WEBVTT 00:00.000 --> 00:03.320 The following is a conversation with Ariol Vanialis. 00:03.320 --> 00:05.920 He's a senior research scientist at Google DeepMind, 00:05.920 --> 00:09.120 and before that, he was at Google Brain and Berkeley. 00:09.120 --> 00:13.320 His research has been cited over 39,000 times. 00:13.320 --> 00:16.520 He's truly one of the most brilliant and impactful minds 00:16.520 --> 00:18.160 in the field of deep learning. 00:18.160 --> 00:20.960 He's behind some of the biggest papers and ideas in AI, 00:20.960 --> 00:23.080 including sequence to sequence learning, 00:23.080 --> 00:25.480 audio generation, image captioning, 00:25.480 --> 00:27.000 neural machine translation, 00:27.000 --> 00:29.640 and, of course, reinforcement learning. 00:29.640 --> 00:32.800 He's a lead researcher of the AlphaStar project, 00:32.800 --> 00:35.760 creating an agent that defeated a top professional 00:35.760 --> 00:38.040 at the game of StarCraft. 00:38.040 --> 00:39.800 This conversation is part 00:39.800 --> 00:41.800 of the artificial intelligence podcast. 00:41.800 --> 00:44.920 If you enjoy it, subscribe on YouTube, iTunes, 00:44.920 --> 00:47.000 or simply connect with me on Twitter, 00:47.000 --> 00:51.200 at Lex Freedman, spelled F R I D. 00:51.200 --> 00:55.440 And now, here's my conversation with Ariol Vanialis. 00:56.520 --> 00:58.480 You spearheaded the DeepMind team 00:58.480 --> 01:00.640 behind AlphaStar that recently beat 01:00.640 --> 01:02.840 a top professional player at StarCraft. 01:04.000 --> 01:07.680 So you have an incredible wealth of work 01:07.680 --> 01:09.440 and deep learning and a bunch of fields, 01:09.440 --> 01:11.840 but let's talk about StarCraft first. 01:11.840 --> 01:13.720 Let's go back to the very beginning, 01:13.720 --> 01:16.680 even before AlphaStar, before DeepMind, 01:16.680 --> 01:18.800 before Deep Learning, first, 01:18.800 --> 01:21.240 what came first for you? 01:21.240 --> 01:24.960 A love for programming or a love for video games? 01:24.960 --> 01:28.560 I think for me, it definitely came first, 01:28.560 --> 01:31.960 the drive to play video games. 01:31.960 --> 01:35.280 I really liked computers. 01:35.280 --> 01:37.800 I didn't really code much, 01:37.800 --> 01:40.680 but what I would do is I would just mess with the computer, 01:40.680 --> 01:42.080 break it and fix it. 01:42.080 --> 01:43.800 That was the level of skills, I guess, 01:43.800 --> 01:46.400 that I gained in my very early days, 01:46.400 --> 01:48.520 I mean, when I was 10 or 11. 01:48.520 --> 01:51.000 And then I really got into video games, 01:51.000 --> 01:53.680 especially StarCraft, actually, the first version. 01:53.680 --> 01:55.800 I spent most of my time just playing, 01:55.800 --> 01:57.080 kind of, pseudo professionally, 01:57.080 --> 02:01.040 as professionally as you could play back in 98 in Europe, 02:01.040 --> 02:03.080 which was not a very main scene, 02:03.080 --> 02:05.840 like what's called nowadays esports. 02:05.840 --> 02:07.400 Right, of course, in the 90s. 02:07.400 --> 02:09.920 So how'd you get into StarCraft? 02:09.920 --> 02:11.680 What was your favorite race? 02:11.680 --> 02:15.080 How did you develop your skill? 02:15.080 --> 02:16.880 What was your strategy? 02:16.880 --> 02:18.000 All that kind of thing. 02:18.000 --> 02:21.520 So as a player, I tended to try to play not many games, 02:21.520 --> 02:25.400 not to disclose the strategies that I developed. 02:25.400 --> 02:27.560 And I like to play random, actually, 02:27.560 --> 02:30.040 not in competitions, but just to... 02:30.040 --> 02:33.400 I think in StarCraft, there's three main races, 02:33.400 --> 02:36.560 and I found it very useful to play with all of them. 02:36.560 --> 02:38.360 So I would choose random many times, 02:38.360 --> 02:40.200 even sometimes in tournaments, 02:40.200 --> 02:42.360 to gain skill on the three races, 02:42.360 --> 02:45.440 because it's not how you play against someone, 02:45.440 --> 02:48.760 but also if you understand the race because you play it, 02:48.760 --> 02:51.120 you also understand what's annoying, 02:51.120 --> 02:52.560 then when you're on the other side, 02:52.560 --> 02:54.240 what to do to annoy that person, 02:54.240 --> 02:57.360 to try to gain advantages here and there and so on. 02:57.360 --> 02:59.160 So I actually played random, 02:59.160 --> 03:02.080 although I must say in terms of favorite race, 03:02.080 --> 03:03.720 I really like Zerk. 03:03.720 --> 03:05.560 I was probably best at Zerk, 03:05.560 --> 03:08.400 and that's probably what I tend to use 03:08.400 --> 03:11.480 towards the end of my career before starting university. 03:11.480 --> 03:13.360 So let's step back a little bit. 03:13.360 --> 03:16.080 Could you try to describe StarCraft to people 03:16.080 --> 03:18.960 that may never have played video games, 03:18.960 --> 03:22.320 especially the massively online variety like StarCraft? 03:22.320 --> 03:25.920 So StarCraft is a real time strategy game, 03:25.920 --> 03:27.800 and the way to think about StarCraft, 03:27.800 --> 03:30.960 perhaps if you understand a bit chess, 03:30.960 --> 03:32.920 is that there's a board, 03:32.920 --> 03:34.240 which is called map, 03:34.240 --> 03:39.120 or the map where people play against each other. 03:39.120 --> 03:41.000 There's obviously many ways you can play, 03:41.000 --> 03:44.640 but the most interesting one is the one versus one setup, 03:44.640 --> 03:47.400 where you just play against someone else, 03:47.400 --> 03:48.880 or even the build in AI, right? 03:48.880 --> 03:52.600 Blizzard put a system that can play the game reasonably well, 03:52.600 --> 03:54.480 if you don't know how to play. 03:54.480 --> 03:56.040 And then in this board, 03:56.040 --> 03:58.680 you have, again, pieces like in chess, 03:58.680 --> 04:01.400 but these pieces are not there initially, 04:01.400 --> 04:02.360 like they are in chess. 04:02.360 --> 04:05.800 You actually need to decide to gather resources, 04:05.800 --> 04:07.920 to decide which pieces to build. 04:07.920 --> 04:10.760 So in a way, you're starting almost with no pieces. 04:10.760 --> 04:13.400 You start gathering resources in StarCraft. 04:13.400 --> 04:16.200 There's minerals and gas that you can gather, 04:16.200 --> 04:19.440 and then you must decide how much do you wanna focus, 04:19.440 --> 04:21.480 for instance, on gathering more resources, 04:21.480 --> 04:24.360 or starting to build units or pieces. 04:24.360 --> 04:27.200 And then once you have enough pieces, 04:27.200 --> 04:32.120 or maybe a good attack composition, 04:32.120 --> 04:35.480 then you go and attack the other side of the map. 04:35.480 --> 04:37.800 And now the other main difference with chess 04:37.800 --> 04:39.920 is that you don't see the other side of the map. 04:39.920 --> 04:43.360 So you're not seeing the moves of the enemy. 04:43.360 --> 04:45.440 It's what we call partially observable. 04:45.440 --> 04:50.160 So as a result, you must not only decide trading off economy 04:50.160 --> 04:52.320 versus building your own units, 04:52.320 --> 04:54.960 but you also must decide whether you wanna scout 04:54.960 --> 04:56.840 to gather information, 04:56.840 --> 04:59.520 but also by scouting you might be giving away some information 04:59.520 --> 05:01.960 that you might be hiding from the enemy. 05:01.960 --> 05:04.960 So there's a lot of complex decision making 05:04.960 --> 05:06.000 all in real time. 05:06.000 --> 05:10.120 There's also unlike chess, this is not a turn based game. 05:10.120 --> 05:13.680 You play basically all the time continuously 05:13.680 --> 05:16.920 and thus some skill in terms of speed and accuracy 05:16.920 --> 05:18.960 of clicking is also very important. 05:18.960 --> 05:20.520 And people that train for this, 05:20.520 --> 05:23.600 really play this game at an amazing skill level. 05:23.600 --> 05:25.840 I've seen many times these, 05:25.840 --> 05:27.400 and if you can witness this life, 05:27.400 --> 05:29.520 it's really, really impressive. 05:29.520 --> 05:31.440 So in a way it's kind of a chess, 05:31.440 --> 05:33.440 where you don't see the other side of the board, 05:33.440 --> 05:35.240 you're building your own pieces, 05:35.240 --> 05:37.240 and you also need to gather resources 05:37.240 --> 05:40.720 to basically get some money to build other buildings, 05:40.720 --> 05:42.920 pieces, technology, and so on. 05:42.920 --> 05:45.160 From the perspective of the human player, 05:45.160 --> 05:47.240 the difference between that and chess, 05:47.240 --> 05:50.800 or maybe that and a game like turn based strategy, 05:50.800 --> 05:52.840 like Heroes of the Might of Magic, 05:52.840 --> 05:55.200 is that there's an anxiety, 05:55.200 --> 05:58.800 because you have to make these decisions really quickly. 05:58.800 --> 06:03.800 And if you are not actually aware of what decisions work, 06:04.400 --> 06:06.520 it's a very stressful balance that you have to, 06:06.520 --> 06:08.920 everything you describe is actually quite stressful, 06:08.920 --> 06:11.760 difficult to balance for amateur human player. 06:11.760 --> 06:14.200 I don't know if it gets easier at the professional level, 06:14.200 --> 06:16.520 like if they're fully aware of what they have to do, 06:16.520 --> 06:18.360 but at the amateur level, 06:18.360 --> 06:20.520 there's this anxiety, oh crap, I'm being attacked, 06:20.520 --> 06:22.840 oh crap, I have to build up resources, 06:22.840 --> 06:24.360 oh, I have to probably expand, 06:24.360 --> 06:28.600 and all these, the real time strategy aspect 06:28.600 --> 06:30.360 is really stressful and computational, 06:30.360 --> 06:32.320 I'm sure, difficult, we'll get into it. 06:32.320 --> 06:36.040 But for me, Battle.net, 06:36.040 --> 06:41.040 so StarCraft was released in 98, 20 years ago, 06:42.640 --> 06:44.680 which is hard to believe, 06:44.680 --> 06:49.680 and Blizzard Battle.net with Diablo 96 came out, 06:50.200 --> 06:52.600 and to me, it might be a narrow perspective, 06:52.600 --> 06:54.520 but it changed online gaming, 06:54.520 --> 06:57.280 and perhaps society forever, 06:57.280 --> 07:00.320 but I may have made way too narrow a viewpoint, 07:00.320 --> 07:02.240 but from your perspective, 07:03.280 --> 07:05.600 can you talk about the history of gaming 07:05.600 --> 07:07.000 over the past 20 years? 07:07.000 --> 07:09.640 Is this, how transformational, 07:09.640 --> 07:12.760 how important is this line of games? 07:12.760 --> 07:16.920 Right, so I think I kind of was an active gamer 07:16.920 --> 07:20.560 whilst this was developing the internet and online gaming, 07:20.560 --> 07:24.040 so for me, the way it came was I played 07:24.040 --> 07:26.400 other games strategy related, 07:26.400 --> 07:28.400 I played a bit of Common and Conquer, 07:28.400 --> 07:31.840 and then I played Warcraft 2, which is from Blizzard, 07:31.840 --> 07:33.040 but at the time, I didn't know, 07:33.040 --> 07:36.000 I didn't understand about what Blizzard was or anything. 07:36.000 --> 07:37.320 Warcraft 2 was just a game, 07:37.320 --> 07:40.240 which was actually very similar to StarCraft in many ways. 07:40.240 --> 07:42.480 It's also a real time strategy game 07:42.480 --> 07:45.400 where there's orcs and humans, so there's only two races. 07:45.400 --> 07:47.960 But it was offline, and it was offline, right? 07:47.960 --> 07:51.600 So I remember a friend of mine came to school, 07:51.600 --> 07:53.960 say, oh, there's this new cool game called StarCraft, 07:53.960 --> 07:55.400 and I just said, oh, this sounds like 07:55.400 --> 07:59.720 just a copy of Warcraft 2, until I kind of installed it, 07:59.720 --> 08:01.960 and at the time, I am from Spain, 08:01.960 --> 08:04.640 so we didn't have like very good internet, right? 08:04.640 --> 08:06.160 So there was, for us, 08:06.160 --> 08:09.520 StarCraft became first kind of an offline experience 08:09.520 --> 08:12.920 where you kind of start to play these missions, right? 08:12.920 --> 08:15.440 You play against some sort of scripted things 08:15.440 --> 08:18.920 to develop the story of the characters in the game, 08:18.920 --> 08:23.480 and then later on, I start playing against the built in AI, 08:23.480 --> 08:26.080 and I thought it was impossible to defeat it. 08:26.080 --> 08:27.400 Then eventually, you defeat one, 08:27.400 --> 08:29.680 and you can actually play against seven built in AIs 08:29.680 --> 08:31.040 at the same time, 08:31.040 --> 08:32.680 which also felt impossible, 08:32.680 --> 08:35.280 but actually, it's not that hard to beat 08:35.280 --> 08:36.960 seven built in AIs at once. 08:36.960 --> 08:39.160 So once we achieved that, 08:39.160 --> 08:42.120 also we discovered that we could play, 08:42.120 --> 08:43.840 as I said, internet wasn't that great, 08:43.840 --> 08:45.920 but we could play with the LAN, right? 08:45.920 --> 08:48.040 Like basically against each other 08:48.040 --> 08:49.920 if we were in the same place, 08:49.920 --> 08:53.640 because you could just connect machines with like cables, right? 08:53.640 --> 08:55.960 So we started playing in LAN mode, 08:55.960 --> 08:58.080 and against, you know, as a group of friends, 08:58.080 --> 09:00.520 and it was really, really like much more entertaining 09:00.520 --> 09:02.280 than playing against the AIs. 09:02.280 --> 09:05.120 And later on, as internet was starting to develop 09:05.120 --> 09:07.400 and being a bit faster and more reliable, 09:07.400 --> 09:09.720 then it's when I started experiencing Battle.net, 09:09.720 --> 09:11.560 which is these amazing universe, 09:11.560 --> 09:13.720 not only because of the fact 09:13.720 --> 09:16.440 that you can play the game against anyone in the world, 09:16.440 --> 09:20.200 but you can also get to know more people. 09:20.200 --> 09:23.080 You just get exposed to now like this vast variety of, 09:23.080 --> 09:25.320 it's kind of a bit when the chats came about, right? 09:25.320 --> 09:27.320 There was a chat system, 09:27.320 --> 09:29.040 you could play against people, 09:29.040 --> 09:30.760 but you could also chat with people, 09:30.760 --> 09:32.520 not only about Stacker, but about anything. 09:32.520 --> 09:36.680 And that became a way of life for kind of two years. 09:36.680 --> 09:38.920 And obviously then it became like kind of, 09:38.920 --> 09:42.280 it exploded in me that I started to play more seriously, 09:42.280 --> 09:44.720 going to tournaments and so on and so forth. 09:44.720 --> 09:49.720 Do you have a sense on a societal sociological level 09:49.880 --> 09:52.280 what's this whole part of society 09:52.280 --> 09:53.840 that many of us are not aware of? 09:53.840 --> 09:56.920 And it's a huge part of society, which is gamers. 09:56.920 --> 10:01.000 I mean, every time I come across that in YouTube 10:01.000 --> 10:03.000 or streaming sites, 10:03.000 --> 10:07.640 I mean, this is a huge number of people play games religiously. 10:07.640 --> 10:08.920 Do you have a sense of those folks, 10:08.920 --> 10:10.880 especially now that you've returned to that realm 10:10.880 --> 10:12.600 a little bit on the AI side? 10:12.600 --> 10:15.880 Yeah, so in fact, even after Stacker, 10:15.880 --> 10:17.640 I actually played World of Warcraft, 10:17.640 --> 10:22.320 which is maybe the main sort of online world and presence 10:22.320 --> 10:24.640 that you get to interact with lots of people. 10:24.640 --> 10:26.400 So I played that for a little bit. 10:26.400 --> 10:29.080 To me, it was a bit less stressful than Starcraft 10:29.080 --> 10:30.920 because winning was kind of a given. 10:30.920 --> 10:32.400 You just put in this world 10:32.400 --> 10:35.040 and you can always complete missions. 10:35.040 --> 10:38.120 But I think it was actually the social aspect 10:38.120 --> 10:40.480 of especially Starcraft first 10:40.480 --> 10:43.440 and then games like World of Warcraft 10:43.440 --> 10:47.000 really shaped me in a very interesting ways 10:47.000 --> 10:48.560 because what you get to experience 10:48.560 --> 10:51.680 is just people you wouldn't usually interact with, right? 10:51.680 --> 10:55.000 So even nowadays, I still have many Facebook friends 10:55.000 --> 10:56.960 from the area where I played online 10:56.960 --> 11:00.120 and their ways of thinking is even political. 11:00.120 --> 11:01.680 They just don't, we don't live in it. 11:01.680 --> 11:03.720 Like we don't interact in the real world, 11:03.720 --> 11:06.800 but we were connected by basically fiber. 11:06.800 --> 11:10.840 And that way I actually get to understand a bit better 11:10.840 --> 11:12.840 that we live in a diverse world. 11:12.840 --> 11:15.640 And these were just connections that were made by, 11:15.640 --> 11:18.120 because I happened to go in a city, 11:18.120 --> 11:20.720 in a virtual city as a priest 11:20.720 --> 11:23.680 and I met this warrior and we became friends. 11:23.680 --> 11:25.720 And then we started like playing together, right? 11:25.720 --> 11:28.800 So I think it's transformative 11:28.800 --> 11:31.320 and more and more and more people are more aware of it. 11:31.320 --> 11:33.520 I mean, it's becoming quite mainstream. 11:33.520 --> 11:35.360 But back in the day, as you were saying, 11:35.360 --> 11:40.360 in 2005 even it was very, still very strange thing to do 11:42.120 --> 11:45.920 especially in Europe, I think there were exceptions 11:45.920 --> 11:47.960 like Korea for instance, it was amazing 11:47.960 --> 11:50.600 like that everything happened so early 11:50.600 --> 11:52.240 in terms of cybercafes. 11:52.240 --> 11:55.120 Like it's, if you go to Seoul, it's a city 11:55.120 --> 11:58.400 that back in the day, StarCraft was kind of, 11:58.400 --> 12:00.640 you could be a celebrity by playing StarCraft 12:00.640 --> 12:03.040 but this was like 99, 2000, right? 12:03.040 --> 12:04.160 It's not like recently. 12:04.160 --> 12:08.560 So yeah, it's quite interesting to look back 12:08.560 --> 12:10.760 and yeah, I think it's changing society. 12:10.760 --> 12:13.120 The same way of course, like technology 12:13.120 --> 12:16.920 and social networks and so on are also transforming things. 12:16.920 --> 12:18.480 And a quick tangent, let me ask, 12:18.480 --> 12:21.000 you're also one of the most productive people 12:21.000 --> 12:26.000 in your particular chosen passion and path in life. 12:26.440 --> 12:29.480 And yet you're also appreciate and enjoy video games. 12:29.480 --> 12:34.480 Do you think it's possible to enjoy video games in moderation? 12:35.800 --> 12:39.920 Someone told me that you could choose two out of three. 12:39.920 --> 12:41.160 When I was playing video games, 12:41.160 --> 12:43.680 you could choose having a girlfriend, 12:43.680 --> 12:46.240 playing video games or studying. 12:46.240 --> 12:50.560 And I think for the most part it was relatively true. 12:50.560 --> 12:52.360 These things do take time. 12:52.360 --> 12:55.400 Games like StarCraft, if you take the game pretty seriously 12:55.400 --> 12:56.520 and you wanna study it, 12:56.520 --> 12:59.080 then you obviously will dedicate more time to it. 12:59.080 --> 13:01.200 And I definitely took gaming 13:01.200 --> 13:03.680 and obviously studying very seriously. 13:03.680 --> 13:08.680 I love learning science and et cetera. 13:08.720 --> 13:13.120 So to me, especially when I started university undergrad, 13:13.120 --> 13:14.920 I kind of stepped off StarCraft. 13:14.920 --> 13:16.800 I actually fully stopped playing. 13:16.800 --> 13:19.040 And then World of Warcraft was a bit more casual. 13:19.040 --> 13:22.920 You could just connect online and I mean, it was fun. 13:22.920 --> 13:26.840 But as I said, that was not as much time investment 13:26.840 --> 13:29.480 as it was for me in StarCraft. 13:29.480 --> 13:31.640 Okay, so let's get into AlphaStar. 13:31.640 --> 13:35.200 What are the, you're behind the team. 13:35.200 --> 13:37.240 So DeepMind has been working on StarCraft 13:37.240 --> 13:39.440 and released a bunch of cool open source agents 13:39.440 --> 13:41.320 and so on in the past few years. 13:41.320 --> 13:43.240 But AlphaStar really is the moment 13:43.240 --> 13:48.240 where the first time you beat a world class player. 13:49.160 --> 13:51.600 So what are the parameters of the challenge 13:51.600 --> 13:53.480 in the way that AlphaStar took it on 13:53.480 --> 13:57.440 and how did you and David and the rest of the DeepMind team 13:57.440 --> 13:58.280 get into it? 13:58.280 --> 14:00.960 Consider that you can even beat the best in the world 14:00.960 --> 14:02.480 or top players. 14:02.480 --> 14:07.480 I think it all started in, back in 2015, actually I'm lying. 14:07.480 --> 14:12.480 I think it was 2014 when DeepMind was acquired by Google 14:12.760 --> 14:14.320 and I at the time was at Google Brain, 14:14.320 --> 14:17.600 which is it was in California, it's still in California. 14:17.600 --> 14:20.600 We had this summit where we got together the two groups. 14:20.600 --> 14:23.400 So Google Brain and Google DeepMind got together 14:23.400 --> 14:25.200 and we gave a series of talks. 14:25.200 --> 14:28.520 And given that they were doing deep reinforcement learning 14:28.520 --> 14:32.200 for games, I decided to bring up part of my past 14:32.200 --> 14:35.000 which I had developed at Berkeley like this thing 14:35.000 --> 14:36.440 which we call Berkeley Overmind 14:36.440 --> 14:39.560 which is really just a StarCraft one bot. 14:39.560 --> 14:41.640 So I talked about that 14:41.640 --> 14:43.800 and I remember them is just came to me and said, 14:43.800 --> 14:46.640 well, maybe not now, it's perhaps a bit too early 14:46.640 --> 14:51.080 but you should just come to DeepMind and do this again 14:51.080 --> 14:53.240 with deep reinforcement learning. 14:53.240 --> 14:56.120 And at the time it sounded very science fiction 14:56.120 --> 14:58.280 for several reasons. 14:58.280 --> 15:01.000 But then in 2016, when I actually moved to London 15:01.000 --> 15:04.280 and joined DeepMind transferring from Brain, 15:04.280 --> 15:07.840 it became apparent that because of the AlphaGo moment 15:07.840 --> 15:11.280 and kind of Blizzard reaching out to us to say, 15:11.280 --> 15:13.000 wait, like, do you want the next challenge 15:13.000 --> 15:15.080 and also me being full time at DeepMind? 15:15.080 --> 15:17.440 So sort of kind of all these came together. 15:17.440 --> 15:21.000 And then I went to Irvine in California 15:21.000 --> 15:23.800 to the Blizzard headquarters to just chat with them 15:23.800 --> 15:26.320 and try to explain how would it all work 15:26.320 --> 15:27.800 before you do anything. 15:27.800 --> 15:31.240 And the approach has always been about 15:32.120 --> 15:33.680 the learning perspective, right? 15:33.680 --> 15:38.680 So in Berkeley, we did a lot of rule based conditioning 15:39.200 --> 15:42.560 and if you have more than three units, then go attack 15:42.560 --> 15:45.040 and if the other has more units than me, I retreat 15:45.040 --> 15:46.400 and so on and so forth. 15:46.400 --> 15:48.840 And of course, the point of deep reinforcement learning, 15:48.840 --> 15:50.520 deep learning, machine learning in general 15:50.520 --> 15:53.480 is that all these should be learned behavior. 15:53.480 --> 15:57.000 So that kind of was the DNA of the project 15:57.000 --> 15:59.520 since its inception in 2016 15:59.520 --> 16:02.920 where we just didn't even have an environment to work with. 16:02.920 --> 16:05.840 And so that's how it all started really. 16:05.840 --> 16:08.600 So if you go back to that conversation with Demis 16:08.600 --> 16:12.200 or even in your own head, how far away did you, 16:12.200 --> 16:14.480 because we're talking about Atari games, 16:14.480 --> 16:16.720 we're talking about Go, which is kind of, 16:16.720 --> 16:19.680 if you're honest about it, really far away from StarCraft. 16:21.120 --> 16:22.160 Well, now that you've beaten it, 16:22.160 --> 16:23.280 maybe you could say it's close, 16:23.280 --> 16:27.160 but it seems like StarCraft is way harder than Go, 16:28.040 --> 16:30.840 philosophically and mathematically speaking. 16:30.840 --> 16:35.040 So how far away did you think you were? 16:35.040 --> 16:38.000 Do you think it's 2019 and 18 you could be doing 16:38.000 --> 16:38.840 as well as you have? 16:38.840 --> 16:40.880 Yeah, when I kind of thought about, 16:40.880 --> 16:44.880 okay, I'm gonna dedicate a lot of my time and focus on this. 16:44.880 --> 16:48.080 And obviously I do a lot of different research 16:48.080 --> 16:50.400 in deep learning, so spending time on it. 16:50.400 --> 16:52.240 I mean, I really had to kind of think 16:52.240 --> 16:55.880 there's gonna be something good happening out of this. 16:55.880 --> 16:59.120 So really I thought, well, this sounds impossible 16:59.120 --> 17:01.600 and it probably is impossible to do the full thing, 17:01.600 --> 17:06.600 like the full game where you play one versus one 17:06.720 --> 17:09.760 and it's only a neural network playing and so on. 17:09.760 --> 17:10.960 So it really felt like, 17:10.960 --> 17:14.000 I just didn't even think it was possible. 17:14.000 --> 17:14.840 But on the other hand, 17:14.840 --> 17:19.080 I could see some stepping stones towards that goal. 17:19.080 --> 17:21.600 Clearly you could define sub problems in StarCraft 17:21.600 --> 17:23.360 and sort of dissect it a bit and say, 17:23.360 --> 17:26.720 okay, here is a part of the game, here is another part. 17:26.720 --> 17:29.400 And also obviously the fact, 17:29.400 --> 17:31.280 so this was really also critical to me, 17:31.280 --> 17:34.400 the fact that we could access human replays. 17:34.400 --> 17:35.720 So Blizzard was very kind 17:35.720 --> 17:38.560 and in fact they open source this for the whole community 17:38.560 --> 17:39.960 where you can just go 17:39.960 --> 17:43.040 and it's not every single StarCraft game ever played, 17:43.040 --> 17:44.200 but it's a lot of them. 17:44.200 --> 17:45.880 You can just go and download 17:45.880 --> 17:47.160 and every day they will, 17:47.160 --> 17:48.960 you can just query a data set and say, 17:48.960 --> 17:51.680 well, give me all the games that were played today. 17:51.680 --> 17:55.800 And given my kind of experience with language 17:55.800 --> 17:57.920 and sequences and supervised learning, 17:57.920 --> 18:00.760 I thought, well, that's definitely gonna be very helpful 18:00.760 --> 18:02.400 and something quite unique now 18:02.400 --> 18:06.720 because ever before we had such a large data set 18:06.720 --> 18:11.000 of replays of people playing the game at this scale 18:11.000 --> 18:12.560 of such a complex video game, right? 18:12.560 --> 18:15.640 So that to me was a precious resource. 18:15.640 --> 18:18.000 And as soon as I knew that Blizzard was able 18:18.000 --> 18:20.960 to kind of give this to the community, 18:20.960 --> 18:22.280 I started to feel positive 18:22.280 --> 18:24.280 about something non trivial happening. 18:24.280 --> 18:27.120 But I also thought the full thing, 18:27.120 --> 18:30.400 like really no rules, no single line of code 18:30.400 --> 18:31.680 that tries to say, well, I mean, 18:31.680 --> 18:33.320 if you see this unit build a detector, 18:33.320 --> 18:36.680 all these, not having any of these specializations 18:36.680 --> 18:39.160 seemed really, really, really difficult to me. 18:39.160 --> 18:40.000 Intuitively. 18:40.000 --> 18:42.680 I do also like that Blizzard was teasing 18:42.680 --> 18:43.960 or even trolling you, 18:45.480 --> 18:48.560 sort of almost pulling you in 18:48.560 --> 18:50.280 into this really difficult challenge. 18:50.280 --> 18:51.840 Do they have any awareness? 18:51.840 --> 18:55.640 What's the interest from the perspective of Blizzard 18:55.640 --> 18:57.280 except just curiosity? 18:57.280 --> 18:59.400 Yeah, I think Blizzard has really understood 18:59.400 --> 19:03.240 and really bring forward this competitiveness 19:03.240 --> 19:04.800 of eSports in games. 19:04.800 --> 19:07.840 The StarCraft really kind of sparked a lot of, 19:07.840 --> 19:10.720 like something that almost was never seen, 19:10.720 --> 19:13.960 especially as I was saying, back in Korea. 19:13.960 --> 19:16.480 So they just probably thought, well, 19:16.480 --> 19:18.880 this is such a pure one versus one setup 19:18.880 --> 19:21.160 that it would be great to see 19:21.160 --> 19:24.840 if something that can play Atari or go 19:24.840 --> 19:27.920 and then later on chess could even tackle 19:27.920 --> 19:30.600 these kind of complex real time strategy game, right? 19:30.600 --> 19:33.880 So for them, they wanted to see first, obviously, 19:33.880 --> 19:36.440 whether it was possible, 19:36.440 --> 19:39.760 if the game they created was in a way solvable, 19:39.760 --> 19:40.840 to some extent. 19:40.840 --> 19:42.200 And I think on the other hand, 19:42.200 --> 19:45.760 they also are a pretty modern company that innovates a lot. 19:45.760 --> 19:48.520 So just starting to understand AI for them 19:48.520 --> 19:50.240 to how to bring AI into games, 19:50.240 --> 19:54.320 is not AI for games, but games for AI, right? 19:54.320 --> 19:56.120 I mean, both ways, I think, can work. 19:56.120 --> 20:00.040 And we obviously had the manuse games for AI, right? 20:00.040 --> 20:01.280 To drive AI progress, 20:01.280 --> 20:03.680 but Blizzard might actually be able to do, 20:03.680 --> 20:04.760 and many other companies, 20:04.760 --> 20:06.800 to start to understand and do the opposite. 20:06.800 --> 20:09.800 So I think that is also something they can get out of this. 20:09.800 --> 20:11.320 And they definitely, 20:11.320 --> 20:13.720 we have brainstormed a lot about this, right? 20:13.720 --> 20:16.080 But one of the interesting things to me about StarCraft 20:16.080 --> 20:19.400 and Diablo and these games that Blizzard has created 20:19.400 --> 20:23.560 is the task of balancing classes, for example, 20:23.560 --> 20:27.480 sort of making the game fair from the starting point, 20:27.480 --> 20:29.920 and then let skill determine the outcome. 20:30.960 --> 20:33.600 Is there, I mean, can you first comment? 20:33.600 --> 20:36.760 There's three races, Zerg, Protoss, and Terran. 20:36.760 --> 20:38.960 I don't know if I've ever said that out loud. 20:38.960 --> 20:40.600 Is that how you pronounce it, Terran? 20:40.600 --> 20:41.600 Yeah, Terran. 20:41.600 --> 20:42.440 Yeah. 20:42.440 --> 20:45.200 Yeah, I don't think I've ever, 20:45.200 --> 20:47.680 in person, interacted with anybody about StarCraft. 20:47.680 --> 20:51.760 That's funny. So they seem to be pretty balanced. 20:51.760 --> 20:56.240 I wonder if the AI, the work that you're doing 20:56.240 --> 20:59.160 with AlphaStar would help balance them even further. 20:59.160 --> 21:00.520 Is that something you think about? 21:00.520 --> 21:03.280 Is that something that Blizzard is thinking about? 21:03.280 --> 21:06.400 Right, so balancing when you add a new unit 21:06.400 --> 21:09.120 or a new spell type is obviously possible, 21:09.120 --> 21:13.160 given that you can always train or pre train at scale, 21:13.160 --> 21:16.680 some agent that might start using that in unintended ways. 21:16.680 --> 21:19.120 But I think actually, if you understand 21:19.120 --> 21:22.200 how StarCraft has kind of co evolved with players, 21:22.200 --> 21:24.280 in a way, I think it's actually very cool, 21:24.280 --> 21:27.400 the ways that many of the things and strategies 21:27.400 --> 21:28.680 that people came up with, right? 21:28.680 --> 21:32.280 So I think it's, we've seen it over and over in StarCraft 21:32.280 --> 21:34.920 that Blizzard comes up with maybe a new unit, 21:34.920 --> 21:37.240 and then some players get creative 21:37.240 --> 21:39.080 and do something kind of unintentional 21:39.080 --> 21:40.840 or something that Blizzard designers 21:40.840 --> 21:43.560 that just simply didn't test or think about. 21:43.560 --> 21:46.960 And then after that becomes kind of mainstream in the community, 21:46.960 --> 21:48.240 Blizzard patches the game, 21:48.240 --> 21:51.880 and then they kind of maybe weaken that strategy 21:51.880 --> 21:53.880 or make it actually more interesting, 21:53.880 --> 21:55.400 but a bit more balanced. 21:55.400 --> 21:58.280 So this kind of continual talk between players and Blizzard 21:58.280 --> 22:01.680 is kind of what has defined them actually, 22:01.680 --> 22:04.040 in actually most games, like in StarCraft, 22:04.040 --> 22:05.760 but also in World of Warcraft, 22:05.760 --> 22:07.440 they would do that, there are several classes 22:07.440 --> 22:10.800 and it would be not good that everyone plays 22:10.800 --> 22:13.200 absolutely the same race and so on, right? 22:13.200 --> 22:17.280 So I think they do care about balancing, of course, 22:17.280 --> 22:19.640 and they do a fair amount of testing, 22:19.640 --> 22:22.160 but it's also beautiful to also see 22:22.160 --> 22:24.520 how players get creative anyways. 22:24.520 --> 22:27.440 And I mean, whether AI can be more creative at this point, 22:27.440 --> 22:28.680 I don't think so, right? 22:28.680 --> 22:31.600 I mean, it's just sometimes something so amazing happens, 22:31.600 --> 22:33.720 like I remember back in the days, 22:33.720 --> 22:36.920 like you have these drop ships that could drop the rivers, 22:36.920 --> 22:39.600 and that was actually not thought about, 22:39.600 --> 22:41.280 that you could drop this unit 22:41.280 --> 22:43.200 that has this what's called splash damage 22:43.200 --> 22:47.800 that would basically eliminate all the enemy's workers at once. 22:47.800 --> 22:50.080 No one thought that you could actually put them 22:50.080 --> 22:53.040 in really early game, do that kind of damage, 22:53.040 --> 22:55.400 and then things change in the game, 22:55.400 --> 22:58.000 but I don't know, I think it's quite an amazing 22:58.000 --> 23:00.280 exploration process from both sides, 23:00.280 --> 23:01.840 players and Blizzard alike. 23:01.840 --> 23:05.000 Well, it's almost like a reinforcement learning exploration, 23:05.000 --> 23:10.000 but the scale of humans that play Blizzard games 23:10.000 --> 23:13.680 is almost on the scale of a large scale, 23:13.680 --> 23:15.360 deep mind RL experiment. 23:15.360 --> 23:17.200 I mean, if you look at the numbers, 23:17.200 --> 23:18.720 that's, I mean, you're talking about, 23:18.720 --> 23:19.560 I don't know how many games, 23:19.560 --> 23:22.080 but hundreds of thousands of games, probably a month. 23:22.080 --> 23:23.880 Yeah, I mean, so you could, 23:23.880 --> 23:28.800 it's almost the same as running RL agents. 23:28.800 --> 23:31.240 What aspect of the problem of Starcraft, 23:31.240 --> 23:32.160 do you think is the hardest? 23:32.160 --> 23:35.400 Is it the, like you said, the imperfect information? 23:35.400 --> 23:38.160 Is it the fact they have to do longterm planning? 23:38.160 --> 23:40.280 Is it the real time aspect? 23:40.280 --> 23:42.240 So you have to do stuff really quickly? 23:42.240 --> 23:44.760 Is it the fact that large action space, 23:44.760 --> 23:47.640 so you can do so many possible things? 23:47.640 --> 23:51.120 Or is it, you know, in the game theoretic sense, 23:51.120 --> 23:52.440 there is no Nash equilibrium. 23:52.440 --> 23:54.240 At least you don't know what the optimal strategy is, 23:54.240 --> 23:56.520 because there's way too many options. 23:56.520 --> 23:57.360 Right. 23:57.360 --> 23:58.360 Is there something that stands out 23:58.360 --> 24:01.000 as just like the hardest, the most annoying thing? 24:01.000 --> 24:04.200 So when we sort of looked at the problem 24:04.200 --> 24:07.640 and start to define the parameters of it, right? 24:07.640 --> 24:08.800 What are the observations? 24:08.800 --> 24:10.520 What are the actions? 24:10.520 --> 24:13.920 It became very apparent that, you know, 24:13.920 --> 24:17.160 the very first barrier that one would hit in Starcraft 24:17.160 --> 24:20.720 would be because of the action space being so large 24:20.720 --> 24:24.880 and as not being able to search like you could in chess 24:24.880 --> 24:27.320 or go even though the search space is vast. 24:28.640 --> 24:30.600 The main problem that we identified 24:30.600 --> 24:32.440 was that of exploration, right? 24:32.440 --> 24:36.720 So without any sort of human knowledge or human prior, 24:36.720 --> 24:38.040 if you think about Starcraft 24:38.040 --> 24:41.440 and you know how deep reinforcement learning algorithm works, 24:41.440 --> 24:45.360 work, which is essentially by issuing random actions 24:45.360 --> 24:47.800 and hoping that they will get some wins sometimes 24:47.800 --> 24:49.200 so they could learn. 24:49.200 --> 24:52.800 So if you think of the action space in Starcraft, 24:52.800 --> 24:55.880 almost anything you can do in the early game is bad 24:55.880 --> 24:58.720 because any action involves taking workers 24:58.720 --> 25:01.360 which are mining minerals for free. 25:01.360 --> 25:03.560 That's something that the game does automatically, 25:03.560 --> 25:04.920 sends them to mine 25:04.920 --> 25:07.720 and you would immediately just take them out of mining 25:07.720 --> 25:09.080 and send them around. 25:09.080 --> 25:13.640 So just thinking how is it gonna be possible 25:13.640 --> 25:16.880 to get to understand these concepts 25:16.880 --> 25:19.280 but even more like expanding, right? 25:19.280 --> 25:21.080 There's these buildings you can place 25:21.080 --> 25:24.160 in other locations in the map to gather more resources 25:24.160 --> 25:26.840 but the location of the building is important 25:26.840 --> 25:28.920 and you have to select a worker, 25:28.920 --> 25:32.680 send it walking to that location, build the building, 25:32.680 --> 25:34.160 wait for the building to be built 25:34.160 --> 25:37.840 and then put extra workers there so they start mining. 25:37.840 --> 25:40.200 That just, that feels like impossible 25:40.200 --> 25:43.680 if you just randomly click to produce that state, 25:43.680 --> 25:47.000 desirable state that then you could hope to learn from 25:47.000 --> 25:49.880 because eventually that may yield to an extra win, right? 25:49.880 --> 25:51.840 So for me, the exploration problem 25:51.840 --> 25:53.840 and due to the action space 25:53.840 --> 25:56.160 and the fact that there's not really turns, 25:56.160 --> 25:57.000 there's so many turns 25:57.000 --> 26:01.440 because the game essentially ticks at 22 times per second. 26:01.440 --> 26:05.560 If you, I mean, that's how they discretize sort of time. 26:05.560 --> 26:07.320 Obviously, you always have to discretize time 26:07.320 --> 26:09.640 where there's no such thing as real time 26:09.640 --> 26:12.560 but it's really a lot of time steps 26:12.560 --> 26:14.280 of things that could go wrong 26:14.280 --> 26:17.960 and that definitely felt a priori like the hardest. 26:17.960 --> 26:19.360 You mentioned many good ones, 26:19.360 --> 26:21.360 I think partial observability, 26:21.360 --> 26:23.440 the fact that there is no perfect strategy 26:23.440 --> 26:25.560 because of the partial observability, 26:25.560 --> 26:26.880 those are very interesting problems. 26:26.880 --> 26:29.400 We start seeing more and more now in terms of 26:29.400 --> 26:31.080 as we saw of the previous ones 26:31.080 --> 26:34.320 but the core problem to me was exploration 26:34.320 --> 26:37.800 and solving it has been basically kind of the focus 26:37.800 --> 26:39.840 and how we saw the first breakthroughs. 26:39.840 --> 26:43.720 So exploration in a multi hierarchical way. 26:43.720 --> 26:46.640 So like 22 times a second exploration 26:46.640 --> 26:48.680 has a very different meaning than it does 26:48.680 --> 26:51.520 in terms of should I gather resources early 26:51.520 --> 26:53.200 or should I wait or so on. 26:53.200 --> 26:56.240 So how do you solve the long term? 26:56.240 --> 26:58.120 Let's talk about the internals of Alpha Star. 26:58.120 --> 27:02.520 So first of all, how do you represent the state 27:02.520 --> 27:05.160 of the game as an input? 27:05.160 --> 27:08.840 How do you then do the long term sequence modeling? 27:08.840 --> 27:10.480 How do you build a policy? 27:10.480 --> 27:12.600 What's the architecture like? 27:12.600 --> 27:16.880 So Alpha Star has obviously several components 27:16.880 --> 27:20.920 but everything passes through what we call the policy 27:20.920 --> 27:22.320 which is a neural network 27:22.320 --> 27:24.320 and that's kind of the beauty of it. 27:24.320 --> 27:27.200 There is, I could just now give you a neural network 27:27.200 --> 27:30.480 and some weights and if you fed the right observations 27:30.480 --> 27:32.600 and you understood the actions the same way we do 27:32.600 --> 27:35.160 you would have basically the agent playing the game. 27:35.160 --> 27:37.280 There's absolutely nothing else needed 27:37.280 --> 27:40.360 other than those weights that were trained. 27:40.360 --> 27:43.400 Now, the first step is observing the game 27:43.400 --> 27:46.680 and we've experimented with a few alternatives. 27:46.680 --> 27:48.800 The one that we currently use 27:48.800 --> 27:51.440 mixes both spatial sort of images 27:51.440 --> 27:53.840 that you would process from the game 27:53.840 --> 27:56.440 that is the zoomed out version of the map 27:56.440 --> 27:58.960 and also a zoomed in version of the camera 27:58.960 --> 28:00.880 or the screen as we call it. 28:00.880 --> 28:04.840 But also we give to the agent the list of units 28:04.840 --> 28:09.000 that it sees more of as a set of objects 28:09.000 --> 28:11.040 that it can operate on. 28:11.040 --> 28:14.760 That is not necessarily required to use it 28:14.760 --> 28:16.840 and we have versions of the game that play well 28:16.840 --> 28:19.080 without this set vision that is a bit 28:19.080 --> 28:21.680 not like how humans perceive the game 28:21.680 --> 28:23.640 but it certainly helps a lot 28:23.640 --> 28:25.040 because it's a very natural way 28:25.040 --> 28:28.480 to encode the game is by just looking at all the units 28:28.480 --> 28:32.960 that they have properties like health, position, 28:32.960 --> 28:36.200 type of unit, whether it's my unit or the enemy's 28:36.200 --> 28:40.800 and that sort of is kind of the summary 28:40.800 --> 28:42.880 of the state of the game, 28:42.880 --> 28:45.520 not that list of units or set of units 28:45.520 --> 28:47.400 that you see all the time. 28:47.400 --> 28:49.600 But that's pretty close to the way humans see the game. 28:49.600 --> 28:51.440 Why do you say it's not, 28:51.440 --> 28:53.240 you're saying the exactness of it 28:53.240 --> 28:55.080 is not similar to humans? 28:55.080 --> 28:57.200 The exactness of it is perhaps not the problem. 28:57.200 --> 28:59.840 I guess maybe the problem if you look at it 28:59.840 --> 29:02.320 from how actually humans play the game 29:02.320 --> 29:05.760 is that they play with a mouse and a keyboard and a screen 29:05.760 --> 29:08.760 and they don't see sort of a structured object 29:08.760 --> 29:09.600 with all the units, 29:09.600 --> 29:12.240 what they see is what they see on the screen, right? 29:12.240 --> 29:14.400 So you remember that there's a certain interrupt, 29:14.400 --> 29:17.000 there's a plot that you showed with camera base 29:17.000 --> 29:19.680 where you do exactly that, right, you move around 29:19.680 --> 29:22.280 and that seems to converge to similar performance. 29:22.280 --> 29:24.760 Yeah, I think that's what we're kind of experimenting 29:24.760 --> 29:28.720 with what's necessary or not, but using the set. 29:28.720 --> 29:32.360 So actually if you look at research in computer vision 29:32.360 --> 29:36.000 where it makes a lot of sense to treat images 29:36.000 --> 29:38.160 as two dimensional arrays, 29:38.160 --> 29:40.360 there's actually a very nice paper from Facebook. 29:40.360 --> 29:42.720 I think, I forgot who the authors are, 29:42.720 --> 29:46.360 but I think it's part of Kmings has group. 29:46.360 --> 29:49.520 And what they do is they take an image, 29:49.520 --> 29:51.960 which is this two dimensional signal 29:51.960 --> 29:54.320 and they actually take pixel by pixel 29:54.320 --> 29:59.160 and scramble the image as if it was just a list of pixels. 29:59.160 --> 30:01.800 Crucially, they encode the position of the pixels 30:01.800 --> 30:03.720 with the XY coordinates. 30:03.720 --> 30:06.160 And this is just kind of a new architecture 30:06.160 --> 30:08.520 which we incidentally also use in stack graph 30:08.520 --> 30:09.880 called the transformer, 30:09.880 --> 30:12.000 which is a very popular paper from last year, 30:12.000 --> 30:15.600 which yielded very nice result in machine translation. 30:15.600 --> 30:18.040 And if you actually believe in this kind of, 30:18.040 --> 30:20.320 oh, it's actually a set of pixels 30:20.320 --> 30:22.520 as long as you encode XY, it's okay. 30:22.520 --> 30:25.560 Then you could argue that the list of units 30:25.560 --> 30:26.960 that we see is precisely that 30:26.960 --> 30:31.480 because we have each unit as a kind of pixel, if you will, 30:31.480 --> 30:33.240 and then their XY coordinates. 30:33.240 --> 30:36.360 So in that perspective, without knowing it, 30:36.360 --> 30:37.680 we use the same architecture 30:37.680 --> 30:39.680 that was shown to work very well 30:39.680 --> 30:41.400 on Pascal and ImageNet and so on. 30:41.400 --> 30:45.440 So the interesting thing here is putting it in that way, 30:45.440 --> 30:47.000 it starts to move it towards 30:47.000 --> 30:49.480 the way you usually work with language. 30:49.480 --> 30:52.760 So what, and especially with your expertise 30:52.760 --> 30:57.000 and work in language, it seems like there's echoes 30:57.000 --> 31:00.720 of a lot of the way you would work with natural language 31:00.720 --> 31:02.440 in the way you've approached AlphaStar. 31:02.440 --> 31:05.080 Right, does that help 31:05.080 --> 31:08.200 with the longterm sequence modeling there somehow? 31:08.200 --> 31:11.200 Exactly, so now that we understand what an observation 31:11.200 --> 31:14.680 for a given time step is, we need to move on to say, 31:14.680 --> 31:17.760 well, there's gonna be a sequence of such observations 31:17.760 --> 31:21.120 and an agent will need to, given all that it's seen, 31:21.120 --> 31:23.720 not only the current time step, but all that it's seen, 31:23.720 --> 31:25.920 why, because there is partial observability. 31:25.920 --> 31:28.400 We must remember whether we saw a worker 31:28.400 --> 31:30.120 going somewhere, for instance, right? 31:30.120 --> 31:31.720 Because then there might be an expansion 31:31.720 --> 31:33.640 on the top right of the map. 31:33.640 --> 31:37.840 So given that, what you must then think about 31:37.840 --> 31:40.400 is there is the problem of, given all the observations, 31:40.400 --> 31:42.640 you have to predict the next action. 31:42.640 --> 31:44.520 And not only given all the observations, 31:44.520 --> 31:45.920 but given all the observations 31:45.920 --> 31:47.920 and given all the actions you've taken, 31:47.920 --> 31:49.360 predict the next action. 31:49.360 --> 31:52.480 And that sounds exactly like machine translation, 31:52.480 --> 31:57.160 where, and that's exactly how kind of I saw the problem, 31:57.160 --> 31:59.960 especially when you are given supervised data 31:59.960 --> 32:01.760 or replaced from humans, 32:01.760 --> 32:03.600 because the problem is exactly the same. 32:03.600 --> 32:06.680 You're translating essentially a prefix 32:06.680 --> 32:08.240 of observations and actions 32:08.240 --> 32:10.160 onto what's gonna happen next, 32:10.160 --> 32:13.000 which is exactly how you would train a model to translate 32:13.000 --> 32:14.760 or to generate language as well, right? 32:14.760 --> 32:16.640 You have a certain prefix. 32:16.640 --> 32:19.000 You must remember everything that comes in the past, 32:19.000 --> 32:20.080 because otherwise, 32:20.080 --> 32:22.640 you might start having non coherent text. 32:22.640 --> 32:25.120 And the same architectures, 32:25.120 --> 32:27.760 we're using LSTMs and transformers 32:27.760 --> 32:29.760 to operate on across time 32:29.760 --> 32:33.080 to kind of integrate all that's happened in the past. 32:33.080 --> 32:35.000 Those architectures that work so well 32:35.000 --> 32:36.880 in translation or language modeling 32:36.880 --> 32:40.640 are exactly the same than what the agent is using 32:40.640 --> 32:42.360 to issue actions in the game. 32:42.360 --> 32:43.880 And the way we train it, moreover, 32:43.880 --> 32:47.120 for imitation, which is step one of alpha studies, 32:47.120 --> 32:49.880 take all the human experience and try to imitate it, 32:49.880 --> 32:52.920 much like you try to imitate translators 32:52.920 --> 32:55.360 that translated many pairs of sentences 32:55.360 --> 32:57.280 from French to English say, 32:57.280 --> 33:00.200 that sort of principle applies exactly the same. 33:00.200 --> 33:02.760 It's almost the same code, 33:02.760 --> 33:04.520 except that instead of words, 33:04.520 --> 33:06.680 you have a slightly more complicated objects, 33:06.680 --> 33:08.280 which are the observations 33:08.280 --> 33:10.240 and the actions are also a bit more complicated 33:10.240 --> 33:11.760 than a word. 33:11.760 --> 33:13.920 Is there a self play component then too? 33:13.920 --> 33:16.480 So once you run out of imitation? 33:16.480 --> 33:21.480 Right, so indeed you can bootstrap from human replays, 33:22.240 --> 33:25.960 but then the agents you get are actually not as good 33:25.960 --> 33:28.160 as the humans you imitated, right? 33:28.160 --> 33:30.440 So how do we imitate? 33:30.440 --> 33:34.240 Well, we take humans from 3000 MMR and higher. 33:34.240 --> 33:37.960 3000 MMR is just a metric of human skill. 33:37.960 --> 33:41.880 And 3000 MMR might be like 50% percentile, right? 33:41.880 --> 33:43.760 So it's just average human. 33:43.760 --> 33:44.600 What's that? 33:44.600 --> 33:45.440 So maybe a quick pause. 33:45.440 --> 33:47.760 MMR is a ranking scale, 33:47.760 --> 33:50.320 the matchmaking rating for players. 33:50.320 --> 33:52.320 So it's 3000, I remember there's like a master 33:52.320 --> 33:54.120 and a grandmaster, what's 3000? 33:54.120 --> 33:56.720 So 3000 is pretty bad. 33:56.720 --> 33:58.440 I think it's kind of gold level. 33:58.440 --> 34:00.680 It just sounds really good relative to chess, I think. 34:00.680 --> 34:02.440 Oh yeah, yeah, no, the ratings, 34:02.440 --> 34:04.480 the best in the world are at 7000 MMR. 34:04.480 --> 34:05.320 7000. 34:05.320 --> 34:07.840 So 3000, it's a bit like Elo indeed, right? 34:07.840 --> 34:12.840 So 3500 just allows us to not filter a lot of the data. 34:13.200 --> 34:15.680 So we like to have a lot of data in deep learning 34:15.680 --> 34:17.320 as you probably know. 34:17.320 --> 34:20.640 So we take these kind of 3500 and above, 34:20.640 --> 34:22.720 but then we do a very interesting trick, 34:22.720 --> 34:25.000 which is we tell the neural network 34:25.000 --> 34:27.560 what level they are imitating. 34:27.560 --> 34:30.800 So we say these replay you're gonna try to imitate 34:30.800 --> 34:33.040 to predict the next action for all the actions 34:33.040 --> 34:36.120 that you're gonna see is a 4000 MMR replay. 34:36.120 --> 34:38.800 This one is a 6000 MMR replay. 34:38.800 --> 34:42.480 And what's cool about this is then we take this policy 34:42.480 --> 34:44.280 that is being trained from human 34:44.280 --> 34:47.400 and then we can ask it to play like a 3000 MMR player 34:47.400 --> 34:49.600 by setting a bit saying, well, okay, 34:49.600 --> 34:53.680 play like a 3000 MMR player or play like a 6000 MMR player. 34:53.680 --> 34:57.320 And you actually see how the policy behaves differently. 34:57.320 --> 35:01.520 It gets worse economy if you play like a gold level player. 35:01.520 --> 35:03.000 It does less actions per minute, 35:03.000 --> 35:05.320 which is the number of clicks or number of actions 35:05.320 --> 35:07.760 that you will issue in a whole minute. 35:07.760 --> 35:09.200 And it's very interesting to see 35:09.200 --> 35:12.360 that it kind of imitates the skill level quite well. 35:12.360 --> 35:15.480 But if we ask it to play like a 6000 MMR player, 35:15.480 --> 35:18.600 we tested of course these policies to see how well they do. 35:18.600 --> 35:20.600 They actually beat all the built in AIs 35:20.600 --> 35:22.400 that Blizzard put in the game, 35:22.400 --> 35:24.960 but they're nowhere near 6000 MMR players, right? 35:24.960 --> 35:29.240 They might be maybe around gold level, platinum perhaps. 35:29.240 --> 35:32.200 So there's still a lot of work to be done for the policy 35:32.200 --> 35:34.960 to truly understand what it means to win. 35:34.960 --> 35:38.160 So far we only ask them, okay, here is the screen 35:38.160 --> 35:41.600 and that's what's happened on the game until this point. 35:41.600 --> 35:46.080 What would the next action be if we ask a pro to now say, 35:46.080 --> 35:49.120 oh, you're gonna click here or here or there? 35:49.120 --> 35:53.680 And the point is experiencing wins and losses 35:53.680 --> 35:56.320 is very important to then start to refine. 35:56.320 --> 35:58.360 Otherwise the policy can get loose, 35:58.360 --> 36:00.440 can just go off policy as we call it. 36:00.440 --> 36:03.400 That's so interesting that you can at least hope eventually 36:03.400 --> 36:06.760 to be able to control a policy 36:06.760 --> 36:09.960 approximately to be at some MMR level. 36:09.960 --> 36:12.240 That's so interesting, especially given 36:12.240 --> 36:15.000 that you have ground truth for a lot of these cases. 36:15.000 --> 36:17.520 Can I ask you a personal question? 36:17.520 --> 36:19.200 What's your MMR? 36:19.200 --> 36:23.600 Well, I haven't played Starcraft 2, so I am unranked, 36:23.600 --> 36:25.360 which is the kind of lowest league. 36:25.360 --> 36:26.200 Okay. 36:26.200 --> 36:28.360 So I used to play Starcraft 1. 36:28.360 --> 36:29.560 The first one and... 36:29.560 --> 36:31.280 But you haven't seriously played Starcraft 2? 36:31.280 --> 36:32.640 No, not Starcraft 2. 36:32.640 --> 36:37.720 So the best player we have at DeepMind is about 5,000 MMR, 36:37.720 --> 36:39.560 which is high masters. 36:39.560 --> 36:42.040 It's not at the Grand Master level. 36:42.040 --> 36:44.640 Grand Master level would be the top 200 players 36:44.640 --> 36:49.120 in a certain region, like Europe or America or Asia. 36:49.120 --> 36:51.560 But for me, it would be hard to say. 36:51.560 --> 36:53.680 I am very bad at the game. 36:53.680 --> 36:56.600 I actually played Alpha Star a bit too late and it beat me. 36:56.600 --> 36:59.720 I remember the whole team was, oh, Oreo, you should play. 36:59.720 --> 37:02.160 And I was, oh, it looks like it's not so good yet. 37:02.160 --> 37:06.600 And then I remember I kind of got busy and waited an extra week 37:06.600 --> 37:09.640 and I played and it really beat me very badly. 37:09.640 --> 37:11.160 How did that feel? 37:11.160 --> 37:12.600 Isn't that an amazing feeling? 37:12.600 --> 37:13.560 That's amazing, yeah. 37:13.560 --> 37:17.920 I mean, obviously, I tried my best and I tried to also impress 37:17.920 --> 37:19.720 because I actually played the first game, 37:19.720 --> 37:23.040 so I'm still pretty good at micro management. 37:23.040 --> 37:25.200 The problem is I just don't understand Starcraft 2. 37:25.200 --> 37:27.200 I understand Starcraft. 37:27.200 --> 37:28.440 And when I played Starcraft, 37:28.440 --> 37:32.640 I probably was consistently like for a couple of years, 37:32.640 --> 37:34.600 top 32 in Europe. 37:34.600 --> 37:36.440 So I was decent, but at the time, 37:36.440 --> 37:40.280 we didn't have this kind of MMR system as well established. 37:40.280 --> 37:43.120 So it would be hard to know what it was back then. 37:43.120 --> 37:45.760 So what's the difference in interface between Alpha Star 37:45.760 --> 37:49.600 and Starcraft and a human player in Starcraft? 37:49.600 --> 37:52.000 Is there any significant differences 37:52.000 --> 37:54.080 between the way they both see the game? 37:54.080 --> 37:55.960 I would say the way they see the game, 37:55.960 --> 37:59.720 there's a few things that are just very hard to simulate. 38:01.000 --> 38:02.640 The main one, perhaps, 38:02.640 --> 38:05.200 which is obvious in hindsight, 38:05.200 --> 38:08.440 is what's called clocked units, 38:08.440 --> 38:10.560 which are invisible units. 38:10.560 --> 38:13.240 So in Starcraft, you can make some units 38:13.240 --> 38:18.040 that you need to have a particular kind of unit to detect it. 38:18.040 --> 38:20.560 So these units are invisible. 38:20.560 --> 38:22.720 If you cannot detect them, you cannot target them. 38:22.720 --> 38:25.720 So they would just destroy your buildings 38:25.720 --> 38:27.720 or kill your workers. 38:27.720 --> 38:31.640 But despite the fact you cannot target the unit, 38:31.640 --> 38:34.600 there's a shimmer that as a human you observe. 38:34.600 --> 38:35.920 I mean, you need to train a little bit. 38:35.920 --> 38:37.400 You need to pay attention, 38:37.400 --> 38:41.840 but you would see this kind of space time distortion 38:41.840 --> 38:44.800 and you wouldn't know, okay, there are, yeah. 38:44.800 --> 38:46.040 Yeah, there's like a wave thing. 38:46.040 --> 38:47.840 Yeah, it's called shimmer. 38:47.840 --> 38:49.120 Space time distortion, I like it. 38:49.120 --> 38:52.440 That's really like the blizzard term is shimmer. 38:52.440 --> 38:56.040 And so these shimmer professional players actually 38:56.040 --> 38:57.160 can see it immediately. 38:57.160 --> 38:59.480 They understand it very well, 38:59.480 --> 39:01.400 but it's still something that requires 39:01.400 --> 39:02.720 certain amount of attention 39:02.720 --> 39:05.640 and it's kind of a bit annoying to deal with. 39:05.640 --> 39:08.640 Whereas for Alpha Star, in terms of vision, 39:08.640 --> 39:11.120 it's very hard for us to simulate sort of, 39:11.120 --> 39:14.160 oh, are you looking at this pixel in the screen and so on? 39:14.160 --> 39:17.520 So the only thing we can do is 39:17.520 --> 39:19.720 there is a unit that's invisible over there. 39:19.720 --> 39:22.520 So Alpha Star would know that immediately. 39:22.520 --> 39:24.040 Obviously still obeys the rules. 39:24.040 --> 39:25.200 You cannot attack the unit. 39:25.200 --> 39:27.400 You must have a detector and so on, 39:27.400 --> 39:29.320 but it's kind of one of the main things 39:29.320 --> 39:32.680 that it just doesn't feel there's a very proper way. 39:32.680 --> 39:35.480 I mean, you could imagine, oh, you don't have hypers. 39:35.480 --> 39:36.960 Maybe you don't know exactly what it is, 39:36.960 --> 39:39.240 or sometimes you see it, sometimes you don't. 39:39.240 --> 39:43.040 But it's just really, really complicated to get it 39:43.040 --> 39:44.280 so that everyone would agree, 39:44.280 --> 39:47.120 oh, that's the best way to simulate this, right? 39:47.120 --> 39:49.280 You know, it seems like a perception problem. 39:49.280 --> 39:50.600 It is a perception problem. 39:50.600 --> 39:54.240 So the only problem is people, you ask, 39:54.240 --> 39:56.760 oh, what's the difference between how humans perceive the game? 39:56.760 --> 39:59.960 I would say they wouldn't be able to tell a shimmer 39:59.960 --> 40:02.240 immediately as it appears on the screen, 40:02.240 --> 40:04.320 whereas Alpha Star, in principle, 40:04.320 --> 40:05.640 sees it very sharply, right? 40:05.640 --> 40:08.680 It sees that the bit turned from zero to one, 40:08.680 --> 40:10.520 meaning there's now a unit there, 40:10.520 --> 40:11.960 although you don't know the unit, 40:11.960 --> 40:15.840 or you know that you cannot attack it and so on. 40:15.840 --> 40:18.080 So from a vision standpoint, 40:18.080 --> 40:23.000 that probably is the one that is kind of the most obvious one. 40:23.000 --> 40:25.200 Then there are things humans cannot do perfectly, 40:25.200 --> 40:28.120 even professionals, which is they might miss a detail 40:28.120 --> 40:30.640 or they might have not seen a unit. 40:30.640 --> 40:32.320 And obviously, as a computer, 40:32.320 --> 40:35.040 if there's a corner of the screen that turns green 40:35.040 --> 40:37.720 because a unit enters the field of view, 40:37.720 --> 40:41.120 that can go into the memory of the agent, the LSTM, 40:41.120 --> 40:42.560 and persists there for a while, 40:42.560 --> 40:45.720 and for however long is relevant, right? 40:45.720 --> 40:49.920 And in terms of action, it seems like the rate of action 40:49.920 --> 40:51.640 from Alpha Star is comparative, 40:51.640 --> 40:54.280 if not slower than professional players, 40:54.280 --> 40:57.160 but it's more precise is what I heard. 40:57.160 --> 40:59.760 So that's really probably the one 40:59.760 --> 41:04.760 that is causing us more issues for a couple of reasons, right? 41:05.040 --> 41:08.440 The first one is StarCraft has been an AI environment 41:08.440 --> 41:09.960 for quite a few years. 41:09.960 --> 41:14.000 In fact, I was participating in the very first competition 41:14.000 --> 41:15.920 back in 2010. 41:15.920 --> 41:19.920 And there's really not been a kind of a very clear set 41:19.920 --> 41:22.320 of rules, how the actions per minute, 41:22.320 --> 41:24.720 the rate of actions that you can issue is. 41:24.720 --> 41:29.280 And as a result, these agents or bots that people build 41:29.280 --> 41:31.080 in a kind of almost very cool way, 41:31.080 --> 41:35.400 they do like 20,000, 40,000 actions per minute. 41:35.400 --> 41:37.200 Now, to put this in perspective, 41:37.200 --> 41:41.640 a very good professional human might do 300 41:41.640 --> 41:45.480 to 800 actions per minute, they might not be as precise. 41:45.480 --> 41:49.040 That's why the range is a bit tricky to identify exactly. 41:49.040 --> 41:51.640 I mean, 300 actions per minute precisely 41:51.640 --> 41:54.600 is probably realistic, 800 is probably not, 41:54.600 --> 41:57.000 but you see humans doing a lot of actions 41:57.000 --> 41:59.480 because they warm up and they kind of select things 41:59.480 --> 42:02.240 and spam and so on, just so that when they need, 42:02.240 --> 42:04.240 they have the accuracy. 42:04.240 --> 42:09.240 So we came into this by not having kind of a standard way 42:09.240 --> 42:12.480 to say, well, how do we measure whether an agent 42:12.480 --> 42:14.360 is at human level or not? 42:15.400 --> 42:17.760 On the other hand, we had a huge advantage, 42:17.760 --> 42:20.960 which is because we do imitation learning, 42:20.960 --> 42:24.040 agents turned out to act like humans 42:24.040 --> 42:26.520 in terms of rate of actions, even precision 42:26.520 --> 42:28.320 and in precision of actions. 42:28.320 --> 42:30.600 In the supervised policy, you could see all these. 42:30.600 --> 42:34.280 You could see how agents like to spam click to move here. 42:34.280 --> 42:36.440 If you played, especially Diablo, you would know what I mean. 42:36.440 --> 42:39.840 I mean, you just like spam, oh, move here, move here, move here. 42:39.840 --> 42:43.680 You're doing literally like maybe five actions in two seconds, 42:43.680 --> 42:46.360 but these actions are not very meaningful. 42:46.360 --> 42:48.280 One would have sufficed. 42:48.280 --> 42:51.640 So on the one hand, we start from this imitation policy 42:51.640 --> 42:55.160 that is at the ballpark of the actions per minute of humans 42:55.160 --> 42:58.440 because it's actually statistically trying to imitate humans. 42:58.440 --> 43:00.560 So we see this very nicely in the curves 43:00.560 --> 43:01.840 that we showed in the blog post. 43:01.840 --> 43:04.120 Like there's these actions per minute 43:04.120 --> 43:07.240 and the distribution looks very human like. 43:07.240 --> 43:10.560 But then of course, as self play kicks in, 43:10.560 --> 43:12.800 and that's the part we haven't talked too much yet, 43:12.800 --> 43:16.760 but of course the agent must play against himself to improve, 43:16.760 --> 43:19.200 then there's almost no guarantees 43:19.200 --> 43:22.000 that these actions will not become more precise 43:22.000 --> 43:25.600 or even the rate of actions is gonna increase over time. 43:25.600 --> 43:29.440 So what we did, and this is probably kind of the first attempt 43:29.440 --> 43:30.760 that we thought was reasonable, 43:30.760 --> 43:33.800 is we looked at the distribution of actions for humans 43:33.800 --> 43:36.440 for certain windows of time. 43:36.440 --> 43:37.800 And just to give a perspective, 43:37.800 --> 43:40.760 because I guess I mentioned that some of these agents 43:40.760 --> 43:42.360 that are programmatic, let's call them, 43:42.360 --> 43:44.640 they do 40,000 actions per minute. 43:44.640 --> 43:47.400 Professionals, as I said, do 300 to 800. 43:47.400 --> 43:49.440 So what we looked is we look at the distribution 43:49.440 --> 43:51.000 over professional gamers 43:51.000 --> 43:54.160 and we took reasonably high actions per minute, 43:54.160 --> 43:57.520 but we kind of identify certain cutoffs 43:57.520 --> 44:00.600 after which even if the agent wanted to act, 44:00.600 --> 44:02.000 these actions would be dropped. 44:02.000 --> 44:06.440 But the problem is this cutoff is probably set a bit too high 44:06.440 --> 44:09.360 and what ends up happening, even though the games, 44:09.360 --> 44:11.400 and when we ask the professionals and the gamers, 44:11.400 --> 44:15.040 by and large, they feel like it's playing human like. 44:15.040 --> 44:17.240 There are some agents that developed 44:18.440 --> 44:21.320 maybe slightly too high APMs, 44:22.040 --> 44:23.520 which is actions per minute, 44:23.520 --> 44:25.680 combined with the precision, 44:25.680 --> 44:28.720 which made people sort of start discussing 44:28.720 --> 44:29.680 a very interesting issue, 44:29.680 --> 44:31.200 which is should we have limited 44:31.200 --> 44:34.000 this, should we just let it loose 44:34.000 --> 44:37.520 and see what cool things it can come up with, right? 44:37.520 --> 44:38.360 Interesting. 44:38.360 --> 44:41.960 So this is in itself an extremely interesting question, 44:41.960 --> 44:43.920 but the same way that modeling the shimmer 44:43.920 --> 44:45.440 would be so difficult, 44:45.440 --> 44:48.800 modeling absolutely all the details about muscles 44:48.800 --> 44:51.600 and precision and tiredness of humans 44:51.600 --> 44:52.880 would be quite difficult, right? 44:52.880 --> 44:56.760 So we're really here in kind of innovating in this sense 44:56.760 --> 45:00.040 of, okay, what could be maybe the next iteration 45:00.040 --> 45:01.720 of putting more rules 45:01.720 --> 45:05.040 that makes the agents more human like 45:05.040 --> 45:06.240 in terms of restrictions. 45:06.240 --> 45:08.040 So yeah, putting constraints that. 45:08.040 --> 45:09.240 More constraints, yeah. 45:09.240 --> 45:11.040 That's really interesting, that's really innovative. 45:11.040 --> 45:15.400 So one of the constraints you put on yourself 45:15.400 --> 45:18.000 or at least focused in is on the Protoss race, 45:18.000 --> 45:19.880 as far as I understand. 45:19.880 --> 45:21.880 Can you tell me about the different races 45:21.880 --> 45:26.000 and how they, so Protoss, Terran and Zerg, 45:26.000 --> 45:27.040 how do they compare? 45:27.040 --> 45:28.120 How do they interact? 45:28.120 --> 45:29.960 Why did you choose Protoss? 45:29.960 --> 45:33.600 In the dynamics of the game 45:33.600 --> 45:35.680 seen from a strategic perspective. 45:35.680 --> 45:39.680 So Protoss, so in Starcraft, there are three races. 45:39.680 --> 45:41.320 Indeed, in the demonstration, 45:41.320 --> 45:43.880 we saw only the Protoss race. 45:43.880 --> 45:45.560 So maybe let's start with that one. 45:45.560 --> 45:49.440 Protoss is kind of the most technologically advanced race. 45:49.440 --> 45:53.800 It has units that are expensive, but powerful, right? 45:53.800 --> 45:57.840 So in general, you wanna kind of conserve your units 45:57.840 --> 45:58.640 as you go attack. 45:58.640 --> 46:01.840 So you wanna, and then you wanna utilize 46:01.840 --> 46:05.120 these tactical advantages of very fancy spells 46:05.120 --> 46:07.240 and so on, so forth. 46:07.240 --> 46:10.320 And at the same time, they're kind of, 46:11.280 --> 46:15.040 people say like they're a bit easier to play perhaps, right? 46:15.040 --> 46:17.160 But that I actually didn't know. 46:17.160 --> 46:20.120 I mean, I just talked to now a lot to the players 46:20.120 --> 46:22.520 that we work with, TLO and Mana. 46:22.520 --> 46:24.120 And they said, oh yeah, Protoss is actually, 46:24.120 --> 46:26.320 people think is actually one of the easiest races. 46:26.320 --> 46:29.360 So perhaps the easier, that doesn't mean 46:29.360 --> 46:32.760 that it's obviously professional players 46:32.760 --> 46:34.080 excel at the three races. 46:34.080 --> 46:37.600 And there's never like a race that dominates 46:37.600 --> 46:38.800 for a very long time anyway. 46:38.800 --> 46:41.760 So if you look at the top, I don't know, 100 in the world, 46:41.760 --> 46:44.360 is there one race that dominates that list? 46:44.360 --> 46:46.840 It would be hard to know because it depends on the regions. 46:46.840 --> 46:50.600 I think it's pretty equal in terms of distribution. 46:50.600 --> 46:52.840 And Blizzard wants it to be equal, right? 46:52.840 --> 46:56.320 They wouldn't want one race like Protoss 46:56.320 --> 46:59.960 to not be representative in the top place. 46:59.960 --> 47:03.880 So definitely like they tried it to be like the balance, right? 47:03.880 --> 47:07.320 So then maybe the opposite race of Protoss is Zerg. 47:07.320 --> 47:10.600 Zerg is a race where you just kind of expand 47:10.600 --> 47:13.840 and take over as many resources as you can. 47:13.840 --> 47:15.720 And they have a very high capacity 47:15.720 --> 47:17.680 to regenerate their units. 47:17.680 --> 47:20.520 So if you have an army, it's not that valuable 47:20.520 --> 47:23.960 in terms of losing the whole army is not a big deal as Zerg 47:23.960 --> 47:25.960 because you can then rebuild it. 47:25.960 --> 47:28.320 And given that you generally accumulate 47:28.320 --> 47:32.000 a huge bank of resources, Zerg's typically play 47:32.000 --> 47:34.240 by applying a lot of pressure, 47:34.240 --> 47:36.160 maybe losing their whole army, 47:36.160 --> 47:37.880 but then rebuilding it quickly. 47:37.880 --> 47:40.480 So although of course every race, 47:40.480 --> 47:43.960 I mean, there's never, I mean, they're pretty diverse. 47:43.960 --> 47:45.160 I mean, there are some units in Zerg 47:45.160 --> 47:46.600 that are technologically advanced 47:46.600 --> 47:48.880 and they do some very interesting spells. 47:48.880 --> 47:51.360 And there's some units in Protoss that are less valuable 47:51.360 --> 47:53.360 and you could lose a lot of them and rebuild them 47:53.360 --> 47:55.160 and it wouldn't be a big deal. 47:55.160 --> 47:57.840 All right, so maybe I'm missing out. 47:57.840 --> 47:59.280 Maybe I'm gonna say some dumb stuff, 47:59.280 --> 48:02.480 but just summary of strategy. 48:02.480 --> 48:05.720 So first there's collection of a lot of resources. 48:05.720 --> 48:06.560 That's one option. 48:06.560 --> 48:11.560 The other one is expanding, so building other bases. 48:11.920 --> 48:15.640 Then the other is obviously building units 48:15.640 --> 48:17.200 and attacking with those units. 48:17.200 --> 48:20.640 And then I don't know what else there is. 48:20.640 --> 48:24.080 Maybe there's the different timing of attacks. 48:24.080 --> 48:26.000 Like do attack early, attack late. 48:26.000 --> 48:28.000 What are the different strategies that emerged 48:28.000 --> 48:29.120 that you've learned about? 48:29.120 --> 48:31.360 I've read that a bunch of people are super happy 48:31.360 --> 48:33.000 that you guys have apparently, 48:33.000 --> 48:35.000 that Alpha Star apparently has discovered 48:35.000 --> 48:38.000 that it's really good to, what is it, saturate. 48:38.000 --> 48:39.600 Oh yeah, the mineral line. 48:39.600 --> 48:41.360 Yeah, the mineral line. 48:41.360 --> 48:42.200 Yeah, yeah. 48:42.200 --> 48:45.600 And that's for greedy amateur players like myself. 48:45.600 --> 48:47.480 That's always been a good strategy. 48:47.480 --> 48:49.000 You just build up a lot of money 48:49.000 --> 48:53.280 and it just feels good to just accumulate and accumulate. 48:53.280 --> 48:56.720 So thank you for discovering that and validating all of us. 48:56.720 --> 48:59.200 But is there other strategies that you discovered 48:59.200 --> 49:01.880 interesting and unique to this game? 49:01.880 --> 49:05.080 Yeah, so if you look at the kind of, 49:05.080 --> 49:06.480 not being a Starcraft 2 player, 49:06.480 --> 49:08.120 but of course Starcraft and Starcraft 2 49:08.120 --> 49:11.120 and real time strategy games in general are very similar. 49:11.120 --> 49:16.120 I would classify perhaps the openings of the game. 49:17.560 --> 49:18.760 They're very important. 49:18.760 --> 49:21.760 And generally I would say there's two kinds of openings. 49:21.760 --> 49:23.400 One that's a standard opening, 49:23.400 --> 49:28.400 that's generally how players find sort of a balance 49:28.840 --> 49:31.520 between risk and economy 49:31.520 --> 49:33.400 and building some units early on 49:33.400 --> 49:34.600 so that they could defend, 49:34.600 --> 49:36.800 but they're not too exposed basically, 49:36.800 --> 49:39.480 but also expanding quite quickly. 49:39.480 --> 49:42.040 So this would be kind of a standard opening. 49:42.040 --> 49:43.680 And within a standard opening, 49:43.680 --> 49:45.480 then what you do choose generally 49:45.480 --> 49:48.400 is what technology are you aiming towards? 49:48.400 --> 49:50.280 So there's a bit of rock, paper, scissors 49:50.280 --> 49:52.920 of you could go for spaceships 49:52.920 --> 49:55.080 or you could go for invisible units 49:55.080 --> 49:56.400 or you could go for, I don't know, 49:56.400 --> 50:00.080 like massive units that attack against certain kinds of units 50:00.080 --> 50:01.640 but they're weak against others. 50:01.640 --> 50:05.760 So standard openings themselves have some choices 50:05.760 --> 50:07.480 like rock, paper, scissors style. 50:07.480 --> 50:09.640 Of course, if you scout and you're good at guessing 50:09.640 --> 50:11.080 what the opponent is doing, 50:11.080 --> 50:12.800 then you can play as an advantage 50:12.800 --> 50:14.480 because if you know you're gonna play rock, 50:14.480 --> 50:16.480 I mean, I'm gonna play paper obviously. 50:16.480 --> 50:19.120 So you can imagine that normal standard games 50:19.120 --> 50:22.920 in Starcraft looks like a continuous rock, paper, 50:22.920 --> 50:26.600 scissors game where you guess what the distribution 50:26.600 --> 50:29.920 of rock, paper, and scissors is from the enemy 50:29.920 --> 50:33.360 and reacting accordingly to try to beat it 50:33.360 --> 50:37.000 or put the paper out before he kind of changes 50:37.000 --> 50:38.880 his mind from rock to scissors 50:38.880 --> 50:40.480 and then you would be in a weak position. 50:40.480 --> 50:42.120 So sorry to pause on that. 50:42.120 --> 50:43.320 I didn't realize this element 50:43.320 --> 50:44.880 because I know it's true with poker. 50:44.880 --> 50:46.440 I looked at Leprata's. 50:47.640 --> 50:52.200 You're also estimating, trying to guess the distribution, 50:52.200 --> 50:54.160 trying to better and better estimate the distribution 50:54.160 --> 50:56.040 what the opponent is likely to be doing. 50:56.040 --> 50:57.440 Yeah, I mean, as a player, 50:57.440 --> 50:59.840 you definitely wanna have a belief state 50:59.840 --> 51:03.000 over what's up on the other side of the map 51:03.000 --> 51:05.600 and when your belief state becomes inaccurate 51:05.600 --> 51:08.040 when you start having serious doubts 51:08.040 --> 51:11.320 whether he's gonna play something that you must know, 51:11.320 --> 51:12.440 that's when you scout. 51:12.440 --> 51:14.560 You wanna then gather information, right? 51:14.560 --> 51:16.440 Is improving the accuracy of the belief 51:16.440 --> 51:19.880 or improving the belief state part of the loss 51:19.880 --> 51:21.040 that you're trying to optimize? 51:21.040 --> 51:22.720 Or is it just an side effect? 51:22.720 --> 51:24.040 It's implicit, but implicit. 51:24.040 --> 51:25.840 You could explicitly model it 51:25.840 --> 51:28.280 and it would be quite good at probably predicting 51:28.280 --> 51:30.360 what's on the other side of the map, 51:30.360 --> 51:32.880 but so far it's all implicit. 51:32.880 --> 51:36.680 There's no additional reward for predicting the enemy. 51:36.680 --> 51:38.800 So there's these standard openings 51:38.800 --> 51:41.640 and then there's what people call cheese, 51:41.640 --> 51:42.800 which is very interesting 51:42.800 --> 51:46.760 and Alpha Star sometimes really likes this kind of cheese. 51:46.760 --> 51:51.120 These cheeses, what they are is kind of an all in strategy. 51:51.120 --> 51:53.240 You're gonna do something sneaky. 51:53.240 --> 51:56.680 You're gonna hide your own buildings 51:56.680 --> 51:58.200 close to the enemy base 51:58.200 --> 52:01.600 or you're gonna go for hiding your technological buildings 52:01.600 --> 52:03.040 so that you do invisible units 52:03.040 --> 52:06.040 and the enemy just cannot react to detect it 52:06.040 --> 52:08.000 and thus lose the game. 52:08.000 --> 52:10.000 And there's quite a few of these cheeses 52:10.000 --> 52:11.800 and variants of them. 52:11.800 --> 52:14.480 And there it's where actually the belief state 52:14.480 --> 52:16.360 becomes even more important 52:16.360 --> 52:18.520 because if I scout your base 52:18.520 --> 52:20.200 and I see no buildings at all, 52:20.200 --> 52:22.480 any human player knows some things up. 52:22.480 --> 52:23.320 They might know, well, 52:23.320 --> 52:25.640 you're hiding something close to my base. 52:25.640 --> 52:28.200 Should I build suddenly a lot of units to defense? 52:28.200 --> 52:31.000 Should I actually block my ramp with workers 52:31.000 --> 52:33.520 so that you cannot come and destroy my base? 52:33.520 --> 52:35.680 So there's all this is happening 52:35.680 --> 52:39.440 and defending against cheeses is extremely important. 52:39.440 --> 52:40.760 And in the Alpha Star League, 52:40.760 --> 52:45.080 many agents actually develop some cheesy strategies. 52:45.080 --> 52:48.000 And in the games we saw against TLO and Mana, 52:48.000 --> 52:49.240 two out of the 10 agents 52:49.240 --> 52:51.760 were actually doing these kind of strategies 52:51.760 --> 52:53.640 which are cheesy strategies. 52:53.640 --> 52:55.600 And then there's a variant of cheesy strategy 52:55.600 --> 52:57.360 which is called all in. 52:57.360 --> 53:00.280 So an all in strategy is not perhaps as drastic 53:00.280 --> 53:02.560 as oh, I'm gonna build cannons on your base 53:02.560 --> 53:03.880 and then bring all my workers 53:03.880 --> 53:06.840 and try to just disrupt your base and game over 53:06.840 --> 53:08.800 or GG as we say in StarCraft. 53:09.840 --> 53:12.000 There's these kind of very cool things 53:12.000 --> 53:14.760 that you can align precisely at a certain time mark. 53:14.760 --> 53:17.400 So for instance, you can generate 53:17.400 --> 53:20.280 exactly 10 unit composition that is perfect. 53:20.280 --> 53:22.960 Like five of this type, five of these other type 53:22.960 --> 53:26.240 and align the upgrade so that at four minutes and a half, 53:26.240 --> 53:28.680 let's say you have these 10 units 53:28.680 --> 53:30.640 and the upgrade just finished. 53:30.640 --> 53:34.000 And at that point, that army is really scary. 53:34.000 --> 53:36.440 And unless the enemy really knows what's going on, 53:36.440 --> 53:40.280 if you push, you might then have an advantage 53:40.280 --> 53:42.480 because maybe the enemy is doing something more standard, 53:42.480 --> 53:45.800 it expanded too much, it developed too much economy 53:45.800 --> 53:49.760 and it trade off badly against having defenses 53:49.760 --> 53:51.120 and the enemy will lose. 53:51.120 --> 53:53.680 But it's called all in because if you don't win, 53:53.680 --> 53:55.080 then you're gonna lose. 53:55.080 --> 53:57.960 So you see players that do these kind of strategies, 53:57.960 --> 54:00.000 if they don't succeed, game is not over. 54:00.000 --> 54:01.240 I mean, they still have a base 54:01.240 --> 54:02.880 and they still gathering minerals, 54:02.880 --> 54:04.800 but they will just GG out of the game 54:04.800 --> 54:06.800 because they know, well, game is over. 54:06.800 --> 54:08.880 I gambled and I failed. 54:08.880 --> 54:11.600 So if we start entering the game 54:11.600 --> 54:14.520 theoretic aspects of the game, it's really rich 54:14.520 --> 54:18.000 and that's why it also makes it quite entertaining to watch. 54:18.000 --> 54:21.800 Even if I don't play, I still enjoy watching the game. 54:21.800 --> 54:26.800 But the agents are trying to do this mostly implicitly, 54:26.800 --> 54:29.320 but one element that we improved in self plays 54:29.320 --> 54:31.320 creating the Alpha Star League. 54:31.320 --> 54:34.600 And the Alpha Star League is not pure self play. 54:34.600 --> 54:37.880 It's trying to create different personalities of agents 54:37.880 --> 54:41.480 so that some of them will become cheesy agents. 54:41.480 --> 54:44.360 Some of them might become very economical, very greedy, 54:44.360 --> 54:46.160 like getting all the resources, 54:46.160 --> 54:48.760 but then maybe early on they're gonna be weak, 54:48.760 --> 54:51.040 but later on they're gonna be very strong. 54:51.040 --> 54:53.400 And by creating this personality of agents, 54:53.400 --> 54:55.400 which sometimes it just happens naturally 54:55.400 --> 54:58.240 that you can see kind of an evolution of agents 54:58.240 --> 55:00.760 that given the previous generation, 55:00.760 --> 55:01.920 they train against all of them 55:01.920 --> 55:04.320 and then they generate kind of the perfect counter 55:04.320 --> 55:05.760 to that distribution. 55:05.760 --> 55:09.280 But these agents, you must have them in the populations 55:09.280 --> 55:11.280 because if you don't have them, 55:11.280 --> 55:13.040 you're not covered against these things, right? 55:13.040 --> 55:17.080 It's kind of, you wanna create all sorts of the opponents 55:17.080 --> 55:18.640 that you will find in the wild. 55:18.640 --> 55:21.800 So you can be exposed to these cheeses, 55:21.800 --> 55:25.720 early aggression, later aggression, more expansions, 55:25.720 --> 55:29.560 dropping units in your base from the side, all these things. 55:29.560 --> 55:32.760 And pure self play is getting a bit stuck 55:32.760 --> 55:36.200 at finding some subset of these, but not all of these. 55:36.200 --> 55:39.480 So the Alpha Star League is a way to kind of 55:39.480 --> 55:41.560 do an ensemble of agents 55:41.560 --> 55:43.480 that they're all playing in a league 55:43.480 --> 55:45.520 much like people play on Battle.net, right? 55:45.520 --> 55:47.440 They play, you play against someone 55:47.440 --> 55:50.240 who does a new cool strategy and you immediately, 55:50.240 --> 55:53.040 oh my God, I wanna try it, I wanna play again. 55:53.040 --> 55:55.960 And these to me was another critical part 55:55.960 --> 55:58.520 of the problem which was, 55:58.520 --> 56:01.240 can we create a Battle.net for agents? 56:01.240 --> 56:02.080 Yeah. 56:02.080 --> 56:03.400 And that's kind of what the Alpha Star League really. 56:03.400 --> 56:04.240 That's fascinating. 56:04.240 --> 56:06.920 And where they stick to their different strategies. 56:06.920 --> 56:09.960 Yeah, wow, that's really, really interesting. 56:09.960 --> 56:13.240 But that said, you were fortunate enough 56:13.240 --> 56:16.280 or just skilled enough to win 5.0. 56:17.320 --> 56:19.280 And so how hard is it to win? 56:19.280 --> 56:20.320 I mean, that's not the goal. 56:20.320 --> 56:21.920 I guess, I don't know what the goal is. 56:21.920 --> 56:25.400 The goal should be to win majority, not 5.0, 56:25.400 --> 56:29.360 but how hard is it in general to win all matchups? 56:29.360 --> 56:31.080 I don't want V1. 56:31.080 --> 56:33.600 So that's a very interesting question 56:33.600 --> 56:37.240 because once you see Alpha Star 56:37.240 --> 56:39.520 and superficially you think, well, okay, 56:39.520 --> 56:42.960 it won, if you sum all the games like 10 to one, right? 56:42.960 --> 56:46.280 It lost the game that it played with the camera interface. 56:46.280 --> 56:48.480 You might think, well, that's done, right? 56:48.480 --> 56:50.840 It's super human at the game. 56:50.840 --> 56:54.800 And that's not really the claim we really can make, actually. 56:56.000 --> 56:58.840 The claim is we beat a professional gamer 56:58.840 --> 57:00.120 for the first time. 57:00.120 --> 57:02.480 Starcraft has really been a thing 57:02.480 --> 57:04.120 that has been going on for a few years, 57:04.120 --> 57:09.120 but a moment like this had not occurred before yet. 57:09.520 --> 57:12.400 But are these agents impossible to beat? 57:12.400 --> 57:13.440 Absolutely not, right? 57:13.440 --> 57:17.360 So that's a bit what's kind of the difference is 57:17.360 --> 57:19.560 the agents play at grandmaster level. 57:19.560 --> 57:21.520 They definitely understand the game enough 57:21.520 --> 57:24.960 to play extremely well, but are they unbeatable? 57:24.960 --> 57:26.600 Do they play perfect? 57:27.920 --> 57:30.320 No, and actually in Starcraft, 57:30.320 --> 57:33.240 because of these sneaky strategies, 57:33.240 --> 57:36.680 it's always possible that you might take a huge risk sometimes, 57:36.680 --> 57:39.200 but you might get wins, right, out of this. 57:39.200 --> 57:44.200 So I think as a domain, it still has a lot of opportunities, 57:44.200 --> 57:47.760 not only because of course we wanna learn with less experience, 57:47.760 --> 57:50.480 we would like to, I mean, if I learn to play Protoss, 57:50.480 --> 57:53.280 I can play Terran and learn it much quicker 57:53.280 --> 57:54.480 than Alpha Star can, right? 57:54.480 --> 57:58.440 So there are obvious interesting research challenges as well. 57:58.440 --> 58:03.080 But even as the raw performance goes, 58:03.080 --> 58:05.960 really the claim here can be we are at pro level 58:05.960 --> 58:09.080 or at high grandmaster level, 58:09.080 --> 58:14.080 but obviously the players also did not know what to expect, 58:14.360 --> 58:17.000 right, their prior distribution was a bit off 58:17.000 --> 58:20.400 because they played this kind of new alien brain 58:20.400 --> 58:22.080 as they like to say it, right? 58:22.080 --> 58:25.080 And that's what makes it exciting for them, 58:25.080 --> 58:28.040 but also I think if you look at the games closely, 58:28.040 --> 58:31.520 you see there were weaknesses in some points, 58:31.520 --> 58:33.320 maybe Alpha Star did not scout 58:33.320 --> 58:36.080 or if it had got invisible units going against 58:36.080 --> 58:38.200 at certain points, it wouldn't have known 58:38.200 --> 58:39.600 and it would have been bad. 58:39.600 --> 58:42.920 So there's still quite a lot of work to do, 58:42.920 --> 58:45.440 but it's really a very exciting moment for us 58:45.440 --> 58:49.120 to be seeing, wow, a single neural net on a GPU 58:49.120 --> 58:52.080 is actually playing against these guys who are amazing. 58:52.080 --> 58:53.760 I mean, you have to see them play in life. 58:53.760 --> 58:55.800 They're really, really amazing players. 58:55.800 --> 59:00.440 Yeah, I'm sure there must be a guy in Poland somewhere 59:00.440 --> 59:02.680 right now training his butt off 59:02.680 --> 59:06.600 to make sure that this never happens again with Alpha Star. 59:06.600 --> 59:09.720 So that's really exciting in terms of Alpha Star 59:09.720 --> 59:12.200 having some holes to exploit, which is great. 59:12.200 --> 59:14.320 And then you build on top of each other 59:14.320 --> 59:18.920 and it feels like Starcraft on let go, even if you win, 59:18.920 --> 59:21.640 it's still not, it's still not, 59:21.640 --> 59:23.120 there's so many different dimensions 59:23.120 --> 59:24.200 in which you can explore. 59:24.200 --> 59:25.560 So that's really, really interesting. 59:25.560 --> 59:28.520 Do you think there's a ceiling to Alpha Star? 59:28.520 --> 59:32.840 You've said that it hasn't reached, this is a big, 59:33.960 --> 59:35.520 let me actually just pause for a second. 59:35.520 --> 59:40.200 How did it feel to come here to this point, 59:40.200 --> 59:42.240 to beat a top professional player? 59:42.240 --> 59:44.600 Like that night, I mean, you know, 59:44.600 --> 59:47.160 Olympic athletes have their gold medal, right? 59:47.160 --> 59:48.840 This is your gold medal in a sense. 59:48.840 --> 59:50.400 Sure, you're cited a lot, 59:50.400 --> 59:53.120 you've published a lot of prestigious papers, whatever, 59:53.120 --> 59:55.280 but this is like a win. 59:55.280 --> 59:56.480 How did it feel? 59:56.480 --> 59:59.440 I mean, it was, for me, it was unbelievable 59:59.440 --> 1:00:04.440 because first the win itself, I mean, it was so exciting. 1:00:04.440 --> 1:00:09.440 I mean, so looking back to those last days of 2018, 1:00:09.840 --> 1:00:12.040 really, that's when the games were played, 1:00:12.040 --> 1:00:14.560 I'm sure I'll look back at that moment and say, 1:00:14.560 --> 1:00:17.280 oh my God, I wanna be in a project like that. 1:00:17.280 --> 1:00:20.440 It's like, I already feel the nostalgia of like, 1:00:20.440 --> 1:00:23.560 yeah, that was huge in terms of the energy 1:00:23.560 --> 1:00:25.720 and the team effort that went into it. 1:00:25.720 --> 1:00:28.520 And so in that sense, as soon as it happened, 1:00:28.520 --> 1:00:30.640 I already knew it was kind of, 1:00:30.640 --> 1:00:32.320 I was losing it a little bit. 1:00:32.320 --> 1:00:35.440 So it is almost like sad that it happened and oh my God, 1:00:35.440 --> 1:00:40.440 like, but on the other hand, it also verifies the approach. 1:00:40.600 --> 1:00:43.160 But to me also, there's so many challenges 1:00:43.160 --> 1:00:45.400 and interesting aspects of intelligence 1:00:45.400 --> 1:00:49.240 that even though we can train a neural network 1:00:49.240 --> 1:00:52.040 to play at the level of the best humans, 1:00:52.040 --> 1:00:53.600 there's still so many challenges. 1:00:53.600 --> 1:00:54.800 So for me, it's also like, 1:00:54.800 --> 1:00:56.800 well, this is really an amazing achievement, 1:00:56.800 --> 1:00:59.280 but I already was also thinking about next steps. 1:00:59.280 --> 1:01:01.720 I mean, as I said, these Asians play Protos, 1:01:01.720 --> 1:01:04.080 they play Protos versus Protos, 1:01:04.080 --> 1:01:07.240 but they should be able to play a different race 1:01:07.240 --> 1:01:08.160 much quicker, right? 1:01:08.160 --> 1:01:10.640 So that would be an amazing achievement. 1:01:10.640 --> 1:01:13.360 Some people call this meta reinforcement learning, 1:01:13.360 --> 1:01:15.200 meta learning and so on, right? 1:01:15.200 --> 1:01:18.960 So there's so many possibilities after that moment, 1:01:18.960 --> 1:01:22.240 but the moment itself, it really felt great. 1:01:22.240 --> 1:01:24.520 It's, we had this bet. 1:01:24.520 --> 1:01:27.720 So I'm kind of a pessimist in general. 1:01:27.720 --> 1:01:30.120 So I kind of send an email to the team and I said, 1:01:30.120 --> 1:01:33.680 okay, let's, against TLO first, right? 1:01:33.680 --> 1:01:35.120 Like what's going to be the result? 1:01:35.120 --> 1:01:38.680 And I really thought we would lose like five zero, right? 1:01:38.680 --> 1:01:41.480 I, I, we had some calibration made 1:01:41.480 --> 1:01:44.080 against the 5,000 MMR player. 1:01:44.080 --> 1:01:47.360 TLO was much stronger than that player. 1:01:47.360 --> 1:01:50.040 Even if he played Protos, which is his off race, 1:01:51.040 --> 1:01:53.120 but yeah, it was not imagining we would win. 1:01:53.120 --> 1:01:55.600 So for me, that was just kind of a test run or something. 1:01:55.600 --> 1:01:59.000 And then it really kind of, he was really surprised. 1:01:59.000 --> 1:02:02.360 And unbelievably, we went to this, 1:02:02.360 --> 1:02:04.560 to this bar to celebrate. 1:02:04.560 --> 1:02:08.360 And Dave tells me, well, why don't we invite someone 1:02:08.360 --> 1:02:11.000 who is a thousand MMR stronger in Protos? 1:02:11.000 --> 1:02:13.040 Like an actual Protos player, like, 1:02:13.040 --> 1:02:16.200 like that it turned out being mana, right? 1:02:16.200 --> 1:02:19.400 And, you know, we had some drinks and I said, sure, why not? 1:02:19.400 --> 1:02:20.240 But then I thought, well, 1:02:20.240 --> 1:02:22.080 that's really going to be impossible to beat. 1:02:22.080 --> 1:02:24.600 I mean, even because it's so much ahead. 1:02:24.600 --> 1:02:28.440 A thousand MMR is really like 99% probability 1:02:28.440 --> 1:02:33.080 that mana would beat TLO as Protos versus Protos, right? 1:02:33.080 --> 1:02:34.240 So we did that. 1:02:34.240 --> 1:02:39.000 And to me, the second game was much more important, 1:02:39.000 --> 1:02:42.120 even though a lot of uncertainty kind of disappeared 1:02:42.120 --> 1:02:43.680 after we kind of beat TLO. 1:02:43.680 --> 1:02:45.680 I mean, he is a professional player. 1:02:45.680 --> 1:02:48.000 So that was kind of, oh, but that's really 1:02:48.000 --> 1:02:49.760 a very nice achievement. 1:02:49.760 --> 1:02:51.800 But mana really was at the top. 1:02:51.800 --> 1:02:53.880 And you could see he played much better, 1:02:53.880 --> 1:02:55.400 but our agents got much better too. 1:02:55.400 --> 1:02:57.440 So it's like, ah. 1:02:57.440 --> 1:02:59.800 And then after the first game, I said, 1:02:59.800 --> 1:03:02.760 if we take a single game, at least we can say we beat A game. 1:03:02.760 --> 1:03:04.320 I mean, even if we don't beat the series, 1:03:04.320 --> 1:03:06.960 for me, that was a huge relief. 1:03:06.960 --> 1:03:09.200 And I mean, I remember the hacking them is. 1:03:09.200 --> 1:03:11.600 And I mean, it was really like this moment, 1:03:11.600 --> 1:03:14.200 for me, will resonate forever as a researcher. 1:03:14.200 --> 1:03:15.880 And I mean, as a person, and yeah, 1:03:15.880 --> 1:03:18.280 it's a really like great accomplishment. 1:03:18.280 --> 1:03:21.360 And it was great also to be there with the team in the room. 1:03:21.360 --> 1:03:23.080 I don't know if you saw like this. 1:03:23.080 --> 1:03:24.760 So it was really like. 1:03:24.760 --> 1:03:26.000 I mean, from my perspective, 1:03:26.000 --> 1:03:29.720 the other interesting thing is just like watching Kasparov, 1:03:29.720 --> 1:03:33.760 now watching mana was also interesting 1:03:33.760 --> 1:03:36.160 because he is kind of at a loss of words. 1:03:36.160 --> 1:03:38.400 I mean, whenever you lose, I've done a lot of sports. 1:03:38.400 --> 1:03:43.560 You sometimes say excuses, you look for reasons. 1:03:43.560 --> 1:03:46.280 And he couldn't really come up with reasons. 1:03:46.280 --> 1:03:50.040 I mean, so with the off race for Protoss, 1:03:50.040 --> 1:03:52.320 you could say, well, it felt awkward, it wasn't, 1:03:52.320 --> 1:03:55.200 but here it was just beaten. 1:03:55.200 --> 1:03:57.960 And it was beautiful to look at a human being 1:03:57.960 --> 1:04:00.320 being superseded by an AI system. 1:04:00.320 --> 1:04:04.480 I mean, it's a beautiful moment for researchers. 1:04:04.480 --> 1:04:05.920 Yeah, for sure it was. 1:04:05.920 --> 1:04:09.960 I mean, probably the highlight of my career so far 1:04:09.960 --> 1:04:11.760 because of its uniqueness and coolness. 1:04:11.760 --> 1:04:14.280 And I don't know, I mean, it's obviously, as you said, 1:04:14.280 --> 1:04:16.240 you can look at paper citations and so on. 1:04:16.240 --> 1:04:19.280 But this really is like a testament 1:04:19.280 --> 1:04:22.400 of the whole machine learning approach 1:04:22.400 --> 1:04:24.640 and using games to advance technology. 1:04:24.640 --> 1:04:27.880 I mean, it really was, everything came together 1:04:27.880 --> 1:04:29.840 at that moment, that's really the summary. 1:04:29.840 --> 1:04:34.040 Also, on the other side, it's a popularization of AI too 1:04:34.040 --> 1:04:38.200 because it's just like traveling to the moon and so on. 1:04:38.200 --> 1:04:41.000 I mean, this is where a very large community of people 1:04:41.000 --> 1:04:45.080 that don't really know AI, they get to really interact with it. 1:04:45.080 --> 1:04:46.000 Which is very important. 1:04:46.000 --> 1:04:50.800 I mean, we must, you know, writing papers helps our peers, 1:04:50.800 --> 1:04:52.520 researchers to understand what we're doing. 1:04:52.520 --> 1:04:55.880 But I think AI is becoming mature enough 1:04:55.880 --> 1:04:59.000 that we must sort of try to explain what it is. 1:04:59.000 --> 1:05:01.440 And perhaps through games is an obvious way 1:05:01.440 --> 1:05:03.640 because these games always had built in AI. 1:05:03.640 --> 1:05:07.680 So it may be everyone experienced an AI playing a video game 1:05:07.680 --> 1:05:08.520 even if they don't know. 1:05:08.520 --> 1:05:10.240 Because there's always some scripted element 1:05:10.240 --> 1:05:13.880 and some people might even call that AI already, right? 1:05:13.880 --> 1:05:16.320 So what are other applications 1:05:16.320 --> 1:05:20.280 of the approaches underlying Alpha Star that you see happening? 1:05:20.280 --> 1:05:23.120 There's a lot of echoes of, you said, transformer 1:05:23.120 --> 1:05:25.680 of language modeling and so on. 1:05:25.680 --> 1:05:29.480 Have you already started thinking where the breakthroughs 1:05:29.480 --> 1:05:32.280 in Alpha Star get expanded to other applications? 1:05:32.280 --> 1:05:34.640 Right, so I thought about a few things 1:05:34.640 --> 1:05:37.280 for like kind of next months, next years. 1:05:38.440 --> 1:05:40.520 The main thing I'm thinking about actually is 1:05:40.520 --> 1:05:43.160 what's next as a kind of a grand challenge 1:05:43.160 --> 1:05:47.120 because for me, like we've seen Atari 1:05:47.120 --> 1:05:50.280 and then there's like the sort of three dimensional walls 1:05:50.280 --> 1:05:52.520 that we've seen also like pretty good performance 1:05:52.520 --> 1:05:54.160 from this capture the flag agents 1:05:54.160 --> 1:05:57.600 that also some people at DeepMind and elsewhere are working on. 1:05:57.600 --> 1:05:59.600 We've also seen some amazing results on like, 1:05:59.600 --> 1:06:03.280 for instance, Dota 2, which is also a very complicated game. 1:06:03.280 --> 1:06:05.960 So for me, like the main thing I'm thinking about 1:06:05.960 --> 1:06:07.960 is what's next in terms of challenge. 1:06:07.960 --> 1:06:12.960 So as a researcher, I see sort of two tensions 1:06:12.960 --> 1:06:16.760 between research and then applications or areas 1:06:16.760 --> 1:06:18.480 or domains where you apply them. 1:06:18.480 --> 1:06:20.480 So on the one hand, we've done, 1:06:20.480 --> 1:06:23.320 thanks to the application of StarCraft is very hard. 1:06:23.320 --> 1:06:25.600 We developed some techniques, some new research 1:06:25.600 --> 1:06:27.480 that now we could look at elsewhere, 1:06:27.480 --> 1:06:30.520 like are there other applications where we can apply this? 1:06:30.520 --> 1:06:32.880 And the obvious ones, absolutely, 1:06:32.880 --> 1:06:37.480 you can think of feeding back to sort of the community 1:06:37.480 --> 1:06:40.240 we took from, which was mostly sequence modeling 1:06:40.240 --> 1:06:41.680 or natural language processing. 1:06:41.680 --> 1:06:46.120 So we've developed an extended things from the transformer 1:06:46.120 --> 1:06:48.120 and we use pointer networks. 1:06:48.120 --> 1:06:51.280 We combine LSTM and transformers in interesting ways. 1:06:51.280 --> 1:06:54.200 So that's perhaps the kind of lowest hanging fruit 1:06:54.200 --> 1:06:58.840 of feeding back to now a different field of machine learning 1:06:58.840 --> 1:07:00.880 that's not playing video games. 1:07:00.880 --> 1:07:05.680 Let me go old school and jump to Mr. Alan Turing. 1:07:05.680 --> 1:07:09.880 So the Turing test is a natural language test, 1:07:09.880 --> 1:07:11.560 a conversational test. 1:07:11.560 --> 1:07:15.760 What's your thought of it as a test for intelligence? 1:07:15.760 --> 1:07:17.360 Do you think it is a grand challenge 1:07:17.360 --> 1:07:18.920 that's worthy of undertaking? 1:07:18.920 --> 1:07:21.960 Maybe if it is, would you reformulate it 1:07:21.960 --> 1:07:23.720 or phrase it somehow differently? 1:07:23.720 --> 1:07:25.640 Right, so I really love the Turing test 1:07:25.640 --> 1:07:29.600 because I also like sequences and language understanding. 1:07:29.600 --> 1:07:32.160 And in fact, some of the early work 1:07:32.160 --> 1:07:33.520 we did in machine translation, 1:07:33.520 --> 1:07:37.320 we tried to apply to kind of a neural chat bot, 1:07:37.320 --> 1:07:40.200 which obviously would never pass the Turing test 1:07:40.200 --> 1:07:42.320 because it was very limited. 1:07:42.320 --> 1:07:45.200 But it is a very fascinating idea 1:07:45.200 --> 1:07:49.440 that you could really have an AI 1:07:49.440 --> 1:07:51.800 that would be indistinguishable from humans 1:07:51.800 --> 1:07:56.040 in terms of asking or conversing with it, right? 1:07:56.040 --> 1:08:00.720 So I think the test itself seems very nice 1:08:00.720 --> 1:08:02.600 and it's kind of well defined actually, 1:08:02.600 --> 1:08:05.000 like the passing it or not. 1:08:05.000 --> 1:08:06.560 I think there's quite a few rules 1:08:06.560 --> 1:08:11.560 that feel like pretty simple and you could really have, 1:08:12.520 --> 1:08:14.800 I mean, I think they have these competitions every year. 1:08:14.800 --> 1:08:15.920 Yeah, so the Leibniz Prize, 1:08:15.920 --> 1:08:20.920 but I don't know if you've seen the kind of bots 1:08:22.240 --> 1:08:24.160 that emerge from that competition. 1:08:24.160 --> 1:08:28.000 They're not quite as what you would, 1:08:28.000 --> 1:08:29.920 so it feels like that there's weaknesses 1:08:29.920 --> 1:08:31.400 with the way Turing formulated it. 1:08:31.400 --> 1:08:35.000 It needs to be that the definition 1:08:35.000 --> 1:08:40.000 of a genuine, rich, fulfilling human conversation 1:08:40.000 --> 1:08:41.640 it needs to be something else. 1:08:41.640 --> 1:08:43.000 Like the Alexa Prize, 1:08:43.000 --> 1:08:44.880 which I'm not as well familiar with, 1:08:44.880 --> 1:08:46.200 has tried to define that more. 1:08:46.200 --> 1:08:48.240 I think by saying you have to continue 1:08:48.240 --> 1:08:50.680 keeping a conversation for 30 minutes, 1:08:50.680 --> 1:08:52.240 something like that. 1:08:52.240 --> 1:08:55.520 So basically forcing the agent not to just fool, 1:08:55.520 --> 1:08:58.000 but to have an engaging conversation kind of thing, 1:08:58.000 --> 1:09:03.000 is that, I mean, have you thought 1:09:03.720 --> 1:09:06.400 about this problem richly? 1:09:06.400 --> 1:09:10.680 And if you have in general, how far away are we from, 1:09:10.680 --> 1:09:14.160 you worked a lot on language understanding, 1:09:14.160 --> 1:09:16.640 language generation, but the full dialogue, 1:09:16.640 --> 1:09:19.920 the conversation, just sitting at the bar, 1:09:19.920 --> 1:09:21.760 having a cup of beers for an hour, 1:09:21.760 --> 1:09:22.960 that kind of conversation. 1:09:22.960 --> 1:09:23.800 Have you thought about it? 1:09:23.800 --> 1:09:26.440 Yeah, so I think you touched here on the critical point, 1:09:26.440 --> 1:09:28.640 which is feasibility, right? 1:09:28.640 --> 1:09:32.880 So there's a great sort of essay by Hamming, 1:09:32.880 --> 1:09:37.400 which describes sort of grand challenges of physics. 1:09:37.400 --> 1:09:41.080 And he argues that, well, okay, for instance, 1:09:41.080 --> 1:09:44.720 teleportation or time travel are great grand challenges 1:09:44.720 --> 1:09:46.600 of physics, but there's no attacks. 1:09:46.600 --> 1:09:50.360 We really don't know or cannot kind of make any progress. 1:09:50.360 --> 1:09:53.360 So that's why most physicists and so on, 1:09:53.360 --> 1:09:55.360 they don't work on these in their PhDs 1:09:55.360 --> 1:09:57.920 and as part of their careers. 1:09:57.920 --> 1:10:01.000 So I see the Turing test as, in the full Turing test, 1:10:01.000 --> 1:10:02.760 as a bit still too early. 1:10:02.760 --> 1:10:06.760 Like I am, I think we're, especially with the current trend 1:10:06.760 --> 1:10:10.080 of deep learning language models, 1:10:10.080 --> 1:10:11.640 we've seen some amazing examples, 1:10:11.640 --> 1:10:14.160 I think GPT2 being the most recent one, 1:10:14.160 --> 1:10:15.840 which is very impressive, 1:10:15.840 --> 1:10:20.840 but to understand to fully solve passing or fooling a human 1:10:21.080 --> 1:10:23.480 to think that there's a human on the other side, 1:10:23.480 --> 1:10:24.960 I think we're quite far. 1:10:24.960 --> 1:10:27.360 So as a result, I don't see myself 1:10:27.360 --> 1:10:30.520 and I probably would not recommend people doing a PhD 1:10:30.520 --> 1:10:31.680 on solving the Turing test, 1:10:31.680 --> 1:10:34.120 because it just feels it's kind of too early 1:10:34.120 --> 1:10:35.520 or too hard of a problem. 1:10:35.520 --> 1:10:37.840 Yeah, but that said, you said the exact same thing 1:10:37.840 --> 1:10:40.480 about StarCraft about a few years ago. 1:10:40.480 --> 1:10:42.600 So to demo, so I pre... 1:10:42.600 --> 1:10:43.920 Yes. 1:10:43.920 --> 1:10:45.600 You'll probably also be the person 1:10:45.600 --> 1:10:48.240 who passes the Turing test in three years. 1:10:48.240 --> 1:10:51.040 I mean, I think the, yeah, so... 1:10:51.040 --> 1:10:52.720 So we have this on record, this is nice. 1:10:52.720 --> 1:10:53.560 It's true. 1:10:53.560 --> 1:10:56.600 I mean, it's true that progress sometimes 1:10:56.600 --> 1:10:57.840 is a bit unpredictable. 1:10:57.840 --> 1:11:00.840 I really wouldn't have not, even six months ago, 1:11:00.840 --> 1:11:02.520 I would not have predicted the level 1:11:02.520 --> 1:11:05.480 that we see that these agents can deliver. 1:11:05.480 --> 1:11:10.120 At grandmaster level, but I have worked on language enough. 1:11:10.120 --> 1:11:13.640 And basically my concern is not that something could happen, 1:11:13.640 --> 1:11:16.440 a breakthrough could happen that would bring us to solving 1:11:16.440 --> 1:11:18.440 or passing the Turing test, 1:11:18.440 --> 1:11:21.680 is that I just think the statistical approach to it, 1:11:21.680 --> 1:11:24.160 like this is not gonna cut it. 1:11:24.160 --> 1:11:25.960 So we need a breakthrough, 1:11:25.960 --> 1:11:28.320 which is great for the community. 1:11:28.320 --> 1:11:31.840 But given that, I think there's quite a more uncertainty. 1:11:31.840 --> 1:11:34.280 Whereas for StarCraft, 1:11:34.280 --> 1:11:38.160 I knew what the steps would be to kind of get us there. 1:11:38.160 --> 1:11:41.640 I think it was clear that using the imitation learning part 1:11:41.640 --> 1:11:44.360 and then using these battle network agents 1:11:44.360 --> 1:11:48.320 were gonna be key and it turned out that this was the case 1:11:48.320 --> 1:11:51.640 and a little more was needed, but not much more. 1:11:51.640 --> 1:11:54.360 For Turing test, I just don't know what the plan 1:11:54.360 --> 1:11:56.000 or execution plan would look like. 1:11:56.000 --> 1:11:59.160 So that's why I myself working on it 1:11:59.160 --> 1:12:01.520 as a grand challenge is hard, 1:12:01.520 --> 1:12:03.920 but there are quite a few sub challenges 1:12:03.920 --> 1:12:05.480 that are related that you could say, 1:12:05.480 --> 1:12:09.080 well, I mean, what if you create a great assistant, 1:12:09.080 --> 1:12:11.400 like Google already has like the Google Assistant. 1:12:11.400 --> 1:12:13.120 So can we make it better 1:12:13.120 --> 1:12:15.440 and can we make it fully neural and so on? 1:12:15.440 --> 1:12:18.200 That I start to believe maybe we're reaching a point 1:12:18.200 --> 1:12:20.760 where we should attempt these challenges. 1:12:20.760 --> 1:12:22.480 I like this conversation so much 1:12:22.480 --> 1:12:24.920 because it echoes very much the StarCraft conversation. 1:12:24.920 --> 1:12:26.920 It's exactly how you approach StarCraft. 1:12:26.920 --> 1:12:29.680 Let's break it down into small pieces and solve those 1:12:29.680 --> 1:12:31.400 and you end up solving the whole game. 1:12:31.400 --> 1:12:34.120 Great, but that said, you're behind some 1:12:34.120 --> 1:12:37.960 of the sort of biggest pieces of work and deep learning 1:12:37.960 --> 1:12:39.360 in the last several years. 1:12:40.280 --> 1:12:42.320 So you mentioned some limits. 1:12:42.320 --> 1:12:44.960 What do you think are the current limits of deep learning 1:12:44.960 --> 1:12:47.080 and how do we overcome those limits? 1:12:47.080 --> 1:12:50.160 So if I had to actually use a single word 1:12:50.160 --> 1:12:53.200 to define the main challenge in deep learning, 1:12:53.200 --> 1:12:55.720 it's a challenge that probably has been the challenge 1:12:55.720 --> 1:12:59.760 for many years and is that of generalization. 1:12:59.760 --> 1:13:04.520 So what that means is that all that we're doing 1:13:04.520 --> 1:13:06.800 is fitting functions to data. 1:13:06.800 --> 1:13:11.800 And when the data we see is not from the same distribution 1:13:12.160 --> 1:13:14.080 or even if there are some times 1:13:14.080 --> 1:13:16.800 that it is very close to distribution 1:13:16.800 --> 1:13:20.240 but because of the way we train it with limited samples, 1:13:20.240 --> 1:13:23.880 we then get to this stage where we just don't see 1:13:23.880 --> 1:13:27.760 generalization as much as we can generalize. 1:13:27.760 --> 1:13:31.240 And I think adversarial examples are a clear example of this 1:13:31.240 --> 1:13:34.640 but if you study machine learning and literature 1:13:34.640 --> 1:13:38.320 and the reason why SVMs came very popular 1:13:38.320 --> 1:13:39.720 were because they were dealing 1:13:39.720 --> 1:13:42.400 and they had some guarantees about generalization 1:13:42.400 --> 1:13:45.600 which is unseen data or out of distribution 1:13:45.600 --> 1:13:47.000 or even within distribution 1:13:47.000 --> 1:13:49.760 where you take an image adding a bit of noise, 1:13:49.760 --> 1:13:51.280 these models fail. 1:13:51.280 --> 1:13:56.280 So I think really I don't see a lot of progress 1:13:56.280 --> 1:14:00.800 on generalization in the strong generalization sense 1:14:00.800 --> 1:14:01.880 of the word. 1:14:01.880 --> 1:14:05.280 I think our neural networks, 1:14:05.280 --> 1:14:08.000 you can always find design examples 1:14:08.000 --> 1:14:11.000 that will make their outputs arbitrary 1:14:11.000 --> 1:14:16.000 which is not good because we humans would never be fooled 1:14:16.000 --> 1:14:19.920 by these kind of images or manipulation of the image. 1:14:19.920 --> 1:14:21.720 And if you look at the mathematics, 1:14:21.720 --> 1:14:23.960 you kind of understand this is a bunch of matrices 1:14:23.960 --> 1:14:27.320 multiplied together, there's probably numerics 1:14:27.320 --> 1:14:30.880 and instability that you can just find corner cases. 1:14:30.880 --> 1:14:34.560 So I think that's really the underlying topic 1:14:34.560 --> 1:14:38.760 many times we see when even at the grand stage 1:14:38.760 --> 1:14:40.840 of like during test generalization, 1:14:40.840 --> 1:14:44.560 I mean, if you start, I mean, passing the during test, 1:14:44.560 --> 1:14:47.920 should it be in English or should it be in any language? 1:14:47.920 --> 1:14:52.320 I mean, as a human, if you ask something 1:14:52.320 --> 1:14:54.120 in a different language, you actually will go 1:14:54.120 --> 1:14:56.280 and do some research and try to translate it 1:14:56.280 --> 1:15:01.080 and so on, should the during test include that, right? 1:15:01.080 --> 1:15:02.920 And it's really a difficult problem 1:15:02.920 --> 1:15:05.360 and very fascinating and very mysterious actually. 1:15:05.360 --> 1:15:06.320 Yeah, absolutely. 1:15:06.320 --> 1:15:09.120 But do you think it's, if you were to try to solve it, 1:15:10.520 --> 1:15:14.280 can you not grow the size of data intelligently 1:15:14.280 --> 1:15:17.400 in such a way that the distribution of your training set 1:15:17.400 --> 1:15:20.360 does include the entirety of the testing set? 1:15:20.360 --> 1:15:21.800 I think is that one path? 1:15:21.800 --> 1:15:23.880 The other path is totally a new methodology. 1:15:23.880 --> 1:15:25.000 That's not statistical. 1:15:25.000 --> 1:15:27.080 So a path that has worked well 1:15:27.080 --> 1:15:29.880 and it worked well in StarCraft and in machine translation 1:15:29.880 --> 1:15:32.800 and in language is scaling up the data and the model. 1:15:32.800 --> 1:15:37.400 And that's kind of been maybe the only single formula 1:15:37.400 --> 1:15:40.440 that still delivers today in deep learning, right? 1:15:40.440 --> 1:15:44.080 It's that scale, data scale and model scale 1:15:44.080 --> 1:15:47.080 really do more and more of the things that we thought, 1:15:47.080 --> 1:15:49.240 oh, there's no way it can generalize to these 1:15:49.240 --> 1:15:51.360 or there's no way it can generalize to that. 1:15:51.360 --> 1:15:54.840 But I don't think fundamentally it will be solved with this. 1:15:54.840 --> 1:15:58.960 And for instance, I'm really liking some style 1:15:58.960 --> 1:16:02.120 or approach that would not only have neural networks 1:16:02.120 --> 1:16:06.400 but it would have programs or some discrete decision making 1:16:06.400 --> 1:16:09.760 because there is where I feel there's a bit more, 1:16:09.760 --> 1:16:12.200 like, I mean, the example of the best example, 1:16:12.200 --> 1:16:14.680 I think for understanding this is, 1:16:14.680 --> 1:16:17.640 I also worked a bit on, oh, like we can learn an algorithm 1:16:17.640 --> 1:16:18.840 with a neural network, right? 1:16:18.840 --> 1:16:20.160 So you give it many examples 1:16:20.160 --> 1:16:22.880 and it's gonna sort the input numbers 1:16:22.880 --> 1:16:24.440 or something like that. 1:16:24.440 --> 1:16:29.520 But really, strong generalization is you give me some numbers 1:16:29.520 --> 1:16:32.360 or you ask me to create an algorithm that sorts numbers 1:16:32.360 --> 1:16:34.760 and instead of creating a neural net which will be fragile 1:16:34.760 --> 1:16:38.000 because it's gonna go out of range at some point, 1:16:38.000 --> 1:16:40.400 you're gonna give it numbers that are too large, 1:16:40.400 --> 1:16:42.680 too small and whatnot, you just, 1:16:42.680 --> 1:16:46.400 if you just create a piece of code that sorts the numbers, 1:16:46.400 --> 1:16:48.760 then you can prove that that will generalize 1:16:48.760 --> 1:16:52.040 to absolutely all the possible inputs you could give. 1:16:52.040 --> 1:16:53.920 So I think that's, the problem comes 1:16:53.920 --> 1:16:56.000 with some exciting prospects. 1:16:56.000 --> 1:16:59.560 I mean, scale is a bit more boring, but it really works. 1:16:59.560 --> 1:17:02.960 And then maybe programs and discrete abstractions 1:17:02.960 --> 1:17:04.920 are a bit less developed, 1:17:04.920 --> 1:17:07.520 but clearly I think they're quite exciting 1:17:07.520 --> 1:17:10.000 in terms of future for the field. 1:17:10.000 --> 1:17:13.560 Do you draw any insight wisdom from the 80s 1:17:13.560 --> 1:17:17.000 and expert systems and symbolic systems, symbolic computing? 1:17:17.000 --> 1:17:18.920 Do you ever go back to those, 1:17:18.920 --> 1:17:20.800 the reasoning, that kind of logic? 1:17:20.800 --> 1:17:23.200 Do you think that might make a comeback? 1:17:23.200 --> 1:17:25.000 You'll have to dust off those books? 1:17:25.000 --> 1:17:30.000 Yeah, I actually love actually adding more inductive biases. 1:17:31.360 --> 1:17:34.360 To me, the problem really is what are you trying to solve? 1:17:34.360 --> 1:17:36.560 If what you're trying to solve is so important 1:17:36.560 --> 1:17:39.240 that try to solve it no matter what, 1:17:39.240 --> 1:17:44.240 then absolutely use rules, use domain knowledge 1:17:44.280 --> 1:17:46.960 and then use a bit of the magic of machine learning 1:17:46.960 --> 1:17:50.160 to empower or to make the system as the best system 1:17:50.160 --> 1:17:55.160 that will detect cancer or detect weather patterns, right? 1:17:56.080 --> 1:17:59.160 Or in terms of StarCraft, it also was a very big challenge. 1:17:59.160 --> 1:18:01.320 So I was definitely happy 1:18:01.320 --> 1:18:04.560 that if we had to cut a corner here and there, 1:18:04.560 --> 1:18:06.920 it could have been interesting to do. 1:18:06.920 --> 1:18:08.400 And in fact, in StarCraft, 1:18:08.400 --> 1:18:10.600 we start thinking about expert systems 1:18:10.600 --> 1:18:12.840 because it's a very, you can define, 1:18:12.840 --> 1:18:15.120 I mean, people actually build StarCraft bots 1:18:15.120 --> 1:18:18.720 by thinking about those principles like state machines 1:18:18.720 --> 1:18:21.600 and rule based and then you could think 1:18:21.600 --> 1:18:24.520 of combining a bit of a rule based system, 1:18:24.520 --> 1:18:27.480 but that has also neural networks incorporated 1:18:27.480 --> 1:18:29.080 to make it generalize a bit better. 1:18:29.080 --> 1:18:31.840 So absolutely, I mean, we should definitely go back 1:18:31.840 --> 1:18:35.440 to those ideas and anything that makes the problem simpler. 1:18:35.440 --> 1:18:38.040 As long as your problem is important, that's okay. 1:18:38.040 --> 1:18:41.080 And that's research driving a very important problem. 1:18:41.080 --> 1:18:42.160 And on the other hand, 1:18:42.160 --> 1:18:45.240 if you wanna really focus on the limits 1:18:45.240 --> 1:18:47.240 of reinforcement learning, then of course, 1:18:47.240 --> 1:18:50.800 you must try not to look at imitation data 1:18:50.800 --> 1:18:54.200 or to look for some rules of the domain 1:18:54.200 --> 1:18:57.040 that would help a lot or even feature engineering, right? 1:18:57.040 --> 1:19:00.760 So this is a tension that depending on what you do, 1:19:00.760 --> 1:19:03.360 I think both ways are definitely fine. 1:19:03.360 --> 1:19:06.080 And I would never not do one or the other 1:19:06.080 --> 1:19:08.040 if you're, as long as what you're doing 1:19:08.040 --> 1:19:10.080 is important and needs to be solved, right? 1:19:10.080 --> 1:19:13.520 All right, so there's a bunch of different ideas 1:19:13.520 --> 1:19:16.920 that you've developed that I really enjoy. 1:19:16.920 --> 1:19:21.920 But one is translating from image captioning, 1:19:22.240 --> 1:19:23.960 translating from image to text. 1:19:23.960 --> 1:19:28.720 Just another beautiful idea, I think, 1:19:28.720 --> 1:19:33.240 that resonates throughout your work, actually. 1:19:33.240 --> 1:19:36.760 So the underlying nature of reality being language always. 1:19:36.760 --> 1:19:37.680 Yeah, somehow. 1:19:37.680 --> 1:19:42.520 So what's the connection between images and text? 1:19:42.520 --> 1:19:45.880 Or rather the visual world and the world of language 1:19:45.880 --> 1:19:46.720 in your view? 1:19:46.720 --> 1:19:51.480 Right, so I think a piece of research that's been central 1:19:51.480 --> 1:19:54.400 to, I would say, even extending into StarCraft 1:19:54.400 --> 1:19:57.680 is this idea of sequence to sequence learning, 1:19:57.680 --> 1:19:59.840 which what we really meant by that 1:19:59.840 --> 1:20:03.520 is that you can now really input anything 1:20:03.520 --> 1:20:06.160 to a neural network as the input X 1:20:06.160 --> 1:20:09.600 and then the neural network will learn a function F 1:20:09.600 --> 1:20:12.840 that will take X as an input and produce any output Y. 1:20:12.840 --> 1:20:16.240 And these X and Ys don't need to be like static 1:20:16.240 --> 1:20:21.240 or like a feature, like a fixed vectors 1:20:21.240 --> 1:20:22.240 or anything like that. 1:20:22.240 --> 1:20:23.800 It could be really sequences 1:20:23.800 --> 1:20:26.600 and now beyond like data structures, right? 1:20:26.600 --> 1:20:31.600 So that paradigm was tested in a very interesting way 1:20:31.600 --> 1:20:35.760 when we moved from translating French to English 1:20:35.760 --> 1:20:37.960 to translating an image to its caption. 1:20:37.960 --> 1:20:40.760 But the beauty of it is that really, 1:20:40.760 --> 1:20:42.160 and that's actually how it happened. 1:20:42.160 --> 1:20:45.240 I ran, I changed a line of code in this thing 1:20:45.240 --> 1:20:47.520 that was doing machine translation 1:20:47.520 --> 1:20:51.800 and I came the next day and I saw how it was producing 1:20:51.800 --> 1:20:54.200 captions that seemed like, oh my God, 1:20:54.200 --> 1:20:56.040 this is really, really working. 1:20:56.040 --> 1:20:57.560 And the principle is the same, right? 1:20:57.560 --> 1:21:02.560 So I think I don't see text, vision, speech, way forms 1:21:02.560 --> 1:21:07.560 as something different, as long as you basically learn 1:21:08.120 --> 1:21:13.120 a function that will vectorize these into, 1:21:13.480 --> 1:21:17.480 and then after we vectorize it, we can then use transformers, 1:21:17.480 --> 1:21:21.160 LSTMs, whatever the flavor of the month of the model is. 1:21:21.160 --> 1:21:24.280 And then as long as we have enough supervised data, 1:21:24.280 --> 1:21:28.280 really this formula will work and will keep working, 1:21:28.280 --> 1:21:30.280 I believe to some extent. 1:21:30.280 --> 1:21:33.360 Model of these generalization issues that I mentioned before. 1:21:33.360 --> 1:21:35.400 So, but the task there is to vectorize 1:21:35.400 --> 1:21:37.880 sort of form a representation that's meaningful, 1:21:37.880 --> 1:21:41.400 and your intuition now, having worked with all this media, 1:21:41.400 --> 1:21:45.240 is that once you are able to form that representation, 1:21:45.240 --> 1:21:47.960 you could basically take anything, any sequence. 1:21:48.960 --> 1:21:51.240 Is there, going back to StarCraft, 1:21:51.240 --> 1:21:52.800 is there limits on the length? 1:21:54.080 --> 1:21:57.960 So we didn't really touch on the long term aspect. 1:21:57.960 --> 1:22:01.640 How did you overcome the whole really long term aspect 1:22:01.640 --> 1:22:02.480 of things here? 1:22:02.480 --> 1:22:03.920 Is there some tricks or is it? 1:22:03.920 --> 1:22:07.000 So the main trick, so StarCraft, 1:22:07.000 --> 1:22:09.360 if you look at absolutely every frame, 1:22:09.360 --> 1:22:11.120 you might think it's quite a long game. 1:22:11.120 --> 1:22:14.440 So we would have to multiply 22 times, 1:22:15.600 --> 1:22:18.200 60 seconds per minute times maybe 1:22:18.200 --> 1:22:20.360 at least 10 minutes per game on average. 1:22:20.360 --> 1:22:24.160 So there are quite a few frames, 1:22:24.160 --> 1:22:26.600 but the trick really was to, 1:22:26.600 --> 1:22:30.760 only observe, in fact, which might be seen as a limitation, 1:22:30.760 --> 1:22:33.600 but it is also a computational advantage. 1:22:33.600 --> 1:22:36.040 Only observe when you act. 1:22:36.040 --> 1:22:38.440 And then what the neural network decides 1:22:38.440 --> 1:22:42.200 is what is the gap gonna be until the next action? 1:22:43.200 --> 1:22:46.520 And if you look at most StarCraft games 1:22:46.520 --> 1:22:50.200 that we have in the data set that Blizzard provided, 1:22:50.200 --> 1:22:54.720 it turns out that most games are actually only, 1:22:54.720 --> 1:22:56.720 I mean, it is still a long sequence, 1:22:56.720 --> 1:23:00.720 but it's maybe like 1,000 to 1,500 actions, 1:23:00.720 --> 1:23:04.720 which if you start looking at LSTMs, 1:23:04.720 --> 1:23:07.720 large LSTMs, transformers, 1:23:07.720 --> 1:23:10.720 it's not that difficult, 1:23:10.720 --> 1:23:13.720 especially if you have supervised learning. 1:23:13.720 --> 1:23:15.720 If you had to do it with reinforcement learning, 1:23:15.720 --> 1:23:16.720 the credit assignment problem, 1:23:16.720 --> 1:23:18.720 what is it that in this game that made you win? 1:23:18.720 --> 1:23:20.720 That would be really difficult. 1:23:20.720 --> 1:23:23.720 But thankfully, because of imitation learning, 1:23:23.720 --> 1:23:26.720 we didn't kind of have to deal with this directly. 1:23:26.720 --> 1:23:28.720 Although if we had to, we tried it, 1:23:28.720 --> 1:23:30.720 and what happened is you just take all your workers 1:23:30.720 --> 1:23:32.720 and attack with them. 1:23:32.720 --> 1:23:35.720 And that sort of is kind of obvious in retrospect, 1:23:35.720 --> 1:23:37.720 because you start trying random actions. 1:23:37.720 --> 1:23:39.720 One of the actions will be a worker 1:23:39.720 --> 1:23:40.720 that goes to the enemy base, 1:23:40.720 --> 1:23:42.720 and because it's self play, 1:23:42.720 --> 1:23:44.720 it's not gonna know how to defend, 1:23:44.720 --> 1:23:46.720 because it basically doesn't know almost anything. 1:23:46.720 --> 1:23:48.720 And eventually what you develop is this, 1:23:48.720 --> 1:23:51.720 take all workers and attack, 1:23:51.720 --> 1:23:54.720 because the credit assignment issue in our rally 1:23:54.720 --> 1:23:55.720 is really, really hard. 1:23:55.720 --> 1:23:57.720 I do believe we could do better, 1:23:57.720 --> 1:24:00.720 and that's maybe a research challenge for the future. 1:24:00.720 --> 1:24:03.720 But yeah, even in StarCraft, 1:24:03.720 --> 1:24:05.720 the sequences are maybe 1,000, 1:24:05.720 --> 1:24:08.720 which I believe is within the realm 1:24:08.720 --> 1:24:10.720 of what transformers can do. 1:24:10.720 --> 1:24:13.720 Yeah, I guess the difference between StarCraft and Go 1:24:13.720 --> 1:24:15.720 is in Go and chess, 1:24:15.720 --> 1:24:17.720 stuff starts happening right away. 1:24:17.720 --> 1:24:18.720 Right. 1:24:18.720 --> 1:24:21.720 Yeah, it's pretty easy to self play, 1:24:21.720 --> 1:24:23.720 not easy, but to self play is possible 1:24:23.720 --> 1:24:25.720 to develop reasonable strategies quickly 1:24:25.720 --> 1:24:27.720 as opposed to StarCraft. 1:24:27.720 --> 1:24:29.720 In Go, there's only 400 actions, 1:24:29.720 --> 1:24:32.720 but one action is what people would call 1:24:32.720 --> 1:24:34.720 the God action that would be, 1:24:34.720 --> 1:24:37.720 if you had expanded the whole search tree, 1:24:37.720 --> 1:24:39.720 that's the best action if you did minimax 1:24:39.720 --> 1:24:41.720 or whatever algorithm you would do 1:24:41.720 --> 1:24:43.720 if you had the computational capacity. 1:24:43.720 --> 1:24:45.720 But in StarCraft, 1:24:45.720 --> 1:24:48.720 400 is minuscule. 1:24:48.720 --> 1:24:51.720 In 400, you couldn't even click 1:24:51.720 --> 1:24:53.720 on the pixels around a unit, right? 1:24:53.720 --> 1:24:55.720 So I think the problem there 1:24:55.720 --> 1:24:58.720 is in terms of action space size 1:24:58.720 --> 1:25:00.720 is way harder, 1:25:00.720 --> 1:25:03.720 and that search is impossible. 1:25:03.720 --> 1:25:05.720 So there's quite a few challenges indeed 1:25:05.720 --> 1:25:08.720 that make this kind of a step up 1:25:08.720 --> 1:25:10.720 in terms of machine learning. 1:25:10.720 --> 1:25:12.720 For humans, maybe playing StarCraft 1:25:12.720 --> 1:25:14.720 seems more intuitive 1:25:14.720 --> 1:25:16.720 because it looks real, 1:25:16.720 --> 1:25:18.720 the graphics and everything moves smoothly, 1:25:18.720 --> 1:25:20.720 whereas I don't know how to... 1:25:20.720 --> 1:25:22.720 Go is a game that I wouldn't really need to study. 1:25:22.720 --> 1:25:24.720 It feels quite complicated, 1:25:24.720 --> 1:25:26.720 but for machines, maybe it's the reverse, yes. 1:25:26.720 --> 1:25:28.720 Which shows you the gap, actually, 1:25:28.720 --> 1:25:30.720 between deep learning and however the heck 1:25:30.720 --> 1:25:32.720 our brains work. 1:25:32.720 --> 1:25:35.720 So you developed a lot of really interesting ideas. 1:25:35.720 --> 1:25:37.720 It's interesting to just ask, 1:25:37.720 --> 1:25:40.720 what's your process of developing new ideas? 1:25:40.720 --> 1:25:42.720 Do you like brainstorming with others? 1:25:42.720 --> 1:25:44.720 Do you like thinking alone? 1:25:44.720 --> 1:25:46.720 Do you like... 1:25:46.720 --> 1:25:48.720 Like what was it? 1:25:48.720 --> 1:25:50.720 Ian Goodfellow said he came up with Gans 1:25:50.720 --> 1:25:52.720 after a few beers. 1:25:52.720 --> 1:25:54.720 He thinks beers are essential 1:25:54.720 --> 1:25:56.720 for coming up with new ideas. 1:25:56.720 --> 1:25:58.720 We had beers to decide to play another game 1:25:58.720 --> 1:26:00.720 of StarCraft after a week, 1:26:00.720 --> 1:26:02.720 so it's really similar to that story. 1:26:02.720 --> 1:26:04.720 Actually, I explained this 1:26:04.720 --> 1:26:06.720 in a deep mind retreat, 1:26:06.720 --> 1:26:08.720 and I said this is the same as the Gans story. 1:26:08.720 --> 1:26:10.720 I mean, we were on a bar and we decided, 1:26:10.720 --> 1:26:12.720 we were on a week and that's what happened. 1:26:12.720 --> 1:26:14.720 I feel like we're giving the wrong message 1:26:14.720 --> 1:26:16.720 to young undergrads. 1:26:16.720 --> 1:26:18.720 But in general, do you like brainstorming? 1:26:18.720 --> 1:26:20.720 Do you like thinking alone, working stuff out? 1:26:20.720 --> 1:26:22.720 So I think throughout the years 1:26:22.720 --> 1:26:24.720 also things changed, right? 1:26:24.720 --> 1:26:26.720 So initially, I was 1:26:26.720 --> 1:26:28.720 very fortunate to be 1:26:28.720 --> 1:26:30.720 with great minds like 1:26:30.720 --> 1:26:32.720 Jeff Hinton, 1:26:32.720 --> 1:26:34.720 Jeff Dean, Ilya Tsutskiber. 1:26:34.720 --> 1:26:36.720 I was really fortunate to join Brain 1:26:36.720 --> 1:26:38.720 at a very good time. 1:26:38.720 --> 1:26:40.720 At that point, ideas, 1:26:40.720 --> 1:26:42.720 I was just kind of brainstorming with my colleagues 1:26:42.720 --> 1:26:44.720 and learned a lot, 1:26:44.720 --> 1:26:46.720 and keep learning is actually 1:26:46.720 --> 1:26:48.720 something you should never stop doing, right? 1:26:48.720 --> 1:26:50.720 So learning implies 1:26:50.720 --> 1:26:52.720 reading papers and also discussing ideas 1:26:52.720 --> 1:26:54.720 with others. It's very hard 1:26:54.720 --> 1:26:56.720 at some point to not communicate 1:26:56.720 --> 1:26:58.720 that being reading a paper from someone 1:26:58.720 --> 1:27:00.720 or actually discussing, right? 1:27:00.720 --> 1:27:02.720 So definitely 1:27:02.720 --> 1:27:04.720 that communication aspect 1:27:04.720 --> 1:27:06.720 needs to be there, whether it's written 1:27:06.720 --> 1:27:08.720 or oral. 1:27:08.720 --> 1:27:10.720 Nowadays, 1:27:10.720 --> 1:27:12.720 I'm also trying to be a bit more strategic 1:27:12.720 --> 1:27:14.720 about what research to do. 1:27:14.720 --> 1:27:16.720 So I was describing 1:27:16.720 --> 1:27:18.720 a little bit this sort of tension between 1:27:18.720 --> 1:27:20.720 research for the sake of research, 1:27:20.720 --> 1:27:22.720 and then you have, on the other hand, 1:27:22.720 --> 1:27:24.720 applications that can drive the research, right? 1:27:24.720 --> 1:27:26.720 And honestly, 1:27:26.720 --> 1:27:28.720 the formula that has worked best for me is 1:27:28.720 --> 1:27:30.720 just find a hard problem 1:27:30.720 --> 1:27:32.720 and then try to 1:27:32.720 --> 1:27:34.720 see how research fits into it, 1:27:34.720 --> 1:27:36.720 how it doesn't fit into it, 1:27:36.720 --> 1:27:38.720 and then you must innovate. 1:27:38.720 --> 1:27:40.720 So I think machine translation 1:27:40.720 --> 1:27:42.720 drove sequence to sequence. 1:27:42.720 --> 1:27:44.720 Then maybe 1:27:44.720 --> 1:27:46.720 learning algorithms 1:27:46.720 --> 1:27:48.720 that had to, like combinatorial algorithms 1:27:48.720 --> 1:27:50.720 led to pointer networks. 1:27:50.720 --> 1:27:52.720 Starcraft led to really scaling up 1:27:52.720 --> 1:27:54.720 imitation learning and the Alpha Star League. 1:27:54.720 --> 1:27:56.720 So that's been a formula 1:27:56.720 --> 1:27:58.720 that I personally like, 1:27:58.720 --> 1:28:00.720 but the other one is also valid, 1:28:00.720 --> 1:28:02.720 and I see it succeed a lot of the times 1:28:02.720 --> 1:28:04.720 where you just want to investigate 1:28:04.720 --> 1:28:06.720 model based RL 1:28:06.720 --> 1:28:08.720 as a kind of a research topic, 1:28:08.720 --> 1:28:10.720 and then you must then start to think, 1:28:10.720 --> 1:28:12.720 well, how are the tests? 1:28:12.720 --> 1:28:14.720 How are you going to test these ideas? 1:28:14.720 --> 1:28:16.720 You need kind of a minimal 1:28:16.720 --> 1:28:18.720 environment to try things. 1:28:18.720 --> 1:28:20.720 You need to read a lot of papers and so on, 1:28:20.720 --> 1:28:22.720 and that's also very fun to do, 1:28:22.720 --> 1:28:24.720 and something I've also done quite a few times, 1:28:24.720 --> 1:28:26.720 both at Brain, at DeepMind, 1:28:26.720 --> 1:28:28.720 and obviously as a PhD. 1:28:28.720 --> 1:28:30.720 So I think 1:28:30.720 --> 1:28:32.720 the ideas and discussions, 1:28:32.720 --> 1:28:34.720 I think it's important also 1:28:34.720 --> 1:28:36.720 because you start sort of 1:28:36.720 --> 1:28:38.720 guiding not only 1:28:38.720 --> 1:28:40.720 your own goals, but 1:28:40.720 --> 1:28:42.720 other people's goals 1:28:42.720 --> 1:28:44.720 to the next breakthrough, so 1:28:44.720 --> 1:28:46.720 you must really kind of understand 1:28:46.720 --> 1:28:48.720 this feasibility also 1:28:48.720 --> 1:28:50.720 as we were discussing before, right? 1:28:50.720 --> 1:28:52.720 Whether this domain is ready 1:28:52.720 --> 1:28:54.720 to be tackled or not, and you don't want 1:28:54.720 --> 1:28:56.720 to be too early, you obviously don't want 1:28:56.720 --> 1:28:58.720 to be too late, so it's really interesting 1:28:58.720 --> 1:29:00.720 this strategic component of research, 1:29:00.720 --> 1:29:02.720 which I think as a grad student 1:29:02.720 --> 1:29:04.720 I just had no idea, 1:29:04.720 --> 1:29:06.720 I just read papers and discussed 1:29:06.720 --> 1:29:08.720 ideas, and I think this has been maybe 1:29:08.720 --> 1:29:10.720 the major change, and I recommend 1:29:10.720 --> 1:29:12.720 people kind of 1:29:12.720 --> 1:29:14.720 fit forward to success, how it looks like 1:29:14.720 --> 1:29:16.720 and try to backtrack, other than just 1:29:16.720 --> 1:29:18.720 kind of looking out, this looks cool, 1:29:18.720 --> 1:29:20.720 this looks cool, and then you do a bit of 1:29:20.720 --> 1:29:22.720 random work, which sometimes you stumble upon 1:29:22.720 --> 1:29:24.720 some interesting things, but 1:29:24.720 --> 1:29:26.720 in general it's also good to plan a bit. 1:29:26.720 --> 1:29:28.720 Yeah, I like it. 1:29:28.720 --> 1:29:30.720 Especially like your approach of taking 1:29:30.720 --> 1:29:32.720 on really hard problems, stepping right in 1:29:32.720 --> 1:29:34.720 and then being super skeptical about 1:29:34.720 --> 1:29:36.720 being able to solve the problem. 1:29:36.720 --> 1:29:38.720 I mean, there's a 1:29:38.720 --> 1:29:40.720 balance of both, right? There's a silly 1:29:40.720 --> 1:29:42.720 optimism 1:29:42.720 --> 1:29:44.720 and a critical 1:29:44.720 --> 1:29:46.720 sort of skepticism 1:29:46.720 --> 1:29:48.720 that's good to balance, which 1:29:48.720 --> 1:29:50.720 is why it's good to have a team of people 1:29:50.720 --> 1:29:52.720 that balance that. 1:29:52.720 --> 1:29:54.720 You don't do that on your own, you have both 1:29:54.720 --> 1:29:56.720 mentors that have seen 1:29:56.720 --> 1:29:58.720 or you obviously want to chat and 1:29:58.720 --> 1:30:00.720 discuss whether it's the right time. 1:30:00.720 --> 1:30:02.720 I mean, Damis 1:30:02.720 --> 1:30:04.720 came in 2014 and he said 1:30:04.720 --> 1:30:06.720 maybe in a bit we'll do StarCraft and 1:30:06.720 --> 1:30:08.720 maybe he knew 1:30:08.720 --> 1:30:10.720 and I'm just following his lead, which 1:30:10.720 --> 1:30:12.720 is great because he's brilliant, right? 1:30:12.720 --> 1:30:14.720 So, these things are 1:30:14.720 --> 1:30:16.720 obviously quite 1:30:16.720 --> 1:30:18.720 important that you want to 1:30:18.720 --> 1:30:20.720 be surrounded by people 1:30:20.720 --> 1:30:22.720 who are diverse, they 1:30:22.720 --> 1:30:24.720 have their knowledge. There's also 1:30:24.720 --> 1:30:26.720 important to... 1:30:26.720 --> 1:30:28.720 I've learned a lot from people 1:30:28.720 --> 1:30:30.720 who actually 1:30:30.720 --> 1:30:32.720 have an idea that I might not think it's good 1:30:32.720 --> 1:30:34.720 but if I give them the space to try it 1:30:34.720 --> 1:30:36.720 I've been proven wrong many, many times 1:30:36.720 --> 1:30:38.720 as well. So, that's great. 1:30:38.720 --> 1:30:40.720 I think it's... 1:30:40.720 --> 1:30:42.720 Your colleagues are more important than yourself 1:30:42.720 --> 1:30:44.720 I think so. 1:30:44.720 --> 1:30:46.720 Now, let's real quick 1:30:46.720 --> 1:30:48.720 talk about another impossible problem. 1:30:48.720 --> 1:30:50.720 AGI. 1:30:50.720 --> 1:30:52.720 What do you think it takes to build a system 1:30:52.720 --> 1:30:54.720 that's human level intelligence? 1:30:54.720 --> 1:30:56.720 We talked a little bit about the touring test, StarCraft 1:30:56.720 --> 1:30:58.720 all these have echoes of general intelligence 1:30:58.720 --> 1:31:00.720 but if you think about 1:31:00.720 --> 1:31:02.720 just something that you would sit back 1:31:02.720 --> 1:31:04.720 and say, wow, this is 1:31:04.720 --> 1:31:06.720 really something that resembles 1:31:06.720 --> 1:31:08.720 human level intelligence, what do you think it takes 1:31:08.720 --> 1:31:10.720 to build that? 1:31:10.720 --> 1:31:12.720 So, I find that 1:31:12.720 --> 1:31:14.720 AGI oftentimes is maybe not 1:31:14.720 --> 1:31:16.720 very well defined 1:31:16.720 --> 1:31:18.720 so what I'm trying to 1:31:18.720 --> 1:31:20.720 then come up with for myself is 1:31:20.720 --> 1:31:22.720 what would be a result 1:31:22.720 --> 1:31:24.720 look like that 1:31:24.720 --> 1:31:26.720 you would start to believe that 1:31:26.720 --> 1:31:28.720 you would have agents or neural nets 1:31:28.720 --> 1:31:30.720 that no longer sort of overfit 1:31:30.720 --> 1:31:32.720 to a single task, right? 1:31:32.720 --> 1:31:34.720 But actually 1:31:34.720 --> 1:31:36.720 kind of learn 1:31:36.720 --> 1:31:38.720 the skill of learning, so to speak 1:31:38.720 --> 1:31:40.720 and that actually is a field that I 1:31:40.720 --> 1:31:42.720 am fascinated by which is 1:31:42.720 --> 1:31:44.720 the learning to learn or meta learning 1:31:44.720 --> 1:31:46.720 which is about no longer 1:31:46.720 --> 1:31:48.720 learning about a single domain 1:31:48.720 --> 1:31:50.720 so you can think about the learning algorithm 1:31:50.720 --> 1:31:52.720 itself is general, right? 1:31:52.720 --> 1:31:54.720 So the same formula we applied for 1:31:54.720 --> 1:31:56.720 Alpha Star or StarCraft 1:31:56.720 --> 1:31:58.720 we can now apply to kind of almost any 1:31:58.720 --> 1:32:00.720 video game or you could apply to 1:32:00.720 --> 1:32:02.720 many other problems and domains 1:32:02.720 --> 1:32:04.720 but the algorithm 1:32:04.720 --> 1:32:06.720 is what's kind of generalizing 1:32:06.720 --> 1:32:08.720 but the neural network, the weights 1:32:08.720 --> 1:32:10.720 those weights are useless even 1:32:10.720 --> 1:32:12.720 to play another race, right? I train 1:32:12.720 --> 1:32:14.720 a network to play very well at PROTOS vs PROTOS 1:32:14.720 --> 1:32:16.720 I need to throw away those weights 1:32:16.720 --> 1:32:18.720 if I want to play 1:32:18.720 --> 1:32:20.720 now Terran vs Terran 1:32:20.720 --> 1:32:22.720 I would need to retrain 1:32:22.720 --> 1:32:24.720 a network from scratch with the same algorithm 1:32:24.720 --> 1:32:26.720 that's beautiful but the network 1:32:26.720 --> 1:32:28.720 itself will not be useful 1:32:28.720 --> 1:32:30.720 so I think when I, if I see 1:32:30.720 --> 1:32:32.720 an approach that 1:32:32.720 --> 1:32:34.720 can absorb or start 1:32:34.720 --> 1:32:36.720 solving new problems 1:32:36.720 --> 1:32:38.720 without the need to kind of restart 1:32:38.720 --> 1:32:40.720 the process I think that 1:32:40.720 --> 1:32:42.720 to me would be a nice way to define 1:32:42.720 --> 1:32:44.720 some form of AGI 1:32:44.720 --> 1:32:46.720 again, I don't know 1:32:46.720 --> 1:32:48.720 the grandiose like age, I mean 1:32:48.720 --> 1:32:50.720 during tests we solve before AGI 1:32:50.720 --> 1:32:52.720 I mean, I don't know, I think concretely 1:32:52.720 --> 1:32:54.720 I would like to see clearly 1:32:54.720 --> 1:32:56.720 that meta learning happen 1:32:56.720 --> 1:32:58.720 meaning there is 1:32:58.720 --> 1:33:00.720 an architecture or a network 1:33:00.720 --> 1:33:02.720 that as it sees new problem 1:33:02.720 --> 1:33:04.720 or new data it solves it 1:33:04.720 --> 1:33:06.720 and to make it 1:33:06.720 --> 1:33:08.720 kind of a benchmark it should 1:33:08.720 --> 1:33:10.720 solve it at the same speed that we do solve 1:33:10.720 --> 1:33:12.720 new problems when I define 1:33:12.720 --> 1:33:14.720 a new object and you have to recognize it 1:33:14.720 --> 1:33:16.720 when you start playing a new game 1:33:16.720 --> 1:33:18.720 you played all the Atari games but now you play a new Atari game 1:33:18.720 --> 1:33:20.720 well, you're going to be 1:33:20.720 --> 1:33:22.720 pretty quickly pretty good at the game 1:33:22.720 --> 1:33:24.720 so that's perhaps 1:33:24.720 --> 1:33:26.720 what's the domain and what's the exact benchmark 1:33:26.720 --> 1:33:28.720 it's a bit difficult, I think as a community 1:33:28.720 --> 1:33:30.720 we might need to do some work to define it 1:33:32.720 --> 1:33:34.720 but I think this first step 1:33:34.720 --> 1:33:36.720 I could see it happen relatively soon 1:33:36.720 --> 1:33:38.720 but then the whole 1:33:38.720 --> 1:33:40.720 what AGI means and so on 1:33:40.720 --> 1:33:42.720 I am a bit more confused about 1:33:42.720 --> 1:33:44.720 what I think people mean different things 1:33:44.720 --> 1:33:46.720 there's an emotional psychological level 1:33:48.720 --> 1:33:50.720 that 1:33:50.720 --> 1:33:52.720 like even the Turing test, passing the Turing test 1:33:52.720 --> 1:33:54.720 is something that we just pass judgment 1:33:54.720 --> 1:33:56.720 on as human beings what it means to be 1:33:56.720 --> 1:33:58.720 you know, as a 1:33:58.720 --> 1:34:00.720 as a dog 1:34:00.720 --> 1:34:02.720 an AGI system 1:34:02.720 --> 1:34:04.720 like what level, what does it mean 1:34:04.720 --> 1:34:06.720 what does it mean 1:34:06.720 --> 1:34:08.720 but I like the generalization 1:34:08.720 --> 1:34:10.720 and maybe as a community we converge towards 1:34:10.720 --> 1:34:12.720 a group of domains 1:34:12.720 --> 1:34:14.720 that are sufficiently far away 1:34:14.720 --> 1:34:16.720 that would be really damn impressive 1:34:16.720 --> 1:34:18.720 if we're able to generalize 1:34:18.720 --> 1:34:20.720 so perhaps not as close as Protoss and Zerg 1:34:20.720 --> 1:34:22.720 but like Wikipedia 1:34:22.720 --> 1:34:24.720 that would be a good step 1:34:24.720 --> 1:34:26.720 and then a really good step 1:34:26.720 --> 1:34:28.720 but then from Starcraft to Wikipedia 1:34:28.720 --> 1:34:30.720 and back 1:34:30.720 --> 1:34:32.720 that kind of thing 1:34:32.720 --> 1:34:34.720 and that feels also quite hard and far 1:34:34.720 --> 1:34:36.720 I think this 1:34:36.720 --> 1:34:38.720 as long as you put the benchmark out 1:34:38.720 --> 1:34:40.720 as we discovered for instance with ImageNet 1:34:40.720 --> 1:34:42.720 then tremendous progress can be had 1:34:42.720 --> 1:34:44.720 so I think maybe there's a lack of 1:34:44.720 --> 1:34:46.720 benchmark 1:34:46.720 --> 1:34:48.720 but I'm sure we'll find one and the community 1:34:48.720 --> 1:34:50.720 will then work towards that 1:34:52.720 --> 1:34:54.720 and then beyond what AGI might mean 1:34:54.720 --> 1:34:56.720 or would imply 1:34:56.720 --> 1:34:58.720 I really am hopeful to see 1:34:58.720 --> 1:35:00.720 basically machine learning 1:35:00.720 --> 1:35:02.720 or AI just scaling up 1:35:02.720 --> 1:35:04.720 and helping 1:35:04.720 --> 1:35:06.720 people that might not have the resources 1:35:06.720 --> 1:35:08.720 to hire an assistant 1:35:08.720 --> 1:35:10.720 or that 1:35:10.720 --> 1:35:12.720 they might not even know what the weather is like 1:35:12.720 --> 1:35:14.720 but so I think there's 1:35:14.720 --> 1:35:16.720 in terms of the impact 1:35:16.720 --> 1:35:18.720 the positive impact of AI 1:35:18.720 --> 1:35:20.720 I think that's maybe what we should also 1:35:20.720 --> 1:35:22.720 not lose focus 1:35:22.720 --> 1:35:24.720 the research community building AGI 1:35:24.720 --> 1:35:26.720 that's a real nice goal 1:35:26.720 --> 1:35:28.720 and I think the way that DeepMind puts it 1:35:28.720 --> 1:35:30.720 is and then use it to solve everything else 1:35:30.720 --> 1:35:32.720 so I think we should paralyze 1:35:32.720 --> 1:35:34.720 yeah we shouldn't forget 1:35:34.720 --> 1:35:36.720 of all the positive things that are actually 1:35:36.720 --> 1:35:38.720 coming out of AI already and are going 1:35:38.720 --> 1:35:40.720 to be coming out 1:35:40.720 --> 1:35:42.720 right 1:35:42.720 --> 1:35:44.720 and then let me ask 1:35:44.720 --> 1:35:46.720 relative to popular perception 1:35:46.720 --> 1:35:48.720 do you have any worry about the existential 1:35:48.720 --> 1:35:50.720 threat of artificial intelligence 1:35:50.720 --> 1:35:52.720 in the near or far future 1:35:52.720 --> 1:35:54.720 that some people have 1:35:54.720 --> 1:35:56.720 I think in the near future 1:35:56.720 --> 1:35:58.720 I'm skeptical so I hope 1:35:58.720 --> 1:36:00.720 I'm not wrong but 1:36:00.720 --> 1:36:02.720 I'm not concerned 1:36:02.720 --> 1:36:04.720 but I appreciate efforts 1:36:04.720 --> 1:36:06.720 ongoing efforts 1:36:06.720 --> 1:36:08.720 and even like a whole research field on 1:36:08.720 --> 1:36:10.720 AI safety emerging and in conferences 1:36:10.720 --> 1:36:12.720 and so on I think that's great 1:36:12.720 --> 1:36:14.720 in the long term 1:36:14.720 --> 1:36:16.720 I really hope we 1:36:16.720 --> 1:36:18.720 just can simply 1:36:18.720 --> 1:36:20.720 have the benefits outweigh the potential dangers 1:36:20.720 --> 1:36:22.720 I am hopeful for that 1:36:22.720 --> 1:36:24.720 but also we must 1:36:24.720 --> 1:36:26.720 remain vigilant to kind of monitor 1:36:26.720 --> 1:36:28.720 and assess whether the tradeoffs 1:36:28.720 --> 1:36:30.720 are there and we have 1:36:30.720 --> 1:36:32.720 enough 1:36:32.720 --> 1:36:34.720 also lead time to prevent 1:36:34.720 --> 1:36:36.720 or to redirect our efforts 1:36:36.720 --> 1:36:38.720 if need be 1:36:38.720 --> 1:36:40.720 but I'm quite optimistic 1:36:40.720 --> 1:36:42.720 about the technology 1:36:42.720 --> 1:36:44.720 and definitely more fearful 1:36:44.720 --> 1:36:46.720 of other threats in terms of 1:36:46.720 --> 1:36:48.720 planetary level 1:36:48.720 --> 1:36:50.720 at this point but obviously 1:36:50.720 --> 1:36:52.720 that's the one I kind of have more 1:36:52.720 --> 1:36:54.720 power on so clearly 1:36:54.720 --> 1:36:56.720 start thinking more and more about this 1:36:56.720 --> 1:36:58.720 and it's kind of 1:36:58.720 --> 1:37:00.720 it's grown in me actually to 1:37:00.720 --> 1:37:02.720 start reading more about AI safety 1:37:02.720 --> 1:37:04.720 which is a field that so far I have not 1:37:04.720 --> 1:37:06.720 really contributed to but maybe 1:37:06.720 --> 1:37:08.720 there's something to be done there as well 1:37:08.720 --> 1:37:10.720 I think it's really important 1:37:10.720 --> 1:37:12.720 I talk about this with a few folks 1:37:12.720 --> 1:37:14.720 but it's important to ask you 1:37:14.720 --> 1:37:16.720 and shove it in your head because you're at the 1:37:16.720 --> 1:37:18.720 leading edge of actually 1:37:18.720 --> 1:37:20.720 what people are excited about in AI 1:37:20.720 --> 1:37:22.720 I mean the work with AlphaStar 1:37:22.720 --> 1:37:24.720 at the very cutting edge of the kind 1:37:24.720 --> 1:37:26.720 of thing that people are afraid of 1:37:26.720 --> 1:37:28.720 and so you speaking 1:37:28.720 --> 1:37:30.720 to that fact and 1:37:30.720 --> 1:37:32.720 that we're actually quite far away 1:37:32.720 --> 1:37:34.720 to the kind of thing that people might be 1:37:34.720 --> 1:37:36.720 afraid of but it's still 1:37:36.720 --> 1:37:38.720 worthwhile to think about 1:37:38.720 --> 1:37:40.720 and it's also good that you're 1:37:40.720 --> 1:37:42.720 that you're not as worried 1:37:42.720 --> 1:37:44.720 and you're also open to 1:37:44.720 --> 1:37:46.720 I mean there's two aspects 1:37:46.720 --> 1:37:48.720 I mean me not being worried but obviously 1:37:48.720 --> 1:37:50.720 we should prepare 1:37:50.720 --> 1:37:52.720 for it 1:37:52.720 --> 1:37:54.720 for things that could 1:37:54.720 --> 1:37:56.720 go wrong, misuse of the technologies 1:37:56.720 --> 1:37:58.720 as with any technologies 1:37:58.720 --> 1:38:00.720 so I think 1:38:00.720 --> 1:38:02.720 there's always tradeoffs 1:38:02.720 --> 1:38:04.720 and as a society we've kind of 1:38:04.720 --> 1:38:06.720 solved this to some extent 1:38:06.720 --> 1:38:08.720 in the past so I'm hoping that 1:38:08.720 --> 1:38:10.720 by having the researchers 1:38:10.720 --> 1:38:12.720 and the whole community 1:38:12.720 --> 1:38:14.720 brainstorm and come up with 1:38:14.720 --> 1:38:16.720 interesting solutions to the new things 1:38:16.720 --> 1:38:18.720 that will happen in the future 1:38:18.720 --> 1:38:20.720 that we can still also push the research 1:38:20.720 --> 1:38:22.720 to the avenue that 1:38:22.720 --> 1:38:24.720 I think is kind of the greatest avenue 1:38:24.720 --> 1:38:26.720 which is to 1:38:26.720 --> 1:38:28.720 understand intelligence, right? How are we doing 1:38:28.720 --> 1:38:30.720 what we're doing and 1:38:30.720 --> 1:38:32.720 obviously from a scientific standpoint 1:38:32.720 --> 1:38:34.720 that is kind of the drive 1:38:34.720 --> 1:38:36.720 my personal drive of 1:38:36.720 --> 1:38:38.720 all the time that I spend doing 1:38:38.720 --> 1:38:40.720 what I'm doing really. 1:38:40.720 --> 1:38:42.720 Where do you see the deep learning as a field heading 1:38:42.720 --> 1:38:44.720 where do you think the next big 1:38:44.720 --> 1:38:46.720 breakthrough might be? 1:38:46.720 --> 1:38:48.720 So I think deep learning 1:38:48.720 --> 1:38:50.720 I discussed a little of this before 1:38:50.720 --> 1:38:52.720 deep learning has to be 1:38:52.720 --> 1:38:54.720 combined with some form of discretization 1:38:54.720 --> 1:38:56.720 program synthesis 1:38:56.720 --> 1:38:58.720 I think that's kind of as a research 1:38:58.720 --> 1:39:00.720 in itself is an interesting topic 1:39:00.720 --> 1:39:02.720 to expand and start doing more research 1:39:02.720 --> 1:39:04.720 and then 1:39:04.720 --> 1:39:06.720 as kind of what will deep learning 1:39:06.720 --> 1:39:08.720 enable to do in the future 1:39:08.720 --> 1:39:10.720 I don't think that's going to be what's going to happen 1:39:10.720 --> 1:39:12.720 this year but also this 1:39:12.720 --> 1:39:14.720 idea of 1:39:14.720 --> 1:39:16.720 not to throw away all the weights 1:39:16.720 --> 1:39:18.720 this idea of learning to learn 1:39:18.720 --> 1:39:20.720 and really having 1:39:20.720 --> 1:39:22.720 these agents 1:39:22.720 --> 1:39:24.720 not having to restart their weights 1:39:24.720 --> 1:39:26.720 and you can have an agent 1:39:26.720 --> 1:39:28.720 that is kind of solving 1:39:28.720 --> 1:39:30.720 or classifying images on ImageNet 1:39:30.720 --> 1:39:32.720 but also generating speech 1:39:32.720 --> 1:39:34.720 if you ask it to generate some speech 1:39:34.720 --> 1:39:36.720 and it should really be kind of 1:39:36.720 --> 1:39:38.720 almost the same 1:39:38.720 --> 1:39:40.720 network but 1:39:40.720 --> 1:39:42.720 might not be a neural network it might be a neural network 1:39:42.720 --> 1:39:44.720 with an optimization algorithm 1:39:44.720 --> 1:39:46.720 attached to it but I think this idea 1:39:46.720 --> 1:39:48.720 of generalization to new task 1:39:48.720 --> 1:39:50.720 is something that we first 1:39:50.720 --> 1:39:52.720 must define good benchmarks but then 1:39:52.720 --> 1:39:54.720 I think that's going to be exciting 1:39:54.720 --> 1:39:56.720 and I'm not sure how close we are 1:39:56.720 --> 1:39:58.720 but I think there's 1:39:58.720 --> 1:40:00.720 if you have a very limited domain 1:40:00.720 --> 1:40:02.720 I think we can start doing some progress 1:40:02.720 --> 1:40:04.720 and 1:40:04.720 --> 1:40:06.720 much like how we did a lot of programs 1:40:06.720 --> 1:40:08.720 in computer vision we should start thinking 1:40:08.720 --> 1:40:10.720 I really like a talk that 1:40:10.720 --> 1:40:12.720 Leon Boutou gave at ICML 1:40:12.720 --> 1:40:14.720 a few years ago which is 1:40:14.720 --> 1:40:16.720 this train test paradigm should be broken 1:40:16.720 --> 1:40:18.720 we should stop 1:40:18.720 --> 1:40:20.720 thinking about a training test 1:40:20.720 --> 1:40:22.720 sorry a training set and a test set 1:40:22.720 --> 1:40:24.720 and these are closed 1:40:24.720 --> 1:40:26.720 things that are untouchable 1:40:26.720 --> 1:40:28.720 I think we should go beyond these and 1:40:28.720 --> 1:40:30.720 in meta learning we call these the meta training set 1:40:30.720 --> 1:40:32.720 and the meta test set which is 1:40:32.720 --> 1:40:34.720 really thinking about 1:40:34.720 --> 1:40:36.720 if I know about ImageNet 1:40:36.720 --> 1:40:38.720 why would that network 1:40:38.720 --> 1:40:40.720 not work on MNIST which is a much 1:40:40.720 --> 1:40:42.720 simpler problem but right now it really doesn't 1:40:42.720 --> 1:40:44.720 it you know 1:40:44.720 --> 1:40:46.720 but it just feels wrong right so I think 1:40:46.720 --> 1:40:48.720 that's kind of the 1:40:48.720 --> 1:40:50.720 there's the on the application 1:40:50.720 --> 1:40:52.720 or the benchmark sites we probably 1:40:52.720 --> 1:40:54.720 will see quite a few 1:40:54.720 --> 1:40:56.720 more interest and progress and hopefully 1:40:56.720 --> 1:40:58.720 people defining new 1:40:58.720 --> 1:41:00.720 and exciting challenges really 1:41:00.720 --> 1:41:02.720 do you have any hope or 1:41:02.720 --> 1:41:04.720 interest in knowledge graphs 1:41:04.720 --> 1:41:06.720 within this context so it's kind of 1:41:06.720 --> 1:41:08.720 constructing graphs 1:41:08.720 --> 1:41:10.720 going back to graphs 1:41:10.720 --> 1:41:12.720 well neural networks are graphs but I mean 1:41:12.720 --> 1:41:14.720 a different kind of knowledge graph 1:41:14.720 --> 1:41:16.720 sort of like semantic graphs 1:41:16.720 --> 1:41:18.720 where there's concepts 1:41:18.720 --> 1:41:20.720 so I think 1:41:20.720 --> 1:41:22.720 the idea of graphs 1:41:22.720 --> 1:41:24.720 is so I've been quite interested 1:41:24.720 --> 1:41:26.720 in sequences first and then more 1:41:26.720 --> 1:41:28.720 interesting or different data structures 1:41:28.720 --> 1:41:30.720 like graphs and 1:41:30.720 --> 1:41:32.720 I've studied graph neural networks 1:41:32.720 --> 1:41:34.720 in the last three years or so 1:41:34.720 --> 1:41:36.720 I 1:41:36.720 --> 1:41:38.720 found these models just very interesting from 1:41:38.720 --> 1:41:40.720 like deep learning 1:41:40.720 --> 1:41:42.720 standpoint but then 1:41:42.720 --> 1:41:44.720 how what do we want 1:41:44.720 --> 1:41:46.720 why do we want these models and why would we 1:41:46.720 --> 1:41:48.720 use them what's the application 1:41:48.720 --> 1:41:50.720 what's kind of the killer application of graphs 1:41:50.720 --> 1:41:52.720 right and 1:41:52.720 --> 1:41:54.720 perhaps 1:41:54.720 --> 1:41:56.720 if we 1:41:56.720 --> 1:41:58.720 could extract a knowledge graph 1:41:58.720 --> 1:42:00.720 from Wikipedia automatically 1:42:00.720 --> 1:42:02.720 that would be interesting because 1:42:02.720 --> 1:42:04.720 then these graphs have 1:42:04.720 --> 1:42:06.720 this very interesting structure 1:42:06.720 --> 1:42:08.720 that also is a bit more compatible with 1:42:08.720 --> 1:42:10.720 this idea of programs and 1:42:10.720 --> 1:42:12.720 deep learning kind of working together 1:42:12.720 --> 1:42:14.720 jumping neighborhoods and so on 1:42:14.720 --> 1:42:16.720 you could imagine defining some primitives 1:42:16.720 --> 1:42:18.720 to go around graphs right so 1:42:18.720 --> 1:42:20.720 I think 1:42:20.720 --> 1:42:22.720 I really like the idea of a knowledge 1:42:22.720 --> 1:42:24.720 graph and in fact 1:42:24.720 --> 1:42:26.720 when we 1:42:26.720 --> 1:42:28.720 we started or you know 1:42:28.720 --> 1:42:30.720 as part of the research we did for StarCraft 1:42:30.720 --> 1:42:32.720 I thought wouldn't it be cool to give 1:42:32.720 --> 1:42:34.720 the graph of 1:42:34.720 --> 1:42:36.720 you know all the 1:42:36.720 --> 1:42:38.720 all these buildings that depend on each other 1:42:38.720 --> 1:42:40.720 and units that have 1:42:40.720 --> 1:42:42.720 prerequisites of being built by that and so 1:42:42.720 --> 1:42:44.720 this is information 1:42:44.720 --> 1:42:46.720 that the network can learn and extract 1:42:46.720 --> 1:42:48.720 but it would have been great to see 1:42:48.720 --> 1:42:50.720 or to think of 1:42:50.720 --> 1:42:52.720 really StarCraft as a giant graph 1:42:52.720 --> 1:42:54.720 that even also as the game evolves 1:42:54.720 --> 1:42:56.720 you kind of start taking branches 1:42:56.720 --> 1:42:58.720 and so on and we tried 1:42:58.720 --> 1:43:00.720 to do a little bit of research on this 1:43:00.720 --> 1:43:02.720 nothing too relevant 1:43:02.720 --> 1:43:04.720 but I really like the idea 1:43:04.720 --> 1:43:06.720 and it has elements that are 1:43:06.720 --> 1:43:08.720 which something you also worked with in terms of visualizing 1:43:08.720 --> 1:43:10.720 your networks as elements of 1:43:10.720 --> 1:43:12.720 having human interpretable 1:43:12.720 --> 1:43:14.720 being able to generate knowledge 1:43:14.720 --> 1:43:16.720 representations that are human interpretable 1:43:16.720 --> 1:43:18.720 that maybe human experts can then tweak 1:43:18.720 --> 1:43:20.720 or at least understand 1:43:20.720 --> 1:43:22.720 so there's a lot of interesting 1:43:22.720 --> 1:43:24.720 aspect there and for me personally I'm just a huge fan of 1:43:24.720 --> 1:43:26.720 Wikipedia and it's a shame 1:43:26.720 --> 1:43:28.720 that our neural networks 1:43:28.720 --> 1:43:30.720 aren't taking advantage of all the structured 1:43:30.720 --> 1:43:32.720 knowledge that's on the web. 1:43:32.720 --> 1:43:34.720 What's next for you? 1:43:34.720 --> 1:43:36.720 What's next for DeepMind? 1:43:36.720 --> 1:43:38.720 What are you excited about? 1:43:38.720 --> 1:43:40.720 For AlphaStar? 1:43:40.720 --> 1:43:42.720 Yeah so I think 1:43:42.720 --> 1:43:44.720 the obvious next steps 1:43:44.720 --> 1:43:46.720 would be to 1:43:46.720 --> 1:43:48.720 apply AlphaStar to 1:43:48.720 --> 1:43:50.720 other races I mean that sort of 1:43:50.720 --> 1:43:52.720 shows that the algorithm 1:43:52.720 --> 1:43:54.720 works because 1:43:54.720 --> 1:43:56.720 by mistake something in the architecture 1:43:56.720 --> 1:43:58.720 that happens to work for proto's 1:43:58.720 --> 1:44:00.720 but not for other races right so 1:44:00.720 --> 1:44:02.720 as verification I think 1:44:02.720 --> 1:44:04.720 that's an obvious next step that we are working on 1:44:04.720 --> 1:44:06.720 and 1:44:06.720 --> 1:44:08.720 then I would like to see 1:44:08.720 --> 1:44:10.720 so agents and players 1:44:10.720 --> 1:44:12.720 can specialize on 1:44:12.720 --> 1:44:14.720 different skill sets that allow them to be 1:44:14.720 --> 1:44:16.720 very good. I think we've seen 1:44:16.720 --> 1:44:18.720 AlphaStar understanding 1:44:18.720 --> 1:44:20.720 very well when to take battles and when 1:44:20.720 --> 1:44:22.720 to not do that 1:44:22.720 --> 1:44:24.720 also very good at micromanagement 1:44:24.720 --> 1:44:26.720 and moving the units around and so on 1:44:26.720 --> 1:44:28.720 and also very good at producing 1:44:28.720 --> 1:44:30.720 nonstop and trading of economy 1:44:30.720 --> 1:44:32.720 with building units 1:44:32.720 --> 1:44:34.720 but I have not 1:44:34.720 --> 1:44:36.720 perhaps seen as much as I would like 1:44:36.720 --> 1:44:38.720 this idea of the poker idea 1:44:38.720 --> 1:44:40.720 that you mentioned right. 1:44:40.720 --> 1:44:42.720 I'm not sure StarCraft or AlphaStar 1:44:42.720 --> 1:44:44.720 rather has developed a very 1:44:44.720 --> 1:44:46.720 deep understanding of 1:44:46.720 --> 1:44:48.720 what the opponent is doing 1:44:48.720 --> 1:44:50.720 and reacting to that and sort of 1:44:50.720 --> 1:44:52.720 trying to 1:44:52.720 --> 1:44:54.720 trick the player to do something else or that 1:44:54.720 --> 1:44:56.720 you know so this kind of reasoning 1:44:56.720 --> 1:44:58.720 I would like to see more so I think 1:44:58.720 --> 1:45:00.720 purely from a research standpoint 1:45:00.720 --> 1:45:02.720 there's perhaps also quite a few 1:45:02.720 --> 1:45:04.720 things to be done there 1:45:04.720 --> 1:45:06.720 in the domain of StarCraft. Yeah in the 1:45:06.720 --> 1:45:08.720 domain of games I've seen some 1:45:08.720 --> 1:45:10.720 interesting work in sort of 1:45:10.720 --> 1:45:12.720 in even auctions manipulating 1:45:12.720 --> 1:45:14.720 other players sort of forming a belief 1:45:14.720 --> 1:45:16.720 state and just messing with 1:45:16.720 --> 1:45:18.720 people. Yeah it's called theory of mind 1:45:18.720 --> 1:45:20.720 so it's a fast 1:45:20.720 --> 1:45:22.720 theory of mind on StarCraft 1:45:22.720 --> 1:45:24.720 is kind of they're really 1:45:24.720 --> 1:45:26.720 made for each other so 1:45:26.720 --> 1:45:28.720 that would be very exciting to see 1:45:28.720 --> 1:45:30.720 those techniques applied to StarCraft 1:45:30.720 --> 1:45:32.720 or perhaps StarCraft driving 1:45:32.720 --> 1:45:34.720 new techniques as I said 1:45:34.720 --> 1:45:36.720 this is always the tension between the two. 1:45:36.720 --> 1:45:38.720 Wow Oriel thank you so much for talking 1:45:38.720 --> 1:45:48.720 awesome it was great to be here thanks