 11:38.600
 Yeah, betting actions, yeah.

11:38.600 --> 11:39.440
 Baiting actions.

11:39.440 --> 11:41.640
 So there's information abstraction,

11:41.640 --> 11:44.720
 don't talk about general games, information abstraction,

11:44.720 --> 11:47.560
 which is the abstraction of what chance does.

11:47.560 --> 11:50.080
 And this would be the cards in the case of poker.

11:50.080 --> 11:52.480
 And then there's action abstraction,

11:52.480 --> 11:57.000
 which is abstracting the actions of the actual players,

11:57.000 --> 11:59.560
 which would be bets in the case of poker.

11:59.560 --> 12:01.320
 Yourself and the other players?

12:01.320 --> 12:03.680
 Yes, yourself and other players.

12:03.680 --> 12:08.280
 And for information abstraction,

12:08.280 --> 12:11.160
 we were completely automated.

12:11.160 --> 12:13.840
 So these are algorithms,

12:13.840 --> 12:16.760
 but they do what we call potential aware abstraction,

12:16.760 --> 12:19.000
 where we don't just look at the value of the hand,

12:19.000 --> 12:20.840
 but also how it might materialize

12:20.840 --> 12:22.560
 into good or bad hands over time.

12:22.560 --> 12:25.280
 And it's a certain kind of bottom up process

12:25.280 --> 12:27.640
 with integer programming there and clustering

12:27.640 --> 12:31.480
 and various aspects, how do you build this abstraction?

12:31.480 --> 12:34.400
 And then in the action abstraction,

12:34.400 --> 12:39.400
 there it's largely based on how humans and other AIs

12:40.520 --> 12:42.320
 have played this game in the past.

12:42.320 --> 12:43.880
 But in the beginning,

12:43.880 --> 12:47.680
 we actually used an automated action abstraction technology,

12:47.680 --> 12:50.240
 which is provably convergent

12:51.240 --> 12:54.040
 that it finds the optimal combination of bet sizes,

12:54.040 --> 12:55.480
 but it's not very scalable.

12:55.480 --> 12:57.280
 So we couldn't use it for the whole game,

12:57.280 --> 12:59.880
 but we use it for the first couple of betting actions.

12:59.880 --> 13:03.080
 So what's more important, the strength of the hand,

13:03.080 --> 13:08.080
 so the information abstraction or the how you play them,

13:09.320 --> 13:11.640
 the actions, does it, you know,

13:11.640 --> 13:13.200
 the romanticized notion again,

13:13.200 --> 13:15.600
 is that it doesn't matter what hands you have,

13:15.600 --> 13:19.240
 that the actions, the betting may be the way you win

13:19.240 --> 13:20.320
 no matter what hands you have.

13:20.320 --> 13:23.280
 Yeah, so that's why you have to play a lot of hands

13:23.280 --> 13:26.800
 so that the role of luck gets smaller.

13:26.800 --> 13:29.920
 So you could otherwise get lucky and get some good hands

13:29.920 --> 13:31.480
 and then you're gonna win the match.

13:31.480 --> 13:34.400
 Even with thousands of hands, you can get lucky

13:35.280 --> 13:36.720
 because there's so much variance

13:36.720 --> 13:40.880
 in No Limit Texas Holden because if we both go all in,

13:40.880 --> 13:43.640
 it's a huge stack of variance, so there are these

13:43.640 --> 13:47.800
 massive swings in No Limit Texas Holden.

13:47.800 --> 13:50.240
 So that's why you have to play not just thousands,

13:50.240 --> 13:55.000
 but over 100,000 hands to get statistical significance.

13:55.000 --> 13:57.880
 So let me ask another way this question.

13:57.880 --> 14:00.880
 If you didn't even look at your hands,

14:02.000 --> 14:04.560
 but they didn't know that, the opponents didn't know that,

14:04.560 --> 14:06.680
 how well would you be able to do?

14:06.680 --> 14:07.760
 Oh, that's a good question.

14:07.760 --> 14:09.600
 There's actually, I heard this story

14:09.600 --> 14:11.800
 that there's this Norwegian female poker player

14:11.800 --> 14:15.240
 called Annette Oberstad who's actually won a tournament

14:15.240 --> 14:18.640
 by doing exactly that, but that would be extremely rare.

14:18.640 --> 14:23.440
 So you cannot really play well that way.

14:23.440 --> 14:27.840
 Okay, so the hands do have some role to play, okay.

14:27.840 --> 14:32.840
 So Labradus does not use, as far as I understand,

14:33.120 --> 14:35.320
 they use learning methods, deep learning.

14:35.320 --> 14:40.320
 Is there room for learning in,

14:40.600 --> 14:44.120
 there's no reason why Labradus doesn't combine

14:44.120 --> 14:46.400
 with an AlphaGo type approach for estimating

14:46.400 --> 14:49.200
 the quality for function estimator.

14:49.200 --> 14:52.040
 What are your thoughts on this,

14:52.040 --> 14:54.760
 maybe as compared to another algorithm

14:54.760 --> 14:56.720
 which I'm not that familiar with, DeepStack,

14:56.720 --> 14:59.280
 the engine that does use deep learning,

14:59.280 --> 15:01.560
 that it's unclear how well it does,

15:01.560 --> 15:03.480
 but nevertheless uses deep learning.

15:03.480 --> 15:05.400
 So what are your thoughts about learning methods

15:05.400 --> 15:09.280
 to aid in the way that Labradus plays in the game of poker?

15:09.280 --> 15:10.640
 Yeah, so as you said,

15:10.640 --> 15:13.080
 Labradus did not use learning methods

15:13.080 --> 15:15.680
 and played very well without them.

15:15.680 --> 15:17.840
 Since then, we have actually, actually here,

15:17.840 --> 15:20.000
 we have a couple of papers on things

15:20.000 --> 15:22.360
 that do use learning techniques.

15:22.360 --> 15:23.200
 Excellent.

15:24.440 --> 15:26.360
 And deep learning in particular.

15:26.360 --> 15:29.920
 And sort of the way you're talking about

15:29.920 --> 15:33.360
 where it's learning an evaluation function,

15:33.360 --> 15:37.400
 but in imperfect information games,

15:37.400 --> 15:42.400
 unlike let's say in Go or now also in chess and shogi,

15:42.440 --> 15:47.400
 it's not sufficient to learn an evaluation for a state

15:47.400 --> 15:52.400
 because the value of an information set

15:52.920 --> 15:55.400
 depends not only on the exact state,

15:55.400 --> 15:59.200
 but it also depends on both players beliefs.

15:59.200 --> 16:01.240
 Like if I have a bad hand,

16:01.240 --> 16:04.720
 I'm much better off if the opponent thinks I have a good hand

16:04.720 --> 16:05.560
 and vice versa.

16:05.560 --> 16:06.480
 If I have a good hand,

16:06.480 --> 16:09.360
 I'm much better off if the opponent believes

16:09.360 --> 16:10.280
 I have a bad hand.

16:11.360 --> 16:15.640
 So the value of a state is not just a function of the cards.

16:15.640 --> 16:19.600
 It depends on, if you will, the path of play,

16:19.600 --> 16:22.040
 but only to the extent that it's captured

16:22.040 --> 16:23.720
 in the belief distributions.

16:23.720 --> 16:26.240
 So that's why it's not as simple

16:26.240 --> 16:29.320
 as it is in perfect information games.

16:29.320 --> 16:31.080
 And I don't wanna say it's simple there either.

16:31.080 --> 16:34.200
 It's of course very complicated computationally there too,

16:34.200 --> 16:36.520
 but at least conceptually, it's very straightforward.

16:36.520 --> 16:38.760
 There's a state, there's an evaluation function.

16:38.760 --> 16:39.800
 You can try to learn it.

16:39.800 --> 16:43.280
 Here, you have to do something more.

16:43.280 --> 16:47.160
 And what we do is in one of these papers,

16:47.160 --> 16:50.800
 we're looking at where we allow the opponent

16:50.800 --> 16:53.000
 to actually take different strategies

16:53.000 --> 16:56.440
 at the leaf of the search tree, if you will.

16:56.440 --> 16:59.840
 And that is a different way of doing it.

16:59.840 --> 17:02.560
 And it doesn't assume therefore a particular way

17:02.560 --> 17:04.040
 that the opponent plays,

17:04.040 --> 17:05.840
 but it allows the opponent to choose

17:05.840 --> 17:09.800
 from a set of different continuation strategies.

17:09.800 --> 17:13.400
 And that forces us to not be too optimistic

17:13.400 --> 17:15.520
 in a look ahead search.

17:15.520 --> 17:19.040
 And that's one way you can do sound look ahead search

17:19.040 --> 17:21.480
 in imperfect information games,

17:21.480 --> 17:23.360
 which is very difficult.

17:23.360 --> 17:26.080
 And you were asking about DeepStack.

17:26.080 --> 17:29.280
 What they did, it was very different than what we do,

17:29.280 --> 17:32.000
 either in Libratus or in this new work.

17:32.000 --> 17:35.440
 They were randomly generating various situations

17:35.440 --> 17:36.440
 in the game.

17:36.440 --> 17:38.080
 Then they were doing the look ahead

17:38.080 --> 17:39.840
 from there to the end of the game,

17:39.840 --> 17:42.960
 as if that was the start of a different game.

17:42.960 --> 17:44.920
 And then they were using deep learning

17:44.920 --> 17:47.960
 to learn those values of those states,

17:47.960 --> 17:50.280
 but the states were not just the physical states.

17:50.280 --> 17:52.560
 They include belief distributions.

17:52.560 --> 17:56.800
 When you talk about look ahead for DeepStack

17:56.800 --> 17:59.480
 or with Libratus, does it mean,

17:59.480 --> 18:02.680
 considering every possibility that the game can evolve,

18:02.680 --> 18:04.280
 are we talking about extremely,

18:04.280 --> 18:06.880
 sort of this exponentially growth of a tree?

18:06.880 --> 18:09.720
 Yes, so we're talking about exactly that.

18:11.280 --> 18:14.280
 Much like you do in alpha beta search

18:14.280 --> 18:17.480
 or Monte Carlo tree search, but with different techniques.

18:17.480 --> 18:19.280
 So there's a different search algorithm.

18:19.280 --> 18:21.920
 And then we have to deal with the leaves differently.

18:21.920 --> 18:24.000
 So if you think about what Libratus did,

18:24.000 --> 18:25.520
 we didn't have to worry about this

18:25.520 --> 18:28.560
 because we only did it at the end of the game.

18:28.560 --> 18:32.280
 So we would always terminate into a real situation

18:32.280 --> 18:34.000
 and we would know what the payout is.

18:34.000 --> 18:36.880
 It didn't do these depth limited lookaheads,

18:36.880 --> 18:40.680
 but now in this new paper, which is called depth limited,

18:40.680 --> 18:42.120
 I think it's called depth limited search

18:42.120 --> 18:43.880
 for imperfect information games,

18:43.880 --> 18:47.040
 we can actually do sound depth limited lookahead.

18:47.040 --> 18:49.240
 So we can actually start to do the look ahead

18:49.240 --> 18:51.080
 from the beginning of the game on,

18:51.080 --> 18:53.400
 because that's too complicated to do

18:53.400 --> 18:54.920
 for this whole long game.

18:54.920 --> 18:57.680
 So in Libratus, we were just doing it for the end.

18:57.680 --> 19:00.720
 So, and then the other side, this belief distribution,

19:00.720 --> 19:05.320
 so is it explicitly modeled what kind of beliefs

19:05.320 --> 19:07.400
 that the opponent might have?

19:07.400 --> 19:11.840
 Yeah, it is explicitly modeled, but it's not assumed.

19:11.840 --> 19:15.400
 The beliefs are actually output, not input.

19:15.400 --> 19:18.840
 Of course, the starting beliefs are input,

19:18.840 --> 19:20.640
 but they just fall from the rules of the game

19:20.640 --> 19:23.520
 because we know that the dealer deals uniformly

19:23.520 --> 19:27.720
 from the deck, so I know that every pair of cards

19:27.720 --> 19:30.440
 that you might have is equally likely.

19:30.440 --> 19:32.200
 I know that for a fact, that just follows

19:32.200 --> 19:33.160
 from the rules of the game.

19:33.160 --> 19:35.200
 Of course, except the two cards that I have,

19:35.200 --> 19:36.560
 I know you don't have those.

19:36.560 --> 19:37.560
 Yeah.

19:37.560 --> 19:38.720
 You have to take that into account.

19:38.720 --> 19:40.920
 That's called card removal and that's very important.

19:40.920 --> 19:43.760
 Is the dealing always coming from a single deck

19:43.760 --> 19:45.880
 in Heads Up, so you can assume.

19:45.880 --> 19:50.880
 Single deck, so you know that if I have the ace of spades,

19:50.880 --> 19:53.560
 I know you don't have an ace of spades.

19:53.560 --> 19:56.880
 Great, so in the beginning, your belief is basically

19:56.880 --> 19:59.320
 the fact that it's a fair dealing of hands,

19:59.320 --> 20:02.800
 but how do you start to adjust that belief?

20:02.800 --> 20:06.800
 Well, that's where this beauty of game theory comes.

20:06.800 --> 20:10.920
 So Nash equilibrium, which John Nash introduced in 1950,

20:10.920 --> 20:13.800
 introduces what rational play is

20:13.800 --> 20:16.040
 when you have more than one player.

20:16.040 --> 20:18.440
 And these are pairs of strategies

20:18.440 --> 20:20.360
 where strategies are contingency plans,

20:20.360 --> 20:21.600
 one for each player.

20:22.880 --> 20:25.720
 So that neither player wants to deviate

20:25.720 --> 20:26.960
 to a different strategy,

20:26.960 --> 20:29.160
 given that the other doesn't deviate.

20:29.160 --> 20:33.840
 But as a side effect, you get the beliefs from base roll.

20:33.840 --> 20:36.440
 So Nash equilibrium really isn't just deriving

20:36.440 --> 20:38.360
 in these imperfect information games,

20:38.360 --> 20:41.920
 Nash equilibrium, it doesn't just define strategies.

20:41.920 --> 20:44.960
 It also defines beliefs for both of us

20:44.960 --> 20:48.840
 and defines beliefs for each state.

20:48.840 --> 20:53.280
 So at each state, it's called information sets.

20:53.280 --> 20:55.560
 At each information set in the game,

20:55.560 --> 20:59.000
 there's a set of different states that we might be in,

20:59.000 --> 21:00.880
 but I don't know which one we're in.

21:01.760 --> 21:03.400
 Nash equilibrium tells me exactly

21:03.400 --> 21:05.000
 what is the probability distribution

21:05.000 --> 21:08.280
 over those real world states in my mind.

21:08.280 --> 21:11.440
 How does Nash equilibrium give you that distribution?

21:11.440 --> 21:12.280
 So why?

21:12.280 --> 21:13.320
 I'll do a simple example.

21:13.320 --> 21:16.760
 So you know the game Rock, Paper, Scissors?

21:16.760 --> 21:20.000
 So we can draw it as player one moves first

21:20.000 --> 21:21.600
 and then player two moves.

21:21.600 --> 21:24.520
 But of course, it's important that player two

21:24.520 --> 21:26.400
 doesn't know what player one moved,

21:26.400 --> 21:28.600
 otherwise player two would win every time.

21:28.600 --> 21:30.480
 So we can draw that as an information set

21:30.480 --> 21:33.280
 where player one makes one of three moves first,

21:33.280 --> 21:36.200
 and then there's an information set for player two.

21:36.200 --> 21:39.920
 So player two doesn't know which of those nodes

21:39.920 --> 21:41.800
 the world is in.

21:41.800 --> 21:44.920
 But once we know the strategy for player one,

21:44.920 --> 21:47.320
 Nash equilibrium will say that you play 1 3rd Rock,

21:47.320 --> 21:49.400
 1 3rd Paper, 1 3rd Scissors.

21:49.400 --> 21:52.600
 From that, I can derive my beliefs on the information set

21:52.600 --> 21:54.480
 that they're 1 3rd, 1 3rd, 1 3rd.

21:54.480 --> 21:56.280
 So Bayes gives you that.

21:56.280 --> 21:57.560
 Bayes gives you.

21:57.560 --> 21:59.760
 But is that specific to a particular player,

21:59.760 --> 22:03.960
 or is it something you quickly update

22:03.960 --> 22:05.040
 with the specific player?

22:05.040 --> 22:08.800
 No, the game theory isn't really player specific.

22:08.800 --> 22:11.720
 So that's also why we don't need any data.

22:11.720 --> 22:12.760
 We don't need any history

22:12.760 --> 22:14.800
 how these particular humans played in the past

22:14.800 --> 22:17.400
 or how any AI or human had played before.

22:17.400 --> 22:20.240
 It's all about rationality.

22:20.240 --> 22:22.720
 So the AI just thinks about

22:22.720 --> 22:24.880
 what would a rational opponent do?

22:24.880 --> 22:28.000
 And what would I do if I am rational?

22:28.000 --> 22:31.080
 And that's the idea of game theory.

22:31.080 --> 22:35.560
 So it's really a data free, opponent free approach.

22:35.560 --> 22:37.680
 So it comes from the design of the game

22:37.680 --> 22:40.040
 as opposed to the design of the player.

22:40.040 --> 22:43.080
 Exactly, there's no opponent modeling per se.

22:43.080 --> 22:45.600
 I mean, we've done some work on combining opponent modeling

22:45.600 --> 22:48.840
 with game theory so you can exploit weak players even more,

22:48.840 --> 22:50.280
 but that's another strand.

22:50.280 --> 22:52.320
 And in Librarus, we didn't turn that on.

22:52.320 --> 22:55.000
 So I decided that these players are too good.

22:55.000 --> 22:58.080
 And when you start to exploit an opponent,

22:58.080 --> 23:01.800
 you typically open yourself up to exploitation.

23:01.800 --> 23:04.000
 And these guys have so few holes to exploit

23:04.000 --> 23:06.760
 and they're world's leading experts in counter exploitation.

23:06.760 --> 23:09.200
 So I decided that we're not gonna turn that stuff on.

23:09.200 --> 23:12.160
 Actually, I saw a few of your papers exploiting opponents.

23:12.160 --> 23:14.800
 It sounded very interesting to explore.

23:15.720 --> 23:17.880
 Do you think there's room for exploitation

23:17.880 --> 23:19.920
 generally outside of Librarus?

23:19.920 --> 23:24.080
 Is there a subject or people differences

23:24.080 --> 23:27.920
 that could be exploited, maybe not just in poker,

23:27.920 --> 23:30.440
 but in general interactions and negotiations,

23:30.440 --> 23:33.480
 all these other domains that you're considering?

23:33.480 --> 23:34.680
 Yeah, definitely.

23:34.680 --> 23:35.920
 We've done some work on that.

23:35.920 --> 23:39.880
 And I really like the work at hybrid digested too.

23:39.880 --> 23:43.440
 So you figure out what would a rational opponent do.

23:43.440 --> 23:46.280
 And by the way, that's safe in these zero sum games,

23:46.280 --> 23:47.480
 two player zero sum games,

23:47.480 --> 23:49.560
 because if the opponent does something irrational,

23:49.560 --> 23:52.200
 yes, it might throw off my beliefs,

23:53.080 --> 23:55.760
 but the amount that the player can gain

23:55.760 --> 23:59.160
 by throwing off my belief is always less

23:59.160 --> 24:01.800
 than they lose by playing poorly.

24:01.800 --> 24:03.080
 So it's safe.

24:03.080 --> 24:06.720
 But still, if somebody's weak as a player,

24:06.720 --> 24:10.240
 you might wanna play differently to exploit them more.

24:10.240 --> 24:12.040
 So you can think about it this way,

24:12.040 --> 24:15.600
 a game theoretic strategy is unbeatable,

24:15.600 --> 24:19.600
 but it doesn't maximally beat the other opponent.

24:19.600 --> 24:22.800
 So the winnings per hand might be better

24:22.800 --> 24:24.240
 with a different strategy.

24:24.240 --> 24:25.720
 And the hybrid is that you start

24:25.720 --> 24:27.080
 from a game theoretic approach.

24:27.080 --> 24:30.840
 And then as you gain data about the opponent

24:30.840 --> 24:32.600
 in certain parts of the game tree,

24:32.600 --> 24:34.360
 then in those parts of the game tree,

24:34.360 --> 24:37.800
 you start to tweak your strategy more and more

24:37.800 --> 24:40.960
 towards exploitation while still staying fairly close

24:40.960 --> 24:42.160
 to the game theoretic strategy

24:42.160 --> 24:46.840
 so as to not open yourself up to exploitation too much.

24:46.840 --> 24:48.320
 How do you do that?

24:48.320 --> 24:53.320
 Do you try to vary up strategies, make it unpredictable?

24:53.640 --> 24:57.520
 It's like, what is it, tit for tat strategies

24:57.520 --> 25:00.720
 in Prisoner's Dilemma or?

25:00.720 --> 25:03.240
 Well, that's a repeated game.

25:03.240 --> 25:04.080
 Repeated games.

25:04.080 --> 25:06.520
 Simple Prisoner's Dilemma, repeated games.

25:06.520 --> 25:08.760
 But even there, there's no proof that says

25:08.760 --> 25:10.080
 that that's the best thing.

25:10.080 --> 25:13.280
 But experimentally, it actually does well.

25:13.280 --> 25:15.320
 So what kind of games are there, first of all?

25:15.320 --> 25:17.040
 I don't know if this is something

25:17.040 --> 25:18.600
 that you could just summarize.

25:18.600 --> 25:20.360
 There's perfect information games

25:20.360 --> 25:22.400
 where all the information's on the table.

25:22.400 --> 25:25.480
 There is imperfect information games.

25:25.480 --> 25:28.560
 There's repeated games that you play over and over.

25:28.560 --> 25:31.320
 There's zero sum games.

25:31.320 --> 25:34.440
 There's non zero sum games.

25:34.440 --> 25:37.520
 And then there's a really important distinction

25:37.520 --> 25:40.720
 you're making, two player versus more players.

25:40.720 --> 25:44.760
 So what are, what other games are there?

25:44.760 --> 25:46.160
 And what's the difference, for example,

25:46.160 --> 25:50.040
 with this two player game versus more players?

25:50.040 --> 25:51.680
 What are the key differences in your view?

25:51.680 --> 25:54.600
 So let me start from the basics.

25:54.600 --> 25:59.600
 So a repeated game is a game where the same exact game

25:59.600 --> 26:01.800
 is played over and over.

26:01.800 --> 26:05.800
 In these extensive form games, where it's,

26:05.800 --> 26:08.480
 think about three form, maybe with these information sets

26:08.480 --> 26:11.400
 to represent incomplete information,

26:11.400 --> 26:14.840
 you can have kind of repetitive interactions.

26:14.840 --> 26:17.760
 Even repeated games are a special case of that, by the way.

26:17.760 --> 26:21.520
 But the game doesn't have to be exactly the same.

26:21.520 --> 26:23.040
 It's like in sourcing auctions.

26:23.040 --> 26:26.320
 Yes, we're gonna see the same supply base year to year,

26:26.320 --> 26:28.800
 but what I'm buying is a little different every time.

26:28.800 --> 26:31.000
 And the supply base is a little different every time

26:31.000 --> 26:31.840
 and so on.

26:31.840 --> 26:33.400
 So it's not really repeated.

26:33.400 --> 26:35.680
 So to find a purely repeated game

26:35.680 --> 26:37.840
 is actually very rare in the world.

26:37.840 --> 26:42.840
 So they're really a very course model of what's going on.

26:42.840 --> 26:46.360
 Then if you move up from just repeated,

26:46.360 --> 26:49.040
 simple repeated matrix games,

26:49.040 --> 26:50.800
 not all the way to extensive form games,

26:50.800 --> 26:53.600
 but in between, they're stochastic games,

26:53.600 --> 26:57.000
 where, you know, there's these,

26:57.000 --> 27:00.520
 you think about it like these little matrix games.

27:00.520 --> 27:04.200
 And when you take an action and your opponent takes an action,

27:04.200 --> 27:07.680
 they determine not which next state I'm going to,

27:07.680 --> 27:09.120
 next game I'm going to,

27:09.120 --> 27:11.440
 but the distribution over next games

27:11.440 --> 27:13.360
 where I might be going to.

27:13.360 --> 27:15.360
 So that's the stochastic game.

27:15.360 --> 27:19.000
 But it's like matrix games, repeated stochastic games,

27:19.000 --> 27:20.400
 extensive form games.

27:20.400 --> 27:23.040
 That is from less to more general.

27:23.040 --> 27:26.280
 And poker is an example of the last one.

27:26.280 --> 27:28.400
 So it's really in the most general setting.

27:29.560 --> 27:30.640
 Extensive form games.

27:30.640 --> 27:34.520
 And that's kind of what the AI community has been working on

27:34.520 --> 27:36.280
 and being benchmarked on

27:36.280 --> 27:38.040
 with this Heads Up No Limit Texas Holdem.

27:38.040 --> 27:39.760
 Can you describe extensive form games?

27:39.760 --> 27:41.560
 What's the model here?

27:41.560 --> 27:44.320
 Yeah, so if you're familiar with the tree form,

27:44.320 --> 27:45.760
 so it's really the tree form.

27:45.760 --> 27:47.560
 Like in chess, there's a search tree.

27:47.560 --> 27:48.720
 Versus a matrix.

27:48.720 --> 27:50.080
 Versus a matrix, yeah.

27:50.080 --> 27:53.000
 And the matrix is called the matrix form

27:53.000 --> 27:55.320
 or bi matrix form or normal form game.

27:55.320 --> 27:57.080
 And here you have the tree form.

27:57.080 --> 28:00.000
 So you can actually do certain types of reasoning there

28:00.000 --> 28:04.680
 that you lose the information when you go to normal form.

28:04.680 --> 28:07.000
 There's a certain form of equivalence.

28:07.000 --> 28:08.880
 Like if you go from tree form and you say it,

28:08.880 --> 28:12.720
 every possible contingency plan is a strategy.

28:12.720 --> 28:15.080
 Then I can actually go back to the normal form,

28:15.080 --> 28:18.600
 but I lose some information from the lack of sequentiality.

28:18.600 --> 28:21.280
 Then the multiplayer versus two player distinction

28:21.280 --> 28:22.880
 is an important one.

28:22.880 --> 28:27.320
 So two player games in zero sum

28:27.320 --> 28:32.320
 are conceptually easier and computationally easier.

28:32.840 --> 28:36.000
 They're still huge like this one,

28:36.000 --> 28:39.680
 but they're conceptually easier and computationally easier

28:39.680 --> 28:42.920
 in that conceptually, you don't have to worry about

28:42.920 --> 28:45.360
 which equilibrium is the other guy going to play

28:45.360 --> 28:46.640
 when there are multiple,

28:46.640 --> 28:49.920
 because any equilibrium strategy is a best response

28:49.920 --> 28:52.000
 to any other equilibrium strategy.

28:52.000 --> 28:54.360
 So I can play a different equilibrium from you

28:54.360 --> 28:57.320
 and we'll still get the right values of the game.

28:57.320 --> 28:59.240
 That falls apart even with two players

28:59.240 --> 29:01.360
 when you have general sum games.

29:01.360 --> 29:03.120
 Even without cooperation just in general.

29:03.120 --> 29:04.800
 Even without cooperation.

29:04.800 --> 29:07.640
 So there's a big gap from two player zero sum

29:07.640 --> 29:11.160
 to two player general sum or even to three player zero sum.

29:11.160 --> 29:14.280
 That's a big gap, at least in theory.

29:14.280 --> 29:18.920
 Can you maybe non mathematically provide the intuition

29:18.920 --> 29:22.120
 why it all falls apart with three or more players?

29:22.120 --> 29:24.400
 It seems like you should still be able to have

29:24.400 --> 29:29.400
 a Nash equilibrium that's instructive, that holds.

29:31.280 --> 29:36.000
 Okay, so it is true that all finite games

29:36.000 --> 29:38.200
 have a Nash equilibrium.

29:38.200 --> 29:41.080
 So this is what John Nash actually proved.

29:41.080 --> 29:42.920
 So they do have a Nash equilibrium.

29:42.920 --> 29:43.840
 That's not the problem.

29:43.840 --> 29:46.600
 The problem is that there can be many.

29:46.600 --> 29:50.400
 And then there's a question of which equilibrium to select.

29:50.400 --> 29:52.200
 So, and if you select your strategy

29:52.200 --> 29:54.640
 from a different equilibrium and I select mine,

29:57.920 --> 29:59.920
 then what does that mean?

29:59.920 --> 30:02.080
 And in these non zero sum games,

30:02.080 --> 30:05.720
 we may lose some joint benefit

30:05.720 --> 30:07.040
 by being just simply stupid.

30:07.040 --> 30:08.400
 We could actually both be better off

30:08.400 --> 30:09.920
 if we did something else.

30:09.920 --> 30:11.760
 And in three player, you get other problems

30:11.760 --> 30:13.200
 also like collusion.

30:13.200 --> 30:16.560
 Like maybe you and I can gang up on a third player

30:16.560 --> 30:19.800
 and we can do radically better by colluding.

30:19.800 --> 30:22.200
 So there are lots of issues that come up there.

30:22.200 --> 30:25.640
 So Noah Brown, the student you work with on this

30:25.640 --> 30:29.360
 has mentioned, I looked through the AMA on Reddit.

30:29.360 --> 30:31.280
 He mentioned that the ability of poker players

30:31.280 --> 30:33.800
 to collaborate will make the game.

30:33.800 --> 30:35.200
 He was asked the question of,

30:35.200 --> 30:37.920
 how would you make the game of poker,

30:37.920 --> 30:39.280
 or both of you were asked the question,

30:39.280 --> 30:41.560
 how would you make the game of poker

30:41.560 --> 30:46.560
 beyond being solvable by current AI methods?

30:47.000 --> 30:50.560
 And he said that there's not many ways

30:50.560 --> 30:53.120
 of making poker more difficult,

30:53.120 --> 30:57.760
 but a collaboration or cooperation between players

30:57.760 --> 30:59.760
 would make it extremely difficult.

30:59.760 --> 31:03.320
 So can you provide the intuition behind why that is,

31:03.320 --> 31:05.280
 if you agree with that idea?

31:05.280 --> 31:10.200
 Yeah, so I've done a lot of work on coalitional games

31:10.200 --> 31:11.680
 and we actually have a paper here

31:11.680 --> 31:13.680
 with my other student Gabriele Farina

31:13.680 --> 31:16.640
 and some other collaborators at NIPS on that.

31:16.640 --> 31:18.520
 Actually just came back from the poster session

31:18.520 --> 31:19.760
 where we presented this.

31:19.760 --> 31:23.800
 But so when you have a collusion, it's a different problem.

31:23.800 --> 31:26.120
 And it typically gets even harder then.

31:27.520 --> 31:29.600
 Even the game representations,

31:29.600 --> 31:32.320
 some of the game representations don't really allow

31:33.600 --> 31:34.480
 good computation.

31:34.480 --> 31:37.600
 So we actually introduced a new game representation

31:37.600 --> 31:38.720
 for that.

31:38.720 --> 31:42.040
 Is that kind of cooperation part of the model?

31:42.040 --> 31:44.560
 Are you, do you have, do you have information

31:44.560 --> 31:47.040
 about the fact that other players are cooperating

31:47.040 --> 31:50.000
 or is it just this chaos that where nothing is known?

31:50.000 --> 31:52.360
 So there's some things unknown.

31:52.360 --> 31:55.840
 Can you give an example of a collusion type game

31:55.840 --> 31:56.680
 or is it usually?

31:56.680 --> 31:58.360
 So like bridge.

31:58.360 --> 31:59.640
 So think about bridge.

31:59.640 --> 32:02.320
 It's like when you and I are on a team,

32:02.320 --> 32:04.480
 our payoffs are the same.

32:04.480 --> 32:06.400
 The problem is that we can't talk.

32:06.400 --> 32:09.000
 So when I get my cards, I can't whisper to you

32:09.000 --> 32:10.320
 what my cards are.

32:10.320 --> 32:12.480
 That would not be allowed.

32:12.480 --> 32:16.080
 So we have to somehow coordinate our strategies

32:16.080 --> 32:19.920
 ahead of time and only ahead of time.

32:19.920 --> 32:22.760
 And then there's certain signals we can talk about,

32:22.760 --> 32:25.240
 but they have to be such that the other team

32:25.240 --> 32:26.840
 also understands them.

32:26.840 --> 32:30.440
 So that's an example where the coordination

32:30.440 --> 32:33.000
 is already built into the rules of the game.

32:33.000 --> 32:35.640
 But in many other situations like auctions

32:35.640 --> 32:40.640
 or negotiations or diplomatic relationships, poker,

32:40.880 --> 32:44.160
 it's not really built in, but it still can be very helpful

32:44.160 --> 32:45.280
 for the colluders.

32:45.280 --> 32:48.240
 I've read you write somewhere,

32:48.240 --> 32:52.800
 the negotiations you come to the table with prior,

32:52.800 --> 32:56.080
 like a strategy that you're willing to do

32:56.080 --> 32:58.320
 and not willing to do those kinds of things.

32:58.320 --> 33:01.960
 So how do you start to now moving away from poker,

33:01.960 --> 33:04.520
 moving beyond poker into other applications

33:04.520 --> 33:07.000
 like negotiations, how do you start applying this

33:07.000 --> 33:11.640
 to other domains, even real world domains

33:11.640 --> 33:12.520
 that you've worked on?

33:12.520 --> 33:14.440
 Yeah, I actually have two startup companies

33:14.440 --> 33:15.480
 doing exactly that.

33:15.480 --> 33:17.800
 One is called Strategic Machine,

33:17.800 --> 33:20.000
 and that's for kind of business applications,

33:20.000 --> 33:22.880
 gaming, sports, all sorts of things like that.

33:22.880 --> 33:27.200
 Any applications of this to business and to sports

33:27.200 --> 33:32.120
 and to gaming, to various types of things

33:32.120 --> 33:34.240
 in finance, electricity markets and so on.

33:34.240 --> 33:36.600
 And the other is called Strategy Robot,

33:36.600 --> 33:40.640
 where we are taking these to military security,

33:40.640 --> 33:43.520
 cyber security and intelligence applications.

33:43.520 --> 33:46.240
 I think you worked a little bit in,

33:48.000 --> 33:51.000
 how do you put it, advertisement,

33:51.000 --> 33:55.360
 sort of suggesting ads kind of thing, auction.

33:55.360 --> 33:57.800
 That's another company, optimized markets.

33:57.800 --> 34:00.880
 But that's much more about a combinatorial market

34:00.880 --> 34:02.840
 and optimization based technology.

34:02.840 --> 34:06.840
 That's not using these game theoretic reasoning technologies.

34:06.840 --> 34:11.600
 I see, okay, so what sort of high level

34:11.600 --> 34:15.280
 do you think about our ability to use

34:15.280 --> 34:18.040
 game theoretic concepts to model human behavior?

34:18.040 --> 34:21.640
 Do you think human behavior is amenable

34:21.640 --> 34:24.720
 to this kind of modeling outside of the poker games,

34:24.720 --> 34:27.520
 and where have you seen it done successfully in your work?

34:27.520 --> 34:32.520
 I'm not sure the goal really is modeling humans.

34:33.640 --> 34:36.480
 Like for example, if I'm playing a zero sum game,

34:36.480 --> 34:39.840
 I don't really care that the opponent

34:39.840 --> 34:42.960
 is actually following my model of rational behavior,

34:42.960 --> 34:46.400
 because if they're not, that's even better for me.

34:46.400 --> 34:50.200
 Right, so see with the opponents in games,

34:51.120 --> 34:56.120
 the prerequisite is that you formalize

34:56.120 --> 34:57.800
 the interaction in some way

34:57.800 --> 35:01.000
 that can be amenable to analysis.

35:01.000 --> 35:04.160
 And you've done this amazing work with mechanism design,

35:04.160 --> 35:08.160
 designing games that have certain outcomes.

35:10.040 --> 35:12.320
 But, so I'll tell you an example

35:12.320 --> 35:15.460
 from my world of autonomous vehicles, right?

35:15.460 --> 35:17.040
 We're studying pedestrians,

35:17.040 --> 35:20.200
 and pedestrians and cars negotiate

35:20.200 --> 35:22.160
 in this nonverbal communication.

35:22.160 --> 35:25.040
 There's this weird game dance of tension

35:25.040 --> 35:27.280
 where pedestrians are basically saying,

35:27.280 --> 35:28.800
 I trust that you won't kill me,

35:28.800 --> 35:31.840
 and so as a jaywalker, I will step onto the road

35:31.840 --> 35:34.720
 even though I'm breaking the law, and there's this tension.

35:34.720 --> 35:36.640
 And the question is, we really don't know

35:36.640 --> 35:40.720
 how to model that well in trying to model intent.

35:40.720 --> 35:43.080
 And so people sometimes bring up ideas

35:43.080 --> 35:44.880
 of game theory and so on.

35:44.880 --> 35:49.120
 Do you think that aspect of human behavior

35:49.120 --> 35:53.080
 can use these kinds of imperfect information approaches,

35:53.080 --> 35:57.860
 modeling, how do you start to attack a problem like that

35:57.860 --> 36:00.940
 when you don't even know how to design the game

36:00.940 --> 36:04.280
 to describe the situation in order to solve it?

36:04.280 --> 36:06.800
 Okay, so I haven't really thought about jaywalking,

36:06.800 --> 36:10.120
 but one thing that I think could be a good application

36:10.120 --> 36:13.000
 in autonomous vehicles is the following.

36:13.000 --> 36:16.320
 So let's say that you have fleets of autonomous cars

36:16.320 --> 36:18.340
 operating by different companies.

36:18.340 --> 36:22.120
 So maybe here's the Waymo fleet and here's the Uber fleet.

36:22.120 --> 36:24.320
 If you think about the rules of the road,

36:24.320 --> 36:26.560
 they define certain legal rules,

36:26.560 --> 36:30.080
 but that still leaves a huge strategy space open.

36:30.080 --> 36:32.840
 Like as a simple example, when cars merge,

36:32.840 --> 36:36.000
 how humans merge, they slow down and look at each other

36:36.000 --> 36:39.240
 and try to merge.

36:39.240 --> 36:40.920
 Wouldn't it be better if these situations

36:40.920 --> 36:43.480
 would already be prenegotiated

36:43.480 --> 36:45.200
 so we can actually merge at full speed

36:45.200 --> 36:47.440
 and we know that this is the situation,

36:47.440 --> 36:50.540
 this is how we do it, and it's all gonna be faster.

36:50.540 --> 36:54.120
 But there are way too many situations to negotiate manually.

36:54.120 --> 36:56.400
 So you could use automated negotiation,

36:56.400 --> 36:57.780
 this is the idea at least,

36:57.780 --> 36:59.840
 you could use automated negotiation

36:59.840 --> 37:02.060
 to negotiate all of these situations

37:02.060 --> 37:04.320
 or many of them in advance.

37:04.320 --> 37:05.460
 And of course it might be that,

37:05.460 --> 37:09.180
 hey, maybe you're not gonna always let me go first.

37:09.180 --> 37:11.280
 Maybe you said, okay, well, in these situations,

37:11.280 --> 37:13.560
 I'll let you go first, but in exchange,

37:13.560 --> 37:14.520
 you're gonna give me too much,

37:14.520 --> 37:17.260
 you're gonna let me go first in this situation.

37:17.260 --> 37:20.680
 So it's this huge combinatorial negotiation.

37:20.680 --> 37:24.080
 And do you think there's room in that example of merging

37:24.080 --> 37:25.600
 to model this whole situation

37:25.600 --> 37:27.160
 as an imperfect information game

37:27.160 --> 37:30.120
 or do you really want to consider it to be a perfect?

37:30.120 --> 37:32.240
 No, that's a good question, yeah.

37:32.240 --> 37:33.080
 That's a good question.

37:33.080 --> 37:37.080
 Do you pay the price of assuming

37:37.080 --> 37:38.640
 that you don't know everything?

37:39.800 --> 37:40.760
 Yeah, I don't know.

37:40.760 --> 37:42.120
 It's certainly much easier.

37:42.120 --> 37:45.060
 Games with perfect information are much easier.

37:45.060 --> 37:49.280
 So if you can't get away with it, you should.

37:49.280 --> 37:52.640
 But if the real situation is of imperfect information,

37:52.640 --> 37:55.160
 then you're gonna have to deal with imperfect information.

37:55.160 --> 37:58.080
 Great, so what lessons have you learned

37:58.080 --> 38:00.680
 the Annual Computer Poker Competition?

38:00.680 --> 38:03.440
 An incredible accomplishment of AI.

38:03.440 --> 38:07.000
 You look at the history of Deep Blue, AlphaGo,

38:07.000 --> 38:10.400
 these kind of moments when AI stepped up

38:10.400 --> 38:13.960
 in an engineering effort and a scientific effort combined

38:13.960 --> 38:16.400
 to beat the best of human players.

38:16.400 --> 38:19.480
 So what do you take away from this whole experience?

38:19.480 --> 38:22.440
 What have you learned about designing AI systems

38:22.440 --> 38:23.960
 that play these kinds of games?

38:23.960 --> 38:28.280
 And what does that mean for AI in general,

38:28.280 --> 38:30.760
 for the future of AI development?

38:30.760 --> 38:32.800
 Yeah, so that's a good question.

38:32.800 --> 38:34.560
 So there's so much to say about it.

38:35.440 --> 38:39.120
 I do like this type of performance oriented research.

38:39.120 --> 38:42.000
 Although in my group, we go all the way from like idea

38:42.000 --> 38:44.880
 to theory, to experiments, to big system building,

38:44.880 --> 38:47.960
 to commercialization, so we span that spectrum.

38:47.960 --> 38:51.080
 But I think that in a lot of situations in AI,

38:51.080 --> 38:53.440
 you really have to build the big systems

38:53.440 --> 38:55.640
 and evaluate them at scale

38:55.640 --> 38:57.520
 before you know what works and doesn't.

38:57.520 --> 39:00.080
 And we've seen that in the computational

39:00.080 --> 39:02.880
 game theory community, that there are a lot of techniques

39:02.880 --> 39:04.280
 that look good in the small,

39:05.200 --> 39:07.120
 but then they cease to look good in the large.

39:07.120 --> 39:10.160
 And we've also seen that there are a lot of techniques

39:10.160 --> 39:13.280
 that look superior in theory.

39:13.280 --> 39:16.200
 And I really mean in terms of convergence rates,

39:16.200 --> 39:18.440
 like first order methods, better convergence rates,

39:18.440 --> 39:20.880
 like the CFR based algorithms,

39:20.880 --> 39:24.880
 yet the CFR based algorithms are the fastest in practice.

39:24.880 --> 39:28.240
 So it really tells me that you have to test this in reality.

39:28.240 --> 39:30.880
 The theory isn't tight enough, if you will,

39:30.880 --> 39:34.360
 to tell you which algorithms are better than the others.

39:34.360 --> 39:38.600
 And you have to look at these things in the large,

39:38.600 --> 39:41.480
 because any sort of projections you do from the small

39:41.480 --> 39:43.800
 can at least in this domain be very misleading.

39:43.800 --> 39:46.240
 So that's kind of from a kind of a science

39:46.240 --> 39:49.120
 and engineering perspective, from a personal perspective,

39:49.120 --> 39:51.280
 it's been just a wild experience

39:51.280 --> 39:54.160
 in that with the first poker competition,

39:54.160 --> 39:57.200
 the first brains versus AI,

39:57.200 --> 39:59.840
 man machine poker competition that we organized.

39:59.840 --> 40:01.760
 There had been, by the way, for other poker games,

40:01.760 --> 40:03.240
 there had been previous competitions,

40:03.240 --> 40:06.360
 but this was for Heads Up No Limit, this was the first.

40:06.360 --> 40:09.560
 And I probably became the most hated person

40:09.560 --> 40:10.880
 in the world of poker.

40:10.880 --> 40:12.880
 And I didn't mean to, I just saw.

40:12.880 --> 40:13.720
 Why is that?

40:13.720 --> 40:15.840
 For cracking the game, for something.

40:15.840 --> 40:20.000
 Yeah, a lot of people felt that it was a real threat

40:20.000 --> 40:22.760
 to the whole game, the whole existence of the game.

40:22.760 --> 40:26.080
 If AI becomes better than humans,

40:26.080 --> 40:28.520
 people would be scared to play poker

40:28.520 --> 40:30.680
 because there are these superhuman AIs running around

40:30.680 --> 40:32.760
 taking their money and all of that.

40:32.760 --> 40:36.200
 So I just, it's just really aggressive.

40:36.200 --> 40:37.880
 The comments were super aggressive.

40:37.880 --> 40:40.920
 I got everything just short of death threats.

40:40.920 --> 40:44.000
 Do you think the same was true for chess?

40:44.000 --> 40:45.760
 Because right now they just completed

40:45.760 --> 40:47.720
 the world championships in chess,

40:47.720 --> 40:49.560
 and humans just started ignoring the fact

40:49.560 --> 40:52.920
 that there's AI systems now that outperform humans

40:52.920 --> 40:55.520
 and they still enjoy the game, it's still a beautiful game.

40:55.520 --> 40:56.360
 That's what I think.

40:56.360 --> 40:58.800
 And I think the same thing happens in poker.

40:58.800 --> 41:01.040
 And so I didn't think of myself

41:01.040 --> 41:02.360
 as somebody who was gonna kill the game,

41:02.360 --> 41:03.800
 and I don't think I did.

41:03.800 --> 41:05.600
 I've really learned to love this game.

41:05.600 --> 41:06.960
 I wasn't a poker player before,

41:06.960 --> 41:10.520
 but learned so many nuances about it from these AIs,

41:10.520 --> 41:12.480
 and they've really changed how the game is played,

41:12.480 --> 41:13.320
 by the way.

41:13.320 --> 41:16.240
 So they have these very Martian ways of playing poker,

41:16.240 --> 41:18.400
 and the top humans are now incorporating

41:18.400 --> 41:21.400
 those types of strategies into their own play.

41:21.400 --> 41:26.400
 So if anything, to me, our work has made poker

41:26.560 --> 41:29.800
 a richer, more interesting game for humans to play,

41:29.800 --> 41:32.160
 not something that is gonna steer humans

41:32.160 --> 41:34.200
 away from it entirely.

41:34.200 --> 41:35.960
 Just a quick comment on something you said,

41:35.960 --> 41:39.400
 which is, if I may say so,

41:39.400 --> 41:42.400
 in academia is a little bit rare sometimes.

41:42.400 --> 41:45.520
 It's pretty brave to put your ideas to the test

41:45.520 --> 41:47.200
 in the way you described,

41:47.200 --> 41:49.360
 saying that sometimes good ideas don't work

41:49.360 --> 41:52.760
 when you actually try to apply them at scale.

41:52.760 --> 41:54.200
 So where does that come from?

41:54.200 --> 41:58.880
 I mean, if you could do advice for people,

41:58.880 --> 42:00.760
 what drives you in that sense?

42:00.760 --> 42:02.360
 Were you always this way?

42:02.360 --> 42:04.080
 I mean, it takes a brave person.

42:04.080 --> 42:06.760
 I guess is what I'm saying, to test their ideas

42:06.760 --> 42:08.640
 and to see if this thing actually works

42:08.640 --> 42:11.680
 against human top human players and so on.

42:11.680 --> 42:12.960
 Yeah, I don't know about brave,

42:12.960 --> 42:15.000
 but it takes a lot of work.

42:15.000 --> 42:17.320
 It takes a lot of work and a lot of time

42:18.400 --> 42:20.360
 to organize, to make something big

42:20.360 --> 42:22.920
 and to organize an event and stuff like that.

42:22.920 --> 42:24.760
 And what drives you in that effort?

42:24.760 --> 42:26.880
 Because you could still, I would argue,

42:26.880 --> 42:30.280
 get a best paper award at NIPS as you did in 17

42:30.280 --> 42:31.440
 without doing this.

42:31.440 --> 42:32.960
 That's right, yes.

42:32.960 --> 42:37.640
 And so in general, I believe it's very important

42:37.640 --> 42:41.480
 to do things in the real world and at scale.

42:41.480 --> 42:46.160
 And that's really where the pudding, if you will,

42:46.160 --> 42:48.400
 proof is in the pudding, that's where it is.

42:48.400 --> 42:50.080
 In this particular case,

42:50.080 --> 42:55.080
 it was kind of a competition between different groups

42:55.160 --> 42:59.080
 and for many years as to who can be the first one

42:59.080 --> 43:02.040
 to beat the top humans at Heads Up No Limit, Texas Holdem.

43:02.040 --> 43:07.040
 So it became kind of like a competition who can get there.

43:09.560 --> 43:11.800
 Yeah, so a little friendly competition

43:11.800 --> 43:14.040
 could do wonders for progress.

43:14.040 --> 43:15.040
 Yes, absolutely.

43:16.400 --> 43:19.040
 So the topic of mechanism design,

43:19.040 --> 43:22.280
 which is really interesting, also kind of new to me,

43:22.280 --> 43:25.680
 except as an observer of, I don't know, politics and any,

43:25.680 --> 43:27.600
 I'm an observer of mechanisms,

43:27.600 --> 43:31.440
 but you write in your paper an automated mechanism design

43:31.440 --> 43:34.000
 that I quickly read.

43:34.000 --> 43:37.880
 So mechanism design is designing the rules of the game

43:37.880 --> 43:40.200
 so you get a certain desirable outcome.

43:40.200 --> 43:44.920
 And you have this work on doing so in an automatic fashion

43:44.920 --> 43:46.720
 as opposed to fine tuning it.

43:46.720 --> 43:50.680
 So what have you learned from those efforts?

43:50.680 --> 43:52.280
 If you look, say, I don't know,

43:52.280 --> 43:56.200
 at complexes like our political system,

43:56.200 --> 43:58.560
 can we design our political system

43:58.560 --> 44:01.800
 to have, in an automated fashion,

44:01.800 --> 44:03.360
 to have outcomes that we want?

44:03.360 --> 44:08.360
 Can we design something like traffic lights to be smart

44:09.000 --> 44:11.800
 where it gets outcomes that we want?

44:11.800 --> 44:14.840
 So what are the lessons that you draw from that work?

44:14.840 --> 44:17.240
 Yeah, so I still very much believe

44:17.240 --> 44:19.400
 in the automated mechanism design direction.

44:19.400 --> 44:20.840
 Yes.

44:20.840 --> 44:23.000
 But it's not a panacea.

44:23.000 --> 44:26.520
 There are impossibility results in mechanism design

44:26.520 --> 44:30.240
 saying that there is no mechanism that accomplishes

44:30.240 --> 44:33.920
 objective X in class C.

44:33.920 --> 44:36.120
 So it's not going to,

44:36.120 --> 44:39.000
 there's no way using any mechanism design tools,

44:39.000 --> 44:41.000
 manual or automated,

44:41.000 --> 44:42.800
 to do certain things in mechanism design.

44:42.800 --> 44:43.800
 Can you describe that again?

44:43.800 --> 44:47.480
 So meaning it's impossible to achieve that?

44:47.480 --> 44:48.320
 Yeah, yeah.

44:48.320 --> 44:50.440
 And it's unlikely.

44:50.440 --> 44:51.280
 Impossible.

44:51.280 --> 44:52.120
 Impossible.

44:52.120 --> 44:55.240
 So these are not statements about human ingenuity

44:55.240 --> 44:57.120
 who might come up with something smart.

44:57.120 --> 44:59.880
 These are proofs that if you want to accomplish

44:59.880 --> 45:02.480
 properties X in class C,

45:02.480 --> 45:04.880
 that is not doable with any mechanism.

45:04.880 --> 45:07.080
 The good thing about automated mechanism design

45:07.080 --> 45:10.840
 is that we're not really designing for a class.

45:10.840 --> 45:14.160
 We're designing for specific settings at a time.

45:14.160 --> 45:16.720
 So even if there's an impossibility result

45:16.720 --> 45:18.240
 for the whole class,

45:18.240 --> 45:21.360
 it just doesn't mean that all of the cases

45:21.360 --> 45:22.560
 in the class are impossible.

45:22.560 --> 45:25.080
 It just means that some of the cases are impossible.

45:25.080 --> 45:28.240
 So we can actually carve these islands of possibility

45:28.240 --> 45:30.920
 within these known impossible classes.

45:30.920 --> 45:31.960
 And we've actually done that.

45:31.960 --> 45:35.160
 So one of the famous results in mechanism design

45:35.160 --> 45:37.360
 is the Meyerson Satethweight theorem

45:37.360 --> 45:41.000
 by Roger Meyerson and Mark Satethweight from 1983.

45:41.000 --> 45:43.480
 It's an impossibility of efficient trade

45:43.480 --> 45:45.200
 under imperfect information.

45:45.200 --> 45:48.560
 We show that you can, in many settings,

45:48.560 --> 45:51.480
 avoid that and get efficient trade anyway.

45:51.480 --> 45:54.160
 Depending on how they design the game, okay.

45:54.160 --> 45:55.880
 Depending how you design the game.

45:55.880 --> 46:00.240
 And of course, it doesn't in any way

46:00.240 --> 46:01.800
 contradict the impossibility result.

46:01.800 --> 46:03.920
 The impossibility result is still there,

46:03.920 --> 46:08.000
 but it just finds spots within this impossible class

46:08.920 --> 46:12.440
 where in those spots, you don't have the impossibility.

46:12.440 --> 46:14.760
 Sorry if I'm going a bit philosophical,

46:14.760 --> 46:17.480
 but what lessons do you draw towards,

46:17.480 --> 46:20.160
 like I mentioned, politics or human interaction

46:20.160 --> 46:24.880
 and designing mechanisms for outside of just

46:24.880 --> 46:26.960
 these kinds of trading or auctioning

46:26.960 --> 46:31.960
 or purely formal games or human interaction,

46:33.480 --> 46:34.920
 like a political system?

46:34.920 --> 46:39.160
 How, do you think it's applicable to, yeah, politics

46:39.160 --> 46:44.160
 or to business, to negotiations, these kinds of things,

46:46.280 --> 46:49.040
 designing rules that have certain outcomes?

46:49.040 --> 46:51.360
 Yeah, yeah, I do think so.

46:51.360 --> 46:54.200
 Have you seen that successfully done?

46:54.200 --> 46:56.440
 They haven't really, oh, you mean mechanism design

46:56.440 --> 46:57.280
 or automated mechanism design?

46:57.280 --> 46:59.000
 Automated mechanism design.

46:59.000 --> 47:01.520
 So mechanism design itself

47:01.520 --> 47:06.440
 has had fairly limited success so far.

47:06.440 --> 47:07.600
 There are certain cases,

47:07.600 --> 47:10.200
 but most of the real world situations

47:10.200 --> 47:14.680
 are actually not sound from a mechanism design perspective,

47:14.680 --> 47:16.920
 even in those cases where they've been designed

47:16.920 --> 47:20.000
 by very knowledgeable mechanism design people,

47:20.000 --> 47:22.760
 the people are typically just taking some insights

47:22.760 --> 47:25.040
 from the theory and applying those insights

47:25.040 --> 47:26.280
 into the real world,

47:26.280 --> 47:29.280
 rather than applying the mechanisms directly.

47:29.280 --> 47:33.520
 So one famous example of is the FCC spectrum auctions.

47:33.520 --> 47:36.880
 So I've also had a small role in that

47:36.880 --> 47:40.600
 and very good economists have been working,

47:40.600 --> 47:42.560
 excellent economists have been working on that

47:42.560 --> 47:44.040
 with no game theory,

47:44.040 --> 47:47.440
 yet the rules that are designed in practice there,

47:47.440 --> 47:49.840
 they're such that bidding truthfully

47:49.840 --> 47:51.800
 is not the best strategy.

47:51.800 --> 47:52.960
 Usually mechanism design,

47:52.960 --> 47:56.160
 we try to make things easy for the participants.

47:56.160 --> 47:58.560
 So telling the truth is the best strategy,

47:58.560 --> 48:01.480
 but even in those very high stakes auctions

48:01.480 --> 48:03.080
 where you have tens of billions of dollars

48:03.080 --> 48:05.200
 worth of spectrum being auctioned,

48:06.360 --> 48:08.280
 truth telling is not the best strategy.

48:08.280 --> 48:10.040
 And by the way,

48:10.040 --> 48:12.920
 nobody knows even a single optimal bidding strategy

48:12.920 --> 48:14.120
 for those auctions.

48:14.120 --> 48:15.960
 What's the challenge of coming up with an optimal,

48:15.960 --> 48:18.160
 because there's a lot of players and there's imperfect.

48:18.160 --> 48:20.040
 It's not so much that a lot of players,

48:20.040 --> 48:22.320
 but many items for sale,

48:22.320 --> 48:26.000
 and these mechanisms are such that even with just two items

48:26.000 --> 48:28.400
 or one item, bidding truthfully

48:28.400 --> 48:30.400
 wouldn't be the best strategy.

48:31.400 --> 48:34.560
 If you look at the history of AI,

48:34.560 --> 48:37.160
 it's marked by seminal events.

48:37.160 --> 48:40.160
 AlphaGo beating a world champion human Go player,

48:40.160 --> 48:43.680
 I would put Liberatus winning the Heads Up No Limit Holdem

48:43.680 --> 48:45.000
 as one of such event.

48:45.000 --> 48:46.040
 Thank you.

48:46.040 --> 48:51.040
 And what do you think is the next such event,

48:52.560 --> 48:56.640
 whether it's in your life or in the broadly AI community

48:56.640 --> 48:59.040
 that you think might be out there

48:59.040 --> 49:01.640
 that would surprise the world?

49:01.640 --> 49:02.800
 So that's a great question,

49:02.800 --> 49:04.520
 and I don't really know the answer.

49:04.520 --> 49:06.160
 In terms of game solving,

49:07.360 --> 49:08.920
 Heads Up No Limit Texas Holdem

49:08.920 --> 49:13.920
 really was the one remaining widely agreed upon benchmark.

49:14.400 --> 49:15.880
 So that was the big milestone.

49:15.880 --> 49:17.800
 Now, are there other things?

49:17.800 --> 49:18.920
 Yeah, certainly there are,

49:18.920 --> 49:21.080
 but there's not one that the community

49:21.080 --> 49:22.920
 has kind of focused on.

49:22.920 --> 49:25.240
 So what could be other things?

49:25.240 --> 49:27.640
 There are groups working on StarCraft.

49:27.640 --> 49:29.840
 There are groups working on Dota 2.

49:29.840 --> 49:31.560
 These are video games.

49:31.560 --> 49:36.240
 Or you could have like Diplomacy or Hanabi,

49:36.240 --> 49:37.080
 things like that.

49:37.080 --> 49:38.640
 These are like recreational games,

49:38.640 --> 49:42.040
 but none of them are really acknowledged

49:42.040 --> 49:45.840
 as kind of the main next challenge problem,

49:45.840 --> 49:50.000
 like chess or Go or Heads Up No Limit Texas Holdem was.

49:50.000 --> 49:52.360
 So I don't really know in the game solving space

49:52.360 --> 49:55.400
 what is or what will be the next benchmark.

49:55.400 --> 49:57.840
 I kind of hope that there will be a next benchmark

49:57.840 --> 49:59.560
 because really the different groups

49:59.560 --> 50:01.120
 working on the same problem

50:01.120 --> 50:05.120
 really drove these application independent techniques

50:05.120 --> 50:07.480
 forward very quickly over 10 years.

50:07.480 --> 50:09.120
 Do you think there's an open problem

50:09.120 --> 50:11.480
 that excites you that you start moving away

50:11.480 --> 50:15.000
 from games into real world games,

50:15.000 --> 50:17.200
 like say the stock market trading?

50:17.200 --> 50:19.320
 Yeah, so that's kind of how I am.

50:19.320 --> 50:23.120
 So I am probably not going to work

50:23.120 --> 50:27.640
 as hard on these recreational benchmarks.

50:27.640 --> 50:30.200
 I'm doing two startups on game solving technology,

50:30.200 --> 50:32.320
 Strategic Machine and Strategy Robot,

50:32.320 --> 50:34.160
 and we're really interested

50:34.160 --> 50:36.560
 in pushing this stuff into practice.

50:36.560 --> 50:40.080
 What do you think would be really

50:43.160 --> 50:45.920
 a powerful result that would be surprising

50:45.920 --> 50:49.960
 that would be, if you can say,

50:49.960 --> 50:53.280
 I mean, five years, 10 years from now,

50:53.280 --> 50:56.480
 something that statistically you would say

50:56.480 --> 50:57.920
 is not very likely,

50:57.920 --> 51:01.480
 but if there's a breakthrough, would achieve?

51:01.480 --> 51:03.800
 Yeah, so I think that overall,

51:03.800 --> 51:08.800
 we're in a very different situation in game theory

51:09.000 --> 51:11.760
 than we are in, let's say, machine learning.

51:11.760 --> 51:14.360
 So in machine learning, it's a fairly mature technology

51:14.360 --> 51:16.480
 and it's very broadly applied

51:16.480 --> 51:19.680
 and proven success in the real world.

51:19.680 --> 51:22.840
 In game solving, there are almost no applications yet.

51:24.320 --> 51:26.680
 We have just become superhuman,

51:26.680 --> 51:29.600
 which machine learning you could argue happened in the 90s,

51:29.600 --> 51:30.640
 if not earlier,

51:30.640 --> 51:32.960
 and at least on supervised learning,

51:32.960 --> 51:35.400
 certain complex supervised learning applications.

51:36.960 --> 51:39.000
 Now, I think the next challenge problem,

51:39.000 --> 51:40.560
 I know you're not asking about it this way,

51:40.560 --> 51:42.640
 you're asking about the technology breakthrough,

51:42.640 --> 51:44.240
 but I think that big, big breakthrough

51:44.240 --> 51:46.120
 is to be able to show that, hey,

51:46.120 --> 51:48.280
 maybe most of, let's say, military planning

51:48.280 --> 51:50.080
 or most of business strategy

51:50.080 --> 51:52.200
 will actually be done strategically

51:52.200 --> 51:54.120
 using computational game theory.

51:54.120 --> 51:55.800
 That's what I would like to see

51:55.800 --> 51:57.640
 as the next five or 10 year goal.

51:57.640 --> 51:59.520
 Maybe you can explain to me again,

51:59.520 --> 52:01.920
 forgive me if this is an obvious question,

52:01.920 --> 52:04.000
 but machine learning methods,

52:04.000 --> 52:07.840
 neural networks suffer from not being transparent,

52:07.840 --> 52:09.280
 not being explainable.

52:09.280 --> 52:12.400
 Game theoretic methods, Nash equilibria,

52:12.400 --> 52:15.280
 do they generally, when you see the different solutions,

52:15.280 --> 52:19.640
 are they, when you talk about military operations,

52:19.640 --> 52:21.800
 are they, once you see the strategies,

52:21.800 --> 52:23.880
 do they make sense, are they explainable,

52:23.880 --> 52:25.840
 or do they suffer from the same problems

52:25.840 --> 52:27.120
 as neural networks do?

52:27.120 --> 52:28.720
 So that's a good question.

52:28.720 --> 52:31.240
 I would say a little bit yes and no.

52:31.240 --> 52:34.560
 And what I mean by that is that

52:34.560 --> 52:36.160
 these game theoretic strategies,

52:36.160 --> 52:38.520
 let's say, Nash equilibrium,

52:38.520 --> 52:40.320
 it has provable properties.

52:40.320 --> 52:42.360
 So it's unlike, let's say, deep learning

52:42.360 --> 52:44.440
 where you kind of cross your fingers,

52:44.440 --> 52:45.680
 hopefully it'll work.

52:45.680 --> 52:47.880
 And then after the fact, when you have the weights,

52:47.880 --> 52:48.920
 you're still crossing your fingers,

52:48.920 --> 52:50.160
 and I hope it will work.

52:51.160 --> 52:55.400
 Here, you know that the solution quality is there.

52:55.400 --> 52:58.040
 There's provable solution quality guarantees.

52:58.040 --> 53:00.920
 Now, that doesn't necessarily mean

53:00.920 --> 53:03.480
 that the strategies are human understandable.

53:03.480 --> 53:04.720
 That's a whole other problem.

53:04.720 --> 53:08.680
 So I think that deep learning and computational game theory

53:08.680 --> 53:10.720
 are in the same boat in that sense,

53:10.720 --> 53:12.680
 that both are difficult to understand.

53:13.760 --> 53:15.680
 But at least the game theoretic techniques,

53:15.680 --> 53:19.840
 they have these guarantees of solution quality.

53:19.840 --> 53:22.880
 So do you see business operations, strategic operations,

53:22.880 --> 53:26.040
 or even military in the future being

53:26.040 --> 53:28.320
 at least the strong candidates

53:28.320 --> 53:32.760
 being proposed by automated systems?

53:32.760 --> 53:34.000
 Do you see that?

53:34.000 --> 53:35.040
 Yeah, I do, I do.

53:35.040 --> 53:39.640
 But that's more of a belief than a substantiated fact.

53:39.640 --> 53:42.320
 Depending on where you land in optimism or pessimism,

53:42.320 --> 53:45.720
 that's a really, to me, that's an exciting future,

53:45.720 --> 53:48.760
 especially if there's provable things

53:48.760 --> 53:50.560
 in terms of optimality.

53:50.560 --> 53:54.040
 So looking into the future,

53:54.040 --> 53:58.760
 there's a few folks worried about the,

53:58.760 --> 54:01.200
 especially you look at the game of poker,

54:01.200 --> 54:03.360
 which is probably one of the last benchmarks

54:03.360 --> 54:05.480
 in terms of games being solved.

54:05.480 --> 54:07.520
 They worry about the future

54:07.520 --> 54:10.520
 and the existential threats of artificial intelligence.

54:10.520 --> 54:13.840
 So the negative impact in whatever form on society.

54:13.840 --> 54:17.440
 Is that something that concerns you as much,

54:17.440 --> 54:21.600
 or are you more optimistic about the positive impacts of AI?

54:21.600 --> 54:24.720
 Oh, I am much more optimistic about the positive impacts.

54:24.720 --> 54:27.560
 So just in my own work, what we've done so far,

54:27.560 --> 54:29.920
 we run the nationwide kidney exchange.

54:29.920 --> 54:32.960
 Hundreds of people are walking around alive today,

54:32.960 --> 54:34.080
 who would it be?

54:34.080 --> 54:36.120
 And it's increased employment.

54:36.120 --> 54:39.920
 You have a lot of people now running kidney exchanges

54:39.920 --> 54:42.200
 and at the transplant centers,

54:42.200 --> 54:45.560
 interacting with the kidney exchange.

54:45.560 --> 54:49.440
 You have extra surgeons, nurses, anesthesiologists,

54:49.440 --> 54:51.400
 hospitals, all of that.

54:51.400 --> 54:53.560
 So employment is increasing from that

54:53.560 --> 54:55.320
 and the world is becoming a better place.

54:55.320 --> 54:59.040
 Another example is combinatorial sourcing auctions.

54:59.040 --> 55:04.040
 We did 800 large scale combinatorial sourcing auctions

55:04.040 --> 55:08.240
 from 2001 to 2010 in a previous startup of mine

55:08.240 --> 55:09.400
 called CombineNet.

55:09.400 --> 55:13.080
 And we increased the supply chain efficiency

55:13.080 --> 55:18.080
 on that $60 billion of spend by 12.6%.

55:18.080 --> 55:21.440
 So that's over $6 billion of efficiency improvement

55:21.440 --> 55:22.240
 in the world.

55:22.240 --> 55:23.760
 And this is not like shifting value

55:23.760 --> 55:25.240
 from somebody to somebody else,

55:25.240 --> 55:28.200
 just efficiency improvement, like in trucking,

55:28.200 --> 55:31.120
 less empty driving, so there's less waste,

55:31.120 --> 55:33.440
 less carbon footprint and so on.

55:33.440 --> 55:36.720
 So a huge positive impact in the near term,

55:36.720 --> 55:40.680
 but sort of to stay in it for a little longer,

55:40.680 --> 55:43.080
 because I think game theory has a role to play here.

55:43.080 --> 55:45.320
 Oh, let me actually come back on that as one thing.

55:45.320 --> 55:49.400
 I think AI is also going to make the world much safer.

55:49.400 --> 55:53.760
 So that's another aspect that often gets overlooked.

55:53.760 --> 55:54.920
 Well, let me ask this question.

55:54.920 --> 55:56.960
 Maybe you can speak to the safer.

55:56.960 --> 55:59.960
 So I talked to Max Tegmark and Stuart Russell,

55:59.960 --> 56:02.680
 who are very concerned about existential threats of AI.

56:02.680 --> 56:06.240
 And often the concern is about value misalignment.

56:06.240 --> 56:10.240
 So AI systems basically working,

56:11.880 --> 56:14.680
 operating towards goals that are not the same

56:14.680 --> 56:17.920
 as human civilization, human beings.

56:17.920 --> 56:21.160
 So it seems like game theory has a role to play there

56:24.200 --> 56:27.880
 to make sure the values are aligned with human beings.

56:27.880 --> 56:29.960
 I don't know if that's how you think about it.

56:29.960 --> 56:34.960
 If not, how do you think AI might help with this problem?

56:35.280 --> 56:39.240
 How do you think AI might make the world safer?

56:39.240 --> 56:43.000
 Yeah, I think this value misalignment

56:43.000 --> 56:46.480
 is a fairly theoretical worry.

56:46.480 --> 56:49.960
 And I haven't really seen it in,

56:49.960 --> 56:51.840
 because I do a lot of real applications.

56:51.840 --> 56:53.920
 I don't see it anywhere.

56:53.920 --> 56:55.240
 The closest I've seen it

56:55.240 --> 56:57.920
 was the following type of mental exercise really,

56:57.920 --> 57:00.720
 where I had this argument in the late eighties

57:00.720 --> 57:01.560
 when we were building

57:01.560 --> 57:03.560
 these transportation optimization systems.

57:03.560 --> 57:05.360
 And somebody had heard that it's a good idea

57:05.360 --> 57:08.160
 to have high utilization of assets.

57:08.160 --> 57:11.400
 So they told me, hey, why don't you put that as objective?

57:11.400 --> 57:14.720
 And we didn't even put it as an objective

57:14.720 --> 57:16.480
 because I just showed him that,

57:16.480 --> 57:18.480
 if you had that as your objective,

57:18.480 --> 57:20.320
 the solution would be to load your trucks full

57:20.320 --> 57:21.840
 and drive in circles.

57:21.840 --> 57:23.000
 Nothing would ever get delivered.

57:23.000 --> 57:25.120
 You'd have a hundred percent utilization.

57:25.120 --> 57:27.240
 So yeah, I know this phenomenon.

57:27.240 --> 57:29.680
 I've known this for over 30 years,

57:29.680 --> 57:33.360
 but I've never seen it actually be a problem in reality.

57:33.360 --> 57:35.240
 And yes, if you have the wrong objective,

57:35.240 --> 57:37.800
 the AI will optimize that to the hilt

57:37.800 --> 57:39.800
 and it's gonna hurt more than some human

57:39.800 --> 57:43.800
 who's kind of trying to solve it in a half baked way

57:43.800 --> 57:45.480
 with some human insight too.

57:45.480 --> 57:49.160
 But I just haven't seen that materialize in practice.

57:49.160 --> 57:52.720
 There's this gap that you've actually put your finger on

57:52.720 --> 57:57.080
 very clearly just now between theory and reality.

57:57.080 --> 57:59.680
 That's very difficult to put into words, I think.

57:59.680 --> 58:02.240
 It's what you can theoretically imagine,

58:03.240 --> 58:08.000
 the worst possible case or even, yeah, I mean bad cases.

58:08.000 --> 58:10.520
 And what usually happens in reality.

58:10.520 --> 58:11.960
 So for example, to me,

58:11.960 --> 58:15.720
 maybe it's something you can comment on having grown up

58:15.720 --> 58:17.680
 and I grew up in the Soviet Union.

58:19.120 --> 58:22.120
 There's currently 10,000 nuclear weapons in the world.

58:22.120 --> 58:24.200
 And for many decades,

58:24.200 --> 58:28.360
 it's theoretically surprising to me

58:28.360 --> 58:30.880
 that the nuclear war is not broken out.

58:30.880 --> 58:33.760
 Do you think about this aspect

58:33.760 --> 58:36.080
 from a game theoretic perspective in general,

58:36.080 --> 58:38.440
 why is that true?

58:38.440 --> 58:40.720
 Why in theory you could see

58:40.720 --> 58:42.600
 how things would go terribly wrong

58:42.600 --> 58:44.280
 and somehow yet they have not?

58:44.280 --> 58:45.600
 Yeah, how do you think about it?

58:45.600 --> 58:47.240
 So I do think about that a lot.

58:47.240 --> 58:50.320
 I think the biggest two threats that we're facing as mankind,

58:50.320 --> 58:53.320
 one is climate change and the other is nuclear war.

58:53.320 --> 58:57.200
 So those are my main two worries that I worry about.

58:57.200 --> 58:59.920
 And I've tried to do something about climate,

58:59.920 --> 59:01.320
 thought about trying to do something

59:01.320 --> 59:02.880
 for climate change twice.

59:02.880 --> 59:05.040
 Actually, for two of my startups,

59:05.040 --> 59:06.760
 I've actually commissioned studies

59:06.760 --> 59:09.480
 of what we could do on those things.

59:09.480 --> 59:11.040
 And we didn't really find a sweet spot,

59:11.040 --> 59:12.680
 but I'm still keeping an eye out on that.

59:12.680 --> 59:15.160
 If there's something where we could actually

59:15.160 --> 59:17.800
 provide a market solution or optimizations solution

59:17.800 --> 59:20.960
 or some other technology solution to problems.

59:20.960 --> 59:23.360
 Right now, like for example,

59:23.360 --> 59:26.760
 pollution critic markets was what we were looking at then.

59:26.760 --> 59:30.040
 And it was much more the lack of political will

59:30.040 --> 59:32.840
 by those markets were not so successful

59:32.840 --> 59:34.640
 rather than bad market design.

59:34.640 --> 59:37.080
 So I could go in and make a better market design,

59:37.080 --> 59:38.600
 but that wouldn't really move the needle

59:38.600 --> 59:41.160
 on the world very much if there's no political will.

59:41.160 --> 59:43.600
 And in the US, the market,

59:43.600 --> 59:47.520
 at least the Chicago market was just shut down and so on.

59:47.520 --> 59:48.760
 So then it doesn't really help

59:48.760 --> 59:51.040
 how great your market design was.

59:51.040 --> 59:53.560
 And then the nuclear side, it's more,

59:53.560 --> 59:57.560
 so global warming is a more encroaching problem.

1:00:00.840 --> 1:00:03.280
 Nuclear weapons have been here.

1:00:03.280 --> 1:00:05.720
 It's an obvious problem that's just been sitting there.

1:00:05.720 --> 1:00:07.480
 So how do you think about,

1:00:07.480 --> 1:00:09.240
 what is the mechanism design there

1:00:09.240 --> 1:00:12.280
 that just made everything seem stable?

1:00:12.280 --> 1:00:14.800
 And are you still extremely worried?

1:00:14.800 --> 1:00:16.640
 I am still extremely worried.

1:00:16.640 --> 1:00:20.040
 So you probably know the simple game theory of mad.

1:00:20.040 --> 1:00:23.760
 So this was a mutually assured destruction

1:00:23.760 --> 1:00:27.360
 and it doesn't require any computation with small matrices.

1:00:27.360 --> 1:00:28.600
 You can actually convince yourself

1:00:28.600 --> 1:00:31.480
 that the game is such that nobody wants to initiate.

1:00:31.480 --> 1:00:34.600
 Yeah, that's a very coarse grained analysis.

1:00:34.600 --> 1:00:36.880
 And it really works in a situational way.

1:00:36.880 --> 1:00:40.400
 You have two superpowers or small number of superpowers.

1:00:40.400 --> 1:00:41.960
 Now things are very different.

1:00:41.960 --> 1:00:43.080
 You have a smaller nuke.

1:00:43.080 --> 1:00:47.240
 So the threshold of initiating is smaller

1:00:47.240 --> 1:00:51.520
 and you have smaller countries and non nation actors

1:00:51.520 --> 1:00:53.760
 who may get a nuke and so on.

1:00:53.760 --> 1:00:58.320
 So I think it's riskier now than it was maybe ever before.

1:00:58.320 --> 1:01:03.320
 And what idea, application of AI,

1:01:03.640 --> 1:01:04.640
 you've talked about a little bit,

1:01:04.640 --> 1:01:07.560
 but what is the most exciting to you right now?

1:01:07.560 --> 1:01:10.160
 I mean, you're here at NIPS, NeurIPS.

1:01:10.160 --> 1:01:14.920
 Now you have a few excellent pieces of work,

1:01:14.920 --> 1:01:16.680
 but what are you thinking into the future

1:01:16.680 --> 1:01:17.840
 with several companies you're doing?

1:01:17.840 --> 1:01:21.120
 What's the most exciting thing or one of the exciting things?

1:01:21.120 --> 1:01:23.160
 The number one thing for me right now

1:01:23.160 --> 1:01:26.360
 is coming up with these scalable techniques

1:01:26.360 --> 1:01:30.440
 for game solving and applying them into the real world.

1:01:30.440 --> 1:01:33.160
 I'm still very interested in market design as well.

1:01:33.160 --> 1:01:35.400
 And we're doing that in the optimized markets,

1:01:35.400 --> 1:01:37.560
 but I'm most interested if number one right now

1:01:37.560 --> 1:01:40.000
 is strategic machine strategy robot,

1:01:40.000 --> 1:01:41.440
 getting that technology out there

1:01:41.440 --> 1:01:45.560
 and seeing as you were in the trenches doing applications,

1:01:45.560 --> 1:01:47.120
 what needs to be actually filled,

1:01:47.120 --> 1:01:49.800
 what technology gaps still need to be filled.

1:01:49.800 --> 1:01:52.040
 So it's so hard to just put your feet on the table

1:01:52.040 --> 1:01:53.800
 and imagine what needs to be done.

1:01:53.800 --> 1:01:56.280
 But when you're actually doing real applications,

1:01:56.280 --> 1:01:59.120
 the applications tell you what needs to be done.

1:01:59.120 --> 1:02:00.840
 And I really enjoy that interaction.

1:02:00.840 --> 1:02:04.480
 Is it a challenging process to apply

1:02:04.480 --> 1:02:07.760
 some of the state of the art techniques you're working on

1:02:07.760 --> 1:02:12.760
 and having the various players in industry

1:02:14.080 --> 1:02:17.720
 or the military or people who could really benefit from it

1:02:17.720 --> 1:02:19.040
 actually use it?

1:02:19.040 --> 1:02:21.400
 What's that process like of,

1:02:21.400 --> 1:02:23.680
 autonomous vehicles work with automotive companies

1:02:23.680 --> 1:02:28.200
 and they're in many ways are a little bit old fashioned.

1:02:28.200 --> 1:02:29.240
 It's difficult.

1:02:29.240 --> 1:02:31.840
 They really want to use this technology.

1:02:31.840 --> 1:02:34.640
 There's clearly will have a significant benefit,

1:02:34.640 --> 1:02:37.480
 but the systems aren't quite in place

1:02:37.480 --> 1:02:41.080
 to easily have them integrated in terms of data,

1:02:41.080 --> 1:02:43.760
 in terms of compute, in terms of all these kinds of things.

1:02:43.760 --> 1:02:48.680
 So is that one of the bigger challenges that you're facing

1:02:48.680 --> 1:02:50.000
 and how do you tackle that challenge?

1:02:50.000 --> 1:02:52.360
 Yeah, I think that's always a challenge.

1:02:52.360 --> 1:02:54.520
 That's kind of slowness and inertia really

1:02:55.560 --> 1:02:57.920
 of let's do things the way we've always done it.

1:02:57.920 --> 1:03:00.120
 You just have to find the internal champions

1:03:00.120 --> 1:03:02.120
 at the customer who understand that,

1:03:02.120 --> 1:03:04.680
 hey, things can't be the same way in the future.

1:03:04.680 --> 1:03:06.960
 Otherwise bad things are going to happen.

1:03:06.960 --> 1:03:08.600
 And it's in autonomous vehicles.

1:03:08.600 --> 1:03:09.680
 It's actually very interesting

1:03:09.680 --> 1:03:11.120
 that the car makers are doing that

1:03:11.120 --> 1:03:12.440
 and they're very traditional,

1:03:12.440 --> 1:03:14.360
 but at the same time you have tech companies

1:03:14.360 --> 1:03:17.120
 who have nothing to do with cars or transportation

1:03:17.120 --> 1:03:21.880
 like Google and Baidu really pushing on autonomous cars.

1:03:21.880 --> 1:03:23.240
 I find that fascinating.

1:03:23.240 --> 1:03:25.160
 Clearly you're super excited

1:03:25.160 --> 1:03:29.320
 about actually these ideas having an impact in the world.

1:03:29.320 --> 1:03:32.680
 In terms of the technology, in terms of ideas and research,

1:03:32.680 --> 1:03:36.600
 are there directions that you're also excited about?

1:03:36.600 --> 1:03:40.840
 Whether that's on some of the approaches you talked about

1:03:40.840 --> 1:03:42.760
 for the imperfect information games,

1:03:42.760 --> 1:03:44.000
 whether it's applying deep learning

1:03:44.000 --> 1:03:45.120
 to some of these problems,

1:03:45.120 --> 1:03:46.520
 is there something that you're excited

1:03:46.520 --> 1:03:48.840
 in the research side of things?

1:03:48.840 --> 1:03:51.120
 Yeah, yeah, lots of different things

1:03:51.120 --> 1:03:53.240
 in the game solving.

1:03:53.240 --> 1:03:56.400
 So solving even bigger games,

1:03:56.400 --> 1:03:59.760
 games where you have more hidden action

1:03:59.760 --> 1:04:02.040
 of the player actions as well.

1:04:02.040 --> 1:04:05.880
 Poker is a game where really the chance actions are hidden

1:04:05.880 --> 1:04:07.080
 or some of them are hidden,

1:04:07.080 --> 1:04:08.720
 but the player actions are public.

1:04:11.440 --> 1:04:14.000
 Multiplayer games of various sorts,

1:04:14.000 --> 1:04:18.080
 collusion, opponent exploitation,

1:04:18.080 --> 1:04:21.280
 all and even longer games.

1:04:21.280 --> 1:04:23.160
 So games that basically go forever,

1:04:23.160 --> 1:04:24.680
 but they're not repeated.

1:04:24.680 --> 1:04:27.880
 So see extensive fun games that go forever.

1:04:27.880 --> 1:04:30.080
 What would that even look like?

1:04:30.080 --> 1:04:31.040
 How do you represent that?

1:04:31.040 --> 1:04:32.040
 How do you solve that?

1:04:32.040 --> 1:04:33.440
 What's an example of a game like that?

1:04:33.440 --> 1:04:35.600
 Or is this some of the stochastic games

1:04:35.600 --> 1:04:36.440
 that you mentioned?

1:04:36.440 --> 1:04:37.320
 Let's say business strategy.

1:04:37.320 --> 1:04:40.840
 So it's not just modeling like a particular interaction,

1:04:40.840 --> 1:04:44.440
 but thinking about the business from here to eternity.

1:04:44.440 --> 1:04:49.040
 Or let's say military strategy.

1:04:49.040 --> 1:04:51.000
 So it's not like war is gonna go away.

1:04:51.000 --> 1:04:54.280
 How do you think about military strategy

1:04:54.280 --> 1:04:55.520
 that's gonna go forever?

1:04:56.680 --> 1:04:58.080
 How do you even model that?

1:04:58.080 --> 1:05:01.000
 How do you know whether a move was good

1:05:01.000 --> 1:05:05.200
 that somebody made and so on?

1:05:05.200 --> 1:05:06.960
 So that's kind of one direction.

1:05:06.960 --> 1:05:09.800
 I'm also very interested in learning

1:05:09.800 --> 1:05:13.440
 much more scalable techniques for integer programming.

1:05:13.440 --> 1:05:16.560
 So we had an ICML paper this summer on that.

1:05:16.560 --> 1:05:20.280
 The first automated algorithm configuration paper

1:05:20.280 --> 1:05:23.560
 that has theoretical generalization guarantees.

1:05:23.560 --> 1:05:26.200
 So if I see this many training examples

1:05:26.200 --> 1:05:28.560
 and I told my algorithm in this way,

1:05:28.560 --> 1:05:30.560
 it's going to have good performance

1:05:30.560 --> 1:05:33.200
 on the real distribution, which I've not seen.

1:05:33.200 --> 1:05:34.840
 So, which is kind of interesting

1:05:34.840 --> 1:05:37.680
 that algorithm configuration has been going on now

1:05:37.680 --> 1:05:41.200
 for at least 17 years seriously.

1:05:41.200 --> 1:05:45.000
 And there has not been any generalization theory before.

1:05:45.960 --> 1:05:47.200
 Well, this is really exciting

1:05:47.200 --> 1:05:49.840
 and it's a huge honor to talk to you.

1:05:49.840 --> 1:05:51.160
 Thank you so much, Tomas.

1:05:51.160 --> 1:05:52.880
 Thank you for bringing Labradus to the world

1:05:52.880 --> 1:05:54.160
 and all the great work you're doing.

1:05:54.160 --> 1:05:55.000
 Well, thank you very much.

1:05:55.000 --> 1:05:55.840
 It's been fun.

1:05:55.840 --> 1:06:16.840
 No more questions.