1 00:00:01,935 --> 00:00:05,704 (dramatic music) 2 00:00:02,000 --> 00:00:07,000 Downloaded from YTS.MX 3 00:00:08,000 --> 00:00:13,000 Official YIFY movies site: YTS.MX 4 00:00:19,618 --> 00:00:22,252 It's intensely contemplative. 5 00:00:26,824 --> 00:00:28,724 It is almost hypnotic. 6 00:00:33,696 --> 00:00:35,530 It's like putting your hand 7 00:00:35,532 --> 00:00:38,632 on the third rail of the universe. 8 00:00:40,635 --> 00:00:43,503 If you play Go seriously, 9 00:00:43,505 --> 00:00:46,639 there is a chance that you will get exposed 10 00:00:46,641 --> 00:00:52,711 to this experience that is kind of like 11 00:00:52,713 --> 00:00:55,680 nothing else on the planet. 12 00:00:55,682 --> 00:00:59,583 Go is putting you in a place where you're always 13 00:00:59,585 --> 00:01:03,619 at the very farthest reaches of your capacity. 14 00:01:05,188 --> 00:01:07,289 There's a reason that people have been playing Go 15 00:01:07,291 --> 00:01:10,392 for thousands and thousands of years, right? 16 00:01:10,394 --> 00:01:13,561 It's not just that they wanna understand Go. 17 00:01:13,563 --> 00:01:17,498 They wanna understand what understanding is. 18 00:01:18,666 --> 00:01:22,902 And maybe that is truly what it means to be human. 19 00:01:27,707 --> 00:01:31,176 When I was a kid I loved playing games. 20 00:01:31,178 --> 00:01:34,612 I started off with board games like chess 21 00:01:34,614 --> 00:01:37,548 and then I bought my first computer when I was eight 22 00:01:37,550 --> 00:01:40,584 with winnings from a chess tournament. 23 00:01:40,586 --> 00:01:43,620 Ever since then, I felt that computers were this sort 24 00:01:43,622 --> 00:01:47,657 of magical device that could extend the power of your mind. 25 00:01:49,693 --> 00:01:52,327 Virtual environments in games. 26 00:01:52,329 --> 00:01:53,728 We think they're the perfect platform 27 00:01:53,730 --> 00:01:56,697 for developing and testing AI algorithms. 28 00:01:56,699 --> 00:02:00,367 Games are very convenient, in that a lot of them have scores 29 00:02:00,369 --> 00:02:03,603 so it's very easy to measure incremental progress. 30 00:02:03,605 --> 00:02:07,706 I'm gonna show you a few videos of the A-Jen system, the AI. 31 00:02:07,708 --> 00:02:09,641 Let's start off with Break Out. 32 00:02:09,643 --> 00:02:12,377 So here you control the bat and ball and you're trying 33 00:02:12,379 --> 00:02:14,712 to break through this rainbow-colored wall. 34 00:02:14,714 --> 00:02:17,548 The A-Jen system has to learn everything for itself, 35 00:02:17,550 --> 00:02:18,815 just from the raw pixels, 36 00:02:18,817 --> 00:02:20,750 doesn't know what it's controlling, 37 00:02:20,752 --> 00:02:23,552 doesn't even know what the object of the game is. 38 00:02:23,554 --> 00:02:25,754 Now, the beginning, after a hundred games, 39 00:02:25,756 --> 00:02:27,622 you can see the A-Jen is not very good. 40 00:02:27,624 --> 00:02:29,423 It's missing the ball most of the time, 41 00:02:29,425 --> 00:02:31,358 but it's starting to get the hang of the idea 42 00:02:31,360 --> 00:02:33,760 that the bat should go towards the ball. 43 00:02:33,762 --> 00:02:36,262 Now, after 300 games, it's about as good 44 00:02:36,264 --> 00:02:38,631 as any human can play this and pretty much 45 00:02:38,633 --> 00:02:40,532 gets the ball back every time. 46 00:02:40,534 --> 00:02:41,532 We thought, well, that's pretty cool, 47 00:02:41,534 --> 00:02:44,502 but we left the system playing for another 200 games. 48 00:02:44,504 --> 00:02:45,502 And it did this amazing thing. 49 00:02:45,504 --> 00:02:47,872 It found the optimal strategy was to dig a tunnel 50 00:02:47,874 --> 00:02:49,740 around the side and put the ball 51 00:02:49,742 --> 00:02:51,675 around the back of the wall. 52 00:02:51,677 --> 00:02:54,544 The researchers working on this are amazing AI developers, 53 00:02:54,546 --> 00:02:55,978 but they're not so good at Break Out 54 00:02:55,980 --> 00:02:57,813 and they didn't know about that strategy, 55 00:02:57,815 --> 00:02:59,748 so they learned something from their own system, 56 00:02:59,750 --> 00:03:02,617 which is pretty funny and quite instructive, I think, 57 00:03:02,619 --> 00:03:05,553 about the potential for general AI. 58 00:03:05,555 --> 00:03:07,655 So, for us, what's the next step now? 59 00:03:07,657 --> 00:03:11,458 Go is the most complex game pretty much ever devised by man. 60 00:03:11,460 --> 00:03:13,359 Beating a professional player at Go 61 00:03:13,361 --> 00:03:17,596 is a longstanding, grand challenge of AI research. 62 00:03:23,669 --> 00:03:26,404 (bells tolling) 63 00:03:37,782 --> 00:03:38,882 Wow. 64 00:04:40,978 --> 00:04:44,646 Dear Mr Fan, my name is Demis Hassabis. 65 00:04:45,781 --> 00:04:47,348 I run an artificial intelligence company 66 00:04:47,350 --> 00:04:50,017 based in London called DeepMind. 67 00:04:50,019 --> 00:04:52,119 As the strongest Go player in Europe, 68 00:04:52,121 --> 00:04:54,921 we would like to invite you to our offices in London 69 00:04:54,923 --> 00:04:57,390 both to meet you in person and to share with you 70 00:04:57,392 --> 00:05:00,826 an exciting Go project that we're working on. 71 00:05:00,828 --> 00:05:02,627 If you would be interested in coming to visit us, 72 00:05:02,629 --> 00:05:04,528 please let us know. 73 00:05:04,530 --> 00:05:07,130 Many thanks, kind regards, Demis. 74 00:05:49,839 --> 00:05:51,873 We needed to get Fan Hui to DeepMind 75 00:05:51,875 --> 00:05:53,941 to see we were a serious operation 76 00:05:53,943 --> 00:05:58,077 and we were serious people doing proper research. 77 00:05:59,646 --> 00:06:00,979 And the search time is getting better and better 78 00:06:00,981 --> 00:06:02,814 as we go from, for example, one to eight 79 00:06:02,816 --> 00:06:03,815 and any one of these. 80 00:06:03,817 --> 00:06:05,983 It always goes up. 81 00:06:05,985 --> 00:06:08,485 We think of DeepMind as kind of like 82 00:06:08,487 --> 00:06:11,921 an Apollo program effort for AI. 83 00:06:11,923 --> 00:06:14,990 Our mission is to fundamentally understand intelligence 84 00:06:14,992 --> 00:06:17,092 and recreate it artificially. 85 00:06:17,094 --> 00:06:19,427 And then, once we've done that, we feel that we can use 86 00:06:19,429 --> 00:06:21,896 that technology to help society 87 00:06:21,898 --> 00:06:23,898 solve all sorts of other problems. 88 00:06:23,900 --> 00:06:25,566 If you search through the actual game 89 00:06:25,568 --> 00:06:28,101 we can see kind of how an algorithm thinks. 90 00:06:28,103 --> 00:06:31,471 What's the most likely variation that it thinks will happen. 91 00:06:31,473 --> 00:06:34,006 [Demis] We've been working on AlphaGo and our program 92 00:06:34,008 --> 00:06:37,075 to play Go for just under two years now. 93 00:06:37,077 --> 00:06:39,110 All the little patterns cascade together, 94 00:06:39,112 --> 00:06:42,213 layer after layer after layer after layer. 95 00:06:42,215 --> 00:06:44,014 I started talking about the game of Go 96 00:06:44,016 --> 00:06:46,783 with Demis more than 20 years ago. 97 00:06:46,785 --> 00:06:49,218 And so, this has been a really long journey. 98 00:06:49,220 --> 00:06:52,788 The game of Go is the holy grail of artificial intelligence. 99 00:06:52,790 --> 00:06:54,890 For many years, people have looked at this game 100 00:06:54,892 --> 00:06:57,893 and they thought, "Wow, this is just too hard." 101 00:06:57,895 --> 00:06:59,928 Everything we've ever tried in AI, 102 00:06:59,930 --> 00:07:01,996 it just falls over when you try the game of Go 103 00:07:01,998 --> 00:07:03,797 and so that's why it feels like a real 104 00:07:03,799 --> 00:07:05,598 litmus test of progress. 105 00:07:05,600 --> 00:07:10,535 If we can crack Go, we know we've done something special. 106 00:07:10,537 --> 00:07:12,570 So, with Fan Hui, we started talking around 107 00:07:12,572 --> 00:07:14,572 what the real purpose of the visit was. 108 00:07:14,574 --> 00:07:17,107 It wasn't just a Go project we wanted him to help with. 109 00:07:17,109 --> 00:07:19,042 That actually we wanted to play him 110 00:07:19,044 --> 00:07:20,777 and we had a very strong program. 111 00:07:25,983 --> 00:07:28,718 A lot of people have thought that it was decades away. 112 00:07:28,720 --> 00:07:31,487 Some people thought it would be never because they felt 113 00:07:31,489 --> 00:07:35,624 that to succeed at Go you needed human intuition. 114 00:07:36,993 --> 00:07:40,929 Go is the world's oldest, continuously played board game 115 00:07:40,931 --> 00:07:44,932 and, in some sense, it is one of the simplest 116 00:07:44,934 --> 00:07:47,668 and also most abstract. 117 00:07:47,670 --> 00:07:48,802 There's only one type of piece. 118 00:07:48,804 --> 00:07:50,036 There's only one type of move. 119 00:07:50,038 --> 00:07:51,904 You just place that piece on the board 120 00:07:51,906 --> 00:07:56,041 and then your goal is to create a linked group 121 00:07:56,043 --> 00:07:59,744 of your stones that surrounds some empty territory. 122 00:07:59,746 --> 00:08:03,247 And, when you surround enemy stones, you capture them 123 00:08:03,249 --> 00:08:05,882 and remove them from the board. 124 00:08:05,884 --> 00:08:08,050 You earn points by surrounding territory 125 00:08:08,052 --> 00:08:10,819 and, at the end of the game, the person 126 00:08:10,821 --> 00:08:13,688 with the most territory wins. 127 00:08:13,690 --> 00:08:17,258 It seems really simple, but then you sit down to play 128 00:08:17,260 --> 00:08:18,959 and you realize right away, it's like, well, 129 00:08:18,961 --> 00:08:21,761 I technically know what I'm allowed to do, 130 00:08:21,763 --> 00:08:23,729 but I have no clue what I should do. 131 00:08:23,731 --> 00:08:26,131 Go's incredibly challenging for computers to tackle 132 00:08:26,133 --> 00:08:28,533 because compared, to say, chess, 133 00:08:28,535 --> 00:08:31,736 the number of possible moves in a position is much larger. 134 00:08:31,738 --> 00:08:33,003 In chess it's about 20. 135 00:08:33,005 --> 00:08:34,938 In Go it's about 200. 136 00:08:34,940 --> 00:08:38,007 And the number of possible configurations of the board 137 00:08:38,009 --> 00:08:40,642 is more than the number of atoms in the universe. 138 00:08:40,644 --> 00:08:42,977 So, even if you took all the computers in the world 139 00:08:42,979 --> 00:08:45,813 and ran them for a million years, that wouldn't be enough 140 00:08:45,815 --> 00:08:49,249 compute power to calculate all the possible variations. 141 00:08:49,251 --> 00:08:51,784 If you ask a great Go player why they played 142 00:08:51,786 --> 00:08:54,119 a particular move, sometimes they'll just tell you 143 00:08:54,121 --> 00:08:55,787 it felt right. 144 00:08:55,789 --> 00:08:58,122 So we have to come up with some kind of clever algorithm 145 00:08:58,124 --> 00:09:02,192 to mimic what people do with their intuition. 146 00:09:02,194 --> 00:09:03,826 - Nice to meet you. - Oh, hello. 147 00:09:03,828 --> 00:09:04,826 Nice meet you too. 148 00:09:04,828 --> 00:09:07,663 [Demis] So with Fan Hui, we agreed a best of five match 149 00:09:07,665 --> 00:09:09,831 and we agreed it would be filmed. 150 00:09:09,833 --> 00:09:13,835 You know, we would treat it as a serious match. 151 00:09:40,829 --> 00:09:43,697 In that first match, I think something clicked for him 152 00:09:43,699 --> 00:09:47,300 that this wasn't an ordinary Go program. 153 00:09:47,302 --> 00:09:50,269 We weren't just doing the same as everyone else, 154 00:09:50,271 --> 00:09:53,105 that something new was happening. 155 00:10:11,190 --> 00:10:14,025 After losing, losing, losing, 156 00:10:14,027 --> 00:10:17,762 you can feel his pressure is getting heavier and heavier. 157 00:10:17,764 --> 00:10:20,331 And several times after the game he said, 158 00:10:20,333 --> 00:10:24,201 he wanted to go out for fresh air. 159 00:10:24,203 --> 00:10:26,269 I said, "Oh, I can go with you and have a chat." 160 00:10:26,271 --> 00:10:29,105 He said, "No, I want to go myself." 161 00:10:44,120 --> 00:10:46,654 I worried whether he'd even come back. 162 00:10:46,656 --> 00:10:49,923 You know, he seemed very low and he spent 163 00:10:49,925 --> 00:10:51,691 about an hour away from the office. 164 00:10:51,693 --> 00:10:54,226 He came back completely changed. 165 00:11:22,855 --> 00:11:25,389 Artificial intelligence researchers have solved 166 00:11:25,391 --> 00:11:29,259 the game of Go a decade earlier than expected. 167 00:11:29,261 --> 00:11:31,361 The computer named AlphaGo was able 168 00:11:31,363 --> 00:11:33,996 to beat the European human champion. 169 00:11:33,998 --> 00:11:35,464 [Reporter] Artificial intelligence researchers 170 00:11:35,466 --> 00:11:37,699 have made a significant breakthrough. 171 00:11:37,701 --> 00:11:40,101 It really is a big leap forward. 172 00:11:40,103 --> 00:11:43,304 There's a big difference between the way the IBM computer 173 00:11:43,306 --> 00:11:47,474 beat Kasparov, which was programed by expert chess players, 174 00:11:47,476 --> 00:11:50,109 and the way the Go-playing computer 175 00:11:50,111 --> 00:11:51,710 more or less learned itself. 176 00:11:51,712 --> 00:11:54,946 The way we start off training AlphaGo is by showing it 177 00:11:54,948 --> 00:11:58,115 a 100,000 games that strong amateurs have played 178 00:11:58,117 --> 00:12:00,117 that we downloaded from the internet. 179 00:12:00,119 --> 00:12:02,486 And we first, initially, get AlphaGo to mimic 180 00:12:02,488 --> 00:12:04,588 the human player and then, through self-playing 181 00:12:04,590 --> 00:12:07,324 reinforcement learning, it plays against different versions 182 00:12:07,326 --> 00:12:10,860 of itself many millions of times and learns from its errors. 183 00:12:10,862 --> 00:12:14,530 These specific ideas that are driving AlphaGo 184 00:12:14,532 --> 00:12:16,899 are gonna drive our future. 185 00:12:16,901 --> 00:12:19,368 The technologies at the heart of AlphaGo, 186 00:12:19,370 --> 00:12:21,536 they're what are called deep neural networks 187 00:12:21,538 --> 00:12:24,272 which, essentially, mimic the web 188 00:12:24,274 --> 00:12:26,040 of neurons in the human brain. 189 00:12:26,042 --> 00:12:29,343 It's a very old idea, but recently, due to increases 190 00:12:29,345 --> 00:12:33,113 in computing power, these neural networks have become 191 00:12:33,115 --> 00:12:35,148 extremely powerful, almost overnight. 192 00:12:35,150 --> 00:12:38,284 Big neural networks that operate on big data 193 00:12:38,286 --> 00:12:40,319 can achieve surprising things. 194 00:12:40,321 --> 00:12:42,888 AlphaGo found a way how to learn how to play Go. 195 00:12:42,890 --> 00:12:44,256 Learning is the key thing here. 196 00:12:44,258 --> 00:12:45,924 It's machine learning. 197 00:12:45,926 --> 00:12:47,859 The whole beauty of these types of algorithms 198 00:12:47,861 --> 00:12:49,627 is that, because they're learning for themselves, 199 00:12:49,629 --> 00:12:52,363 they can go beyond what we as the programmers 200 00:12:52,365 --> 00:12:55,265 know how to do and allow us to make new breakthroughs 201 00:12:55,267 --> 00:12:57,533 in areas of science and medicine. 202 00:12:57,535 --> 00:13:00,335 So AlphaGo is one significant step 203 00:13:00,337 --> 00:13:02,603 towards that ultimate goal. 204 00:13:35,102 --> 00:13:37,636 The Go world were skeptical about how strong 205 00:13:37,638 --> 00:13:41,006 really was AlphaGo and how much further did it need 206 00:13:41,008 --> 00:13:43,475 to get to beat the top professionals. 207 00:13:43,477 --> 00:13:47,345 Our program is improving over time and we want to push 208 00:13:47,347 --> 00:13:50,181 the AI algorithm to the limit and see how far 209 00:13:50,183 --> 00:13:53,384 this kind of self-improving process can go. 210 00:13:53,386 --> 00:13:57,388 So we needed to look for an even greater challenge. 211 00:13:57,390 --> 00:13:59,323 [Reporter] A match like no other is about 212 00:13:59,325 --> 00:14:01,258 to get underway in South Korea. 213 00:14:01,260 --> 00:14:03,560 [Reporter] Lee Sedol, the long-reigning global champ. 214 00:14:03,562 --> 00:14:05,228 [Reporter] This guy is a genius. 215 00:14:05,230 --> 00:14:07,396 [Reporter] He will take on artificial intelligence program 216 00:14:07,398 --> 00:14:11,599 AlphaGo, in the ultimate human versus machine smack down. 217 00:14:11,601 --> 00:14:13,701 This is a huge moment for both the world 218 00:14:13,703 --> 00:14:17,671 of artificial intelligence and, I think, the world of Go. 219 00:14:17,673 --> 00:14:21,441 So far, AlphaGo has beaten every challenge we've given it, 220 00:14:21,443 --> 00:14:24,210 but we won't know its true strength until we play 221 00:14:24,212 --> 00:14:27,713 somebody who is at the top of the world like Lee Sedol. 222 00:14:27,715 --> 00:14:30,482 We chose Lee Sedol because we wanted 223 00:14:30,484 --> 00:14:32,984 a legendary, historic player. 224 00:14:32,986 --> 00:14:34,552 Somebody who has been acknowledged 225 00:14:34,554 --> 00:14:38,322 as the greatest player of the last decade. 226 00:14:43,294 --> 00:14:45,361 [Translator] I will not be too arrogant, 227 00:14:45,363 --> 00:14:48,664 but I don't think that it will be a very close match. 228 00:14:48,666 --> 00:14:52,367 The level of the player that AlphaGo went against 229 00:14:52,369 --> 00:14:54,736 in October is not the same level as me. 230 00:14:54,738 --> 00:14:58,406 So, given that a couple of months was only passed, 231 00:14:58,408 --> 00:15:00,608 I don't think that it is enough for it 232 00:15:00,610 --> 00:15:02,343 to be able to catch up with me. 233 00:15:02,345 --> 00:15:05,512 My hope is that it will be either five-zero for me 234 00:15:05,514 --> 00:15:07,280 or maybe four to one. 235 00:15:07,282 --> 00:15:08,748 So the critical point for me 236 00:15:08,750 --> 00:15:11,684 is to make sure I do not lose one. 237 00:15:11,686 --> 00:15:14,353 (speaking in foreign language) 238 00:15:14,355 --> 00:15:17,322 We don't know how well our system will play 239 00:15:17,324 --> 00:15:20,491 against someone as creative as Lee Sedol. 240 00:15:20,493 --> 00:15:24,761 Also, he's very famous for very creative fighting play 241 00:15:24,763 --> 00:15:28,364 so this could be difficult for us, but we'll see. 242 00:15:28,366 --> 00:15:29,565 Maybe in some ways he's the most 243 00:15:29,567 --> 00:15:32,134 difficult opponent we could pick. 244 00:15:32,136 --> 00:15:35,537 There was still so much of a question about whether or not 245 00:15:35,539 --> 00:15:39,007 they could beat someone like Lee Sedol. 246 00:15:39,009 --> 00:15:41,442 Fan Hui is a good player, but he's nothing 247 00:15:41,444 --> 00:15:43,544 like the very top players. 248 00:15:43,546 --> 00:15:46,146 Lee Sedol is a nine dan professional. 249 00:15:46,148 --> 00:15:47,514 Nine dan professional is the highest rank 250 00:15:47,516 --> 00:15:51,217 you can achieve in Go, the ranking system. 251 00:15:51,219 --> 00:15:53,152 Fan Hui, who's the European champion, 252 00:15:53,154 --> 00:15:55,320 is just a two dan professional. 253 00:15:55,322 --> 00:15:57,755 On the other hand, at the very top, Lee Sedol 254 00:15:57,757 --> 00:15:59,757 is nine dan professional. 255 00:15:59,759 --> 00:16:02,593 Lee Sedol is to Go what Roger Federer is to tennis. 256 00:16:02,595 --> 00:16:04,595 He's playing in Wimbledon to win the Grand Slam 257 00:16:04,597 --> 00:16:06,063 and it's not just this year. 258 00:16:06,065 --> 00:16:08,031 He'll be there next year and the year after that 259 00:16:08,033 --> 00:16:10,099 and the year after that. 260 00:16:24,447 --> 00:16:27,682 (children chattering) 261 00:16:38,627 --> 00:16:40,628 Go is taken seriously in Korea. 262 00:16:40,630 --> 00:16:43,430 It's so much part of the culture like breathing 263 00:16:43,432 --> 00:16:47,667 or, you know, like taking swimming lessons or something. 264 00:17:01,381 --> 00:17:04,549 In ancient China, Japan, Korea, 265 00:17:04,551 --> 00:17:08,586 Go is one of the four noble things for accomplishment 266 00:17:10,322 --> 00:17:14,124 for literati with music, poetry and painting. 267 00:17:15,726 --> 00:17:18,494 So people think the Go players 268 00:17:19,629 --> 00:17:22,197 are very smart and very noble. 269 00:17:27,669 --> 00:17:31,138 Mr Kwon started this school so that he could 270 00:17:31,140 --> 00:17:34,274 produce great player in Korea. 271 00:17:34,276 --> 00:17:38,244 Lee Sedol, when he was age of eight, he came in here 272 00:17:38,246 --> 00:17:42,681 attending class nine a.m until nine p.m. seven days. 273 00:17:42,683 --> 00:17:45,250 And then he stay with Mr Kwon. 274 00:18:03,569 --> 00:18:07,271 Lee Sedol was a little boy that I remember 275 00:18:07,273 --> 00:18:09,539 when I was a student studying Go. 276 00:18:09,541 --> 00:18:12,642 He was very young and rural enough 277 00:18:12,644 --> 00:18:15,745 to think that pizza grew on trees. 278 00:18:37,534 --> 00:18:41,169 Lee Sedol plays things that are interesting 279 00:18:41,171 --> 00:18:43,638 where you felt like he's beyond winning and losing. 280 00:18:43,640 --> 00:18:45,840 He wants to do something that's innovative 281 00:18:45,842 --> 00:18:47,708 or takes things to the next level. 282 00:18:47,710 --> 00:18:51,244 So every Go player study his game, for sure. 283 00:18:51,246 --> 00:18:54,747 Roughly, 10 years, he dominates the professional Go world. 284 00:18:54,749 --> 00:18:57,316 He win 18 World Championships. 285 00:19:10,596 --> 00:19:14,798 Definitely, Lee Sedol is gonna win the game, five all. 286 00:19:14,800 --> 00:19:16,866 Yeah, all game. 287 00:19:16,868 --> 00:19:19,468 He's gonna win all game. 288 00:19:19,470 --> 00:19:22,571 We had our evaluation match last week. 289 00:19:22,573 --> 00:19:24,506 We won a game and we lost a game. 290 00:19:24,508 --> 00:19:26,341 And we lost a game in a way that would have made us 291 00:19:26,343 --> 00:19:29,677 look extremely foolish if that happened publicly. 292 00:19:29,679 --> 00:19:31,712 It means that we still have work to do 293 00:19:31,714 --> 00:19:33,680 and we need to take this really seriously. 294 00:19:33,682 --> 00:19:37,716 It's just too much risk that actually we could lose overall. 295 00:19:38,818 --> 00:19:40,785 Not only that, but we could lose in a way 296 00:19:40,787 --> 00:19:42,720 that makes us look rather silly. 297 00:19:42,722 --> 00:19:43,854 So I guess... 298 00:19:44,755 --> 00:19:45,888 Yeah, Aja, did you want to say anything 299 00:19:45,890 --> 00:19:46,989 about what you're trying? 300 00:19:46,991 --> 00:19:48,690 Um, I'm working hard. 301 00:19:48,692 --> 00:19:52,293 (group laughing) 302 00:19:52,295 --> 00:19:54,228 [Demis] So we're working around the clock at the moment, 303 00:19:54,230 --> 00:19:55,829 training our algorithms further, 304 00:19:55,831 --> 00:19:57,897 trying to incrementally keep on improving, 305 00:19:57,899 --> 00:20:00,532 right up to the moment where we have the match. 306 00:20:00,534 --> 00:20:01,933 We've collected together people 307 00:20:01,935 --> 00:20:04,568 with different skills sets onto the same team. 308 00:20:04,570 --> 00:20:08,805 So we have researchers, engineers, valuation guys. 309 00:20:13,810 --> 00:20:15,510 [Man] We don't see it. 310 00:20:15,512 --> 00:20:16,410 Okay. 311 00:20:20,648 --> 00:20:23,650 [David] Aja is the lead programmer and built 312 00:20:23,652 --> 00:20:25,551 the original search engine. 313 00:20:25,553 --> 00:20:27,452 [Demis] So Aja's responsibility is quite a big one. 314 00:20:27,454 --> 00:20:29,954 He'll be the one sitting opposite Lee Sedol 315 00:20:29,956 --> 00:20:33,757 and actually playing the moves AlphaGo makes. 316 00:20:33,759 --> 00:20:35,658 I'm feeling excited. 317 00:20:37,460 --> 00:20:39,761 Lee Sedol, he's a great player 318 00:20:39,763 --> 00:20:42,830 and I feel honored to play with him. 319 00:20:44,032 --> 00:20:46,533 Many of my friends, they are very excited about the match. 320 00:20:46,535 --> 00:20:49,802 They keep telling me that the whole world is watching. 321 00:20:49,804 --> 00:20:51,837 Just prepare AlphaGo. 322 00:20:52,705 --> 00:20:53,671 Hello. 323 00:20:53,673 --> 00:20:55,039 Hello, Demis, can you hear us? 324 00:20:55,041 --> 00:20:56,540 Yes, I can hear you guys. 325 00:20:56,542 --> 00:20:58,041 Can you hear me okay? 326 00:20:58,043 --> 00:20:59,809 Yes, perfect. 327 00:20:59,811 --> 00:21:03,612 I was curious to meet you, such an amazing developer 328 00:21:03,614 --> 00:21:05,714 who made AlphaGo. 329 00:21:05,716 --> 00:21:07,282 It's very nice to meet you. 330 00:21:07,284 --> 00:21:08,016 Thank you, thank you. 331 00:21:08,018 --> 00:21:09,817 So likewise. 332 00:21:09,819 --> 00:21:12,386 It's a real pleasure to meet you. 333 00:21:12,388 --> 00:21:13,820 I love the game of Go. 334 00:21:13,822 --> 00:21:17,790 I'm not very good, but I don't know if you know, 335 00:21:17,792 --> 00:21:19,992 but we actually have something in common. 336 00:21:19,994 --> 00:21:23,595 I actually also trained at a game at a young age. 337 00:21:23,597 --> 00:21:26,731 I used to be, when I was young, a professional chess player. 338 00:21:26,733 --> 00:21:29,800 (speaking in foreign language) 339 00:21:29,802 --> 00:21:32,869 I used to play for the England team 340 00:21:32,871 --> 00:21:36,905 and then I stopped playing when I was about 14. 341 00:21:36,907 --> 00:21:38,873 When I was 13, I was the second highest 342 00:21:38,875 --> 00:21:40,774 rated player in the world. 343 00:21:40,776 --> 00:21:43,009 [Translator] After the match, I would like to propose 344 00:21:43,011 --> 00:21:45,711 that I give you one teaching game of Go 345 00:21:45,713 --> 00:21:48,680 and you give me one teaching game of chess. 346 00:21:48,682 --> 00:21:49,747 How does that sound? 347 00:21:49,749 --> 00:21:50,981 That sounds very good. 348 00:21:50,983 --> 00:21:53,583 I would be very honored to do that. 349 00:21:53,585 --> 00:21:57,486 I'm not sure it's okay to ask you this question, 350 00:21:57,488 --> 00:22:01,089 but I saw the games against Fan Hui and I didn't think 351 00:22:01,091 --> 00:22:04,425 it was quite at the level to play with me. 352 00:22:04,427 --> 00:22:07,594 But I heard that it's getting really stronger. 353 00:22:07,596 --> 00:22:10,697 Can I ask how much it got stronger since then? 354 00:22:10,699 --> 00:22:13,599 (Demis laughing) 355 00:22:13,601 --> 00:22:17,502 Um, uh, I can't say too much, of course, 356 00:22:17,504 --> 00:22:21,505 but it's definitely got significantly stronger. 357 00:22:22,807 --> 00:22:25,709 With more time, I think a logical approach 358 00:22:25,711 --> 00:22:29,012 would be to follow up on Aja's local search. 359 00:22:29,014 --> 00:22:31,848 Run that in AlphaGo, generate a whole new data. 360 00:22:31,850 --> 00:22:33,049 [Man] Exactly. 361 00:22:34,117 --> 00:22:35,050 But we're just out of time for that 362 00:22:35,052 --> 00:22:36,951 so we have to be realistic. 363 00:22:36,953 --> 00:22:39,720 We realized that if we wanted to prepare 364 00:22:39,722 --> 00:22:42,956 for a mammoth task like taking on Lee Sedol, 365 00:22:42,958 --> 00:22:46,526 there would be nothing better than talking to a professional 366 00:22:46,528 --> 00:22:50,762 and we couldn't have picked a better person that Fan Hui. 367 00:22:52,731 --> 00:22:55,232 [Aja] We invite him as an advisor 368 00:22:55,234 --> 00:22:57,100 because, during the match, we found 369 00:22:57,102 --> 00:22:59,202 he is a man of good spirit. 370 00:22:59,204 --> 00:23:00,136 [Woman] Welcome back. 371 00:23:00,138 --> 00:23:01,837 Thank you very much. 372 00:23:01,839 --> 00:23:05,140 It's very precious that we have Fan Hui with us. 373 00:23:06,709 --> 00:23:09,977 He was crushed by AlphaGo, but then he became very positive 374 00:23:09,979 --> 00:23:12,012 and a big help to us. 375 00:23:49,983 --> 00:23:53,518 The superheroes always have a hidden vulnerability, right. 376 00:23:53,520 --> 00:23:55,887 And the same is true of AlphaGo. 377 00:23:55,889 --> 00:23:59,089 It's unbelievably superhuman in general, 378 00:24:00,658 --> 00:24:04,060 but it has some particular weaknesses in some situations. 379 00:24:04,062 --> 00:24:05,795 We can think of that being the space 380 00:24:05,797 --> 00:24:07,229 of all the things it knows about. 381 00:24:07,231 --> 00:24:09,664 And it knows about most of it extremely well, 382 00:24:09,666 --> 00:24:11,966 but then there'll be these tricky lumps of knowledge 383 00:24:11,968 --> 00:24:14,969 that it just understands very poorly 384 00:24:14,971 --> 00:24:17,004 and it's really hard for us to characterize 385 00:24:17,006 --> 00:24:19,072 when it's going to enter into one of these lumps. 386 00:24:19,074 --> 00:24:21,307 But, if it does, it could be completely delusional 387 00:24:21,309 --> 00:24:23,208 thinking it's alive on one part of the board 388 00:24:23,210 --> 00:24:25,977 when, in fact, it's dead or vice-versa. 389 00:24:25,979 --> 00:24:28,779 So there is a real risk that we could lose the match. 390 00:24:39,089 --> 00:24:41,990 [Woman] We've got version 18 in the pipeline, haven't we? 391 00:24:41,992 --> 00:24:43,591 Well, we'll only go for version 18 392 00:24:43,593 --> 00:24:45,159 if it makes a significant improvement 393 00:24:45,161 --> 00:24:47,961 which we're having to rerun all the tests 394 00:24:47,963 --> 00:24:49,996 cuz they went wrong unfortunately. 395 00:24:49,998 --> 00:24:52,131 So we're basically a day behind 396 00:24:52,133 --> 00:24:54,633 in our evaluation because of that. 397 00:24:54,635 --> 00:24:56,067 Well, maybe we just need to be realistic 398 00:24:56,069 --> 00:24:58,135 that we've tried a bunch of things 399 00:24:58,137 --> 00:24:59,535 and we've come to the point where we said 400 00:24:59,537 --> 00:25:01,571 we would actually start freezing the code 401 00:25:01,573 --> 00:25:05,641 and saying, "Look, it's actually not panning out." 402 00:25:08,044 --> 00:25:10,078 These delusions are still a realistic 403 00:25:10,080 --> 00:25:12,046 possibility for the match. 404 00:25:12,048 --> 00:25:15,049 We have some weaknesses that I don't think 405 00:25:15,051 --> 00:25:18,052 we're going to fix fully before the match 406 00:25:18,054 --> 00:25:21,921 so that's causing us a little bit of anxiety. 407 00:25:21,923 --> 00:25:24,590 (upbeat music) 408 00:25:27,160 --> 00:25:30,061 [Reporter] Lee Sedol is getting ready to rumble. 409 00:25:30,063 --> 00:25:32,229 On Wednesday live across the internet, 410 00:25:32,231 --> 00:25:34,731 this professional South Korean Go player 411 00:25:34,733 --> 00:25:38,968 will take on artificial intelligence program AlphaGo. 412 00:25:44,206 --> 00:25:46,140 [David] Tomorrow, Julian and George, 413 00:25:46,142 --> 00:25:48,909 they pack up version 18, they stick it on a laptop 414 00:25:48,911 --> 00:25:51,177 and they fly out to Seoul. 415 00:25:51,179 --> 00:25:54,180 (dramatic music) 416 00:25:54,481 --> 00:25:57,516 (overlapping chatter) 417 00:26:08,994 --> 00:26:12,229 Is everything all set for the victory against Lee Sedol? 418 00:26:12,231 --> 00:26:15,098 Everything is set, I think, but when I get to the hotel, 419 00:26:15,100 --> 00:26:16,799 I'm gonna catch up with the team. 420 00:26:16,801 --> 00:26:19,635 What we should do in the time remaining is list things 421 00:26:19,637 --> 00:26:22,070 that could go wrong in the solidity of the system. 422 00:26:26,108 --> 00:26:28,809 [Woman] All right, rehearsal, let's go, please folks. 423 00:26:34,148 --> 00:26:36,883 Yes, there is a terrace and we will have security 424 00:26:36,885 --> 00:26:39,218 so he'll be able to go up there and be by himself. 425 00:26:50,763 --> 00:26:53,364 When I got here I didn't expect the attention on the man. 426 00:26:53,366 --> 00:26:55,766 It was literally front page news. 427 00:26:55,768 --> 00:26:58,769 About eight million Koreans play the game of Go 428 00:26:58,771 --> 00:27:01,138 and, even those who don't recognize Lee Sedol, 429 00:27:01,140 --> 00:27:03,373 he's a national figure. 430 00:27:03,375 --> 00:27:04,874 And so there's that, right. 431 00:27:04,876 --> 00:27:06,275 There's some national pride involved, 432 00:27:06,277 --> 00:27:08,076 but it's more than that. 433 00:27:08,078 --> 00:27:11,345 Just the very thought of a machine playing a human 434 00:27:11,347 --> 00:27:13,447 at something like this, I think is inherently 435 00:27:13,449 --> 00:27:15,248 intriguing to people. 436 00:27:18,218 --> 00:27:21,220 (crowd chattering) 437 00:27:21,222 --> 00:27:24,223 (cameras clicking) 438 00:28:08,768 --> 00:28:11,069 [David] I just really hope we win this first game. 439 00:28:11,071 --> 00:28:12,270 If you lose the first game, you literally 440 00:28:12,272 --> 00:28:14,505 have to win three out of four 441 00:28:14,507 --> 00:28:16,139 which is hard work. 442 00:28:16,141 --> 00:28:17,507 [Demis] Where are you gonna be? 443 00:28:17,509 --> 00:28:18,774 I'll start in the match room 444 00:28:18,776 --> 00:28:19,508 and then I'm gonna come in here, 445 00:28:19,510 --> 00:28:20,909 make sure everything's sound. 446 00:28:20,911 --> 00:28:23,944 Yeah, you should probably be in there. 447 00:28:26,447 --> 00:28:27,814 I know you're nervous. 448 00:28:27,816 --> 00:28:29,515 Yeah, I'm nervous. 449 00:28:29,517 --> 00:28:30,782 I'm nervous. 450 00:28:32,217 --> 00:28:33,050 How are you? 451 00:28:33,052 --> 00:28:34,084 You're not nervous? 452 00:28:34,086 --> 00:28:35,184 - A bit. - A little bit, yeah? 453 00:28:35,186 --> 00:28:39,021 It's good to have a little bit of nerves. 454 00:28:39,023 --> 00:28:40,255 It'll be fine. 455 00:28:42,157 --> 00:28:44,024 Be fine, be fine. 456 00:28:44,026 --> 00:28:47,260 Hello and welcome to the DeepMind challenge, 457 00:28:47,262 --> 00:28:50,830 game one, round one, live from the Four Seasons 458 00:28:50,832 --> 00:28:53,265 here in Seoul, Korea. 459 00:28:53,267 --> 00:28:56,167 I am Chris Garlock of the American Go E-Journal. 460 00:28:56,169 --> 00:28:59,203 I'm here with Michael Redman nine dan professional. 461 00:28:59,205 --> 00:29:00,337 Welcome, Michael. 462 00:29:00,339 --> 00:29:02,072 Want to give a shout out to all 463 00:29:02,074 --> 00:29:04,374 the folks watching around the world. 464 00:29:04,376 --> 00:29:06,175 [Man] Well, the excitement is pretty palpable 465 00:29:06,177 --> 00:29:07,075 here in the hotel. 466 00:29:07,077 --> 00:29:10,245 I've never seen a crush of interested reporters 467 00:29:10,247 --> 00:29:12,146 from around the world. 468 00:29:18,786 --> 00:29:20,052 [Announcer] All right, folks, 469 00:29:20,054 --> 00:29:21,053 you're here, you're gonna see history made. 470 00:29:21,055 --> 00:29:22,154 Stay with us. 471 00:29:22,156 --> 00:29:23,922 [Man] Five minutes, guys. 472 00:29:23,924 --> 00:29:25,056 Five minutes. 473 00:29:33,498 --> 00:29:36,433 I just thought we should take a moment together 474 00:29:36,435 --> 00:29:40,803 and just think about what's about to happen. 475 00:29:40,805 --> 00:29:42,371 Extremely excited to be here. 476 00:29:42,373 --> 00:29:45,374 Extremely proud of everyone of you and what we've done 477 00:29:45,376 --> 00:29:48,076 and, win or loss, I think it's just amazing that we're here. 478 00:30:01,523 --> 00:30:04,491 (cameras clicking) 479 00:30:05,326 --> 00:30:08,027 (soft music) 480 00:30:51,370 --> 00:30:54,205 (dramatic music) 481 00:32:00,138 --> 00:32:02,105 [David] We had worked so hard to make sure 482 00:32:02,107 --> 00:32:05,541 that this would go technically smoothly. 483 00:32:05,543 --> 00:32:08,977 We tested it and tested it and tested it 484 00:32:08,979 --> 00:32:13,380 and still, there comes that moment, when you're live, 485 00:32:13,382 --> 00:32:16,716 all the TV cameras are broadcasting everything 486 00:32:16,718 --> 00:32:20,853 and now it has to do the thing you built it to do. 487 00:32:34,233 --> 00:32:37,235 (cameras clicking) 488 00:32:52,250 --> 00:32:55,085 (dramatic music) 489 00:33:09,400 --> 00:33:11,434 [Aja] I was a bit nervous. 490 00:33:11,436 --> 00:33:14,303 It's the first time that I sit in front 491 00:33:14,305 --> 00:33:17,272 of a world class Go player. 492 00:33:17,274 --> 00:33:20,475 And I actually can feel the spirit and courtesy 493 00:33:20,477 --> 00:33:23,478 of a great Go player like Lee Sedol 494 00:33:23,480 --> 00:33:25,046 because I think it is the first time 495 00:33:25,048 --> 00:33:27,581 he face a strange opponent. 496 00:33:27,583 --> 00:33:30,750 I think it's not a human, has no emotion. 497 00:33:30,752 --> 00:33:33,585 It's cold, but he stay very calm. 498 00:33:35,020 --> 00:33:37,787 And I can feel his mental strength. 499 00:33:40,223 --> 00:33:42,524 (soft music) 500 00:33:48,530 --> 00:33:49,630 - Oh! - Oh. 501 00:34:05,413 --> 00:34:06,379 I'm Andrew Jackson. 502 00:34:06,381 --> 00:34:07,379 With me here is Myungwan Kim, 503 00:34:07,381 --> 00:34:10,383 a nine dan prof from the Korean Baduk Association. 504 00:34:10,385 --> 00:34:12,518 We are here live at the Four Seasons Hotel 505 00:34:12,520 --> 00:34:13,819 on the 21st floor. 506 00:34:13,821 --> 00:34:14,653 What are we looking at? 507 00:34:14,655 --> 00:34:16,687 How's the game going? 508 00:34:19,157 --> 00:34:20,624 From the very beginning? 509 00:34:20,626 --> 00:34:21,524 Yeah, yeah. 510 00:34:21,526 --> 00:34:25,728 AlphaGo play very well, just like a top professional. 511 00:34:25,730 --> 00:34:27,564 Just like a top professional? 512 00:34:27,566 --> 00:34:30,767 [Myungwan] Yeah, it's very aggressive. 513 00:34:38,875 --> 00:34:41,543 Now, the fight is getting really complicated. 514 00:34:41,545 --> 00:34:44,779 This is actually the first time I've seen AlphaGo 515 00:34:44,781 --> 00:34:48,783 playing a game that has this difficult a fight. 516 00:35:10,705 --> 00:35:13,373 [Michael] I just saw him looking at his opponent's face 517 00:35:13,375 --> 00:35:15,274 and that's just a kind of a habit. 518 00:35:15,276 --> 00:35:17,509 [Chris] Just an instinct as a player to look 519 00:35:17,511 --> 00:35:19,511 at the person across the board from you. 520 00:35:19,513 --> 00:35:22,614 [Michael] It's sort of like something Lee Sedol would do 521 00:35:22,616 --> 00:35:25,283 when he was wondering how his opponent was feeling. 522 00:35:25,285 --> 00:35:26,684 It's just a habit. 523 00:35:26,686 --> 00:35:30,587 So, it's not as if Aja Huang is gonna do any giveaway 524 00:35:30,589 --> 00:35:33,556 cuz Aja Huang isn't AlphaGo, right. 525 00:35:39,763 --> 00:35:40,596 - Oo. - Oh! 526 00:35:40,598 --> 00:35:41,563 [Man] Oo. 527 00:36:19,835 --> 00:36:22,336 (soft music) 528 00:37:04,011 --> 00:37:05,378 It's hard to know where to be. 529 00:37:05,380 --> 00:37:06,779 All the different rooms are like exciting 530 00:37:06,781 --> 00:37:08,413 in a different way. 531 00:37:09,748 --> 00:37:11,949 Like nice to be here at the heart of the operation. 532 00:37:11,951 --> 00:37:14,952 [Demis] Yeah, I feel safe here. 533 00:37:19,857 --> 00:37:21,791 Yes, go for it. 534 00:37:21,793 --> 00:37:23,325 [Demis] It's done it. 535 00:37:23,327 --> 00:37:24,492 It's gone in. 536 00:37:24,494 --> 00:37:27,895 Ooh, look at his face, look at his face! 537 00:37:27,897 --> 00:37:29,696 That is not a confident face. 538 00:37:29,698 --> 00:37:32,632 He's pretty horrified by that. 539 00:37:32,634 --> 00:37:35,934 I can't believe what I see right now. 540 00:37:36,835 --> 00:37:40,037 She thought, you know, she's only a bit behind. 541 00:37:40,039 --> 00:37:42,639 It's made a very aggressive move here. 542 00:37:42,641 --> 00:37:44,574 It make it very complicated. 543 00:37:44,576 --> 00:37:48,077 I think Lee Sedol, if he doesn't respond mentally, 544 00:37:48,079 --> 00:37:49,912 then he can collapse. 545 00:38:05,761 --> 00:38:07,394 [Man] That's the maximum number 546 00:38:07,396 --> 00:38:08,795 of moves ahead that AlphaGo 547 00:38:08,797 --> 00:38:11,898 is looking from the current game position. 548 00:38:11,900 --> 00:38:14,100 It's typically over 50. 549 00:38:14,102 --> 00:38:15,935 It's often over 60. 550 00:38:17,070 --> 00:38:20,639 In the games we see often, about move a hundred and fifty, 551 00:38:20,641 --> 00:38:22,073 AlphaGo goes for the kill. 552 00:38:22,075 --> 00:38:23,941 We're move a hundred and fifteen now 553 00:38:23,943 --> 00:38:26,977 so we're going to that critical point. 554 00:38:29,080 --> 00:38:33,650 We are all astonished, just in the middle of the game, 555 00:38:33,652 --> 00:38:38,087 because AlphaGo seems to be doing a much better job 556 00:38:38,089 --> 00:38:40,389 than we all thought it would. 557 00:38:40,391 --> 00:38:42,991 I thought that Lee Sedol would be leading 558 00:38:42,993 --> 00:38:45,893 the game confidently, but it turned out 559 00:38:45,895 --> 00:38:47,995 that he's struggling at the moment. 560 00:38:47,997 --> 00:38:51,898 But I think eventually he will prevail, I hope. 561 00:38:51,900 --> 00:38:57,970 I was more five to zero, but now I'm not sure about that. 562 00:39:51,057 --> 00:39:53,224 Myungwan, it looks like you've gotta count. 563 00:39:53,226 --> 00:39:56,961 - Yeah, right. - What do you think? 564 00:39:56,963 --> 00:39:59,530 White won a lot at this time, if you check that. 565 00:39:59,532 --> 00:40:01,698 If it's like this, white has won by a lot. 566 00:40:14,011 --> 00:40:16,479 Wow, it's so shocking. 567 00:40:16,481 --> 00:40:19,882 I expected AlphaGo to win only one game. 568 00:40:26,756 --> 00:40:27,655 Oh. 569 00:40:33,595 --> 00:40:35,996 [Aja] I can feel his pain. 570 00:40:37,164 --> 00:40:38,631 He couldn't believe, you know. 571 00:40:38,633 --> 00:40:39,965 He couldn't accept it. 572 00:40:39,967 --> 00:40:43,902 It takes time for him to accept the outcome. 573 00:41:08,961 --> 00:41:12,663 [Aja] I think he resigned in a very polite way. 574 00:41:12,665 --> 00:41:16,566 Lee Sedol, he's black, but he put white stone. 575 00:41:23,840 --> 00:41:26,041 [Man] I think he resigned. 576 00:41:27,543 --> 00:41:28,709 Oh my gosh. 577 00:41:30,778 --> 00:41:32,211 Yeah, time has stopped. 578 00:41:32,213 --> 00:41:33,111 Wow. 579 00:41:33,946 --> 00:41:36,814 (upbeat music) 580 00:41:43,120 --> 00:41:45,154 [Woman] Congratulations. 581 00:41:47,190 --> 00:41:48,690 [David] I feel really good. 582 00:41:48,692 --> 00:41:51,692 I feel like I really believe in AlphaGo. 583 00:41:53,094 --> 00:41:54,961 Of course, it's natural that humans want humans to win. 584 00:41:54,963 --> 00:41:56,996 I mean I think that's a natural response. 585 00:41:56,998 --> 00:42:00,132 But AlphaGo is human-created and I think that's the ultimate 586 00:42:00,134 --> 00:42:02,868 sign of human ingenuity and cleverness. 587 00:42:02,870 --> 00:42:05,837 Everything that AlphaGo does, it does because a human 588 00:42:05,839 --> 00:42:08,973 has either created the data that it learns from, 589 00:42:08,975 --> 00:42:11,842 created the learning algorithm that learns from that data, 590 00:42:11,844 --> 00:42:14,211 created the search algorithm, all of these things 591 00:42:14,213 --> 00:42:16,313 have come from humans. 592 00:42:16,315 --> 00:42:19,683 So, really, this is a human endeavor. 593 00:42:27,624 --> 00:42:30,592 (cameras clicking) 594 00:42:34,096 --> 00:42:35,996 [Reporter] In the battle between man versus machine, 595 00:42:35,998 --> 00:42:37,931 a computer just came out the victor. 596 00:42:37,933 --> 00:42:40,633 [Reporter] DeepMind put its computer program to the test 597 00:42:40,635 --> 00:42:43,636 against one of the brightest minds in the world and won. 598 00:42:43,638 --> 00:42:45,938 [Reporter] AlphaGo beat a professional player 599 00:42:45,940 --> 00:42:49,274 who has 18 Go World Championships under his belt. 600 00:42:49,276 --> 00:42:51,242 The victory is considered a breakthrough 601 00:42:51,244 --> 00:42:53,777 in artificial intelligence. 602 00:43:46,663 --> 00:43:49,131 [David] In research, we normally work 603 00:43:49,133 --> 00:43:50,732 to produce an academic paper. 604 00:43:50,734 --> 00:43:52,734 It gets published and maybe we get to talk about it 605 00:43:52,736 --> 00:43:55,269 in a conference, if we're lucky. 606 00:43:56,104 --> 00:43:58,171 This is not normal for research. 607 00:43:58,173 --> 00:44:01,107 In fact, I've never experienced any media attention 608 00:44:01,109 --> 00:44:02,975 remotely close to this. 609 00:44:02,977 --> 00:44:04,910 So it's a special moment for us all 610 00:44:04,912 --> 00:44:08,280 and we're just enjoying it while it lasts. 611 00:44:12,451 --> 00:44:13,951 [Reporter] In a match up between man 612 00:44:13,953 --> 00:44:15,752 and machine, who wins? 613 00:44:15,754 --> 00:44:17,687 Well, so far, it's the machine. 614 00:44:17,689 --> 00:44:21,090 In Seoul, South Korea, the artificially intelligent computer 615 00:44:21,092 --> 00:44:23,859 defeated the global champion in the ancient 616 00:44:23,861 --> 00:44:26,228 Chinese board game Go. 617 00:44:26,230 --> 00:44:28,897 Lee Sedol lost the first match up, 618 00:44:28,899 --> 00:44:30,665 but he's got four more chances. 619 00:44:46,114 --> 00:44:48,148 It is a bit strange being on the front cover 620 00:44:48,150 --> 00:44:50,350 and everything as a computer scientist. 621 00:44:50,352 --> 00:44:52,919 Nobody used to learn codes. 622 00:44:52,921 --> 00:44:54,854 Nobody really knows about it. 623 00:44:54,856 --> 00:44:56,822 Perhaps heard the joke, how can you tell 624 00:44:56,824 --> 00:44:59,224 that a computer scientist is an extrovert 625 00:44:59,226 --> 00:45:01,125 and not an introvert? 626 00:45:01,127 --> 00:45:03,026 If he's an extrovert, he looks at your shoes 627 00:45:03,028 --> 00:45:07,162 when he's talking to you instead of at his own. 628 00:45:07,164 --> 00:45:09,797 [Julian] Like if you look at Aja, he avoids 629 00:45:09,799 --> 00:45:12,132 all the cameras like crazy. 630 00:45:12,134 --> 00:45:13,867 He's like, "The game is finished," 631 00:45:13,869 --> 00:45:17,403 and he's like, "I'll go back in his room." 632 00:45:17,405 --> 00:45:21,406 And I think a lot of computer scientists would be like that. 633 00:45:21,408 --> 00:45:23,441 We're more about doing our work 634 00:45:23,443 --> 00:45:26,744 than standing in the spotlight. 635 00:45:26,746 --> 00:45:29,413 [Announcer] Hello and welcome to game two, 636 00:45:29,415 --> 00:45:34,184 round two in the Google DeepMind challenge throw down 637 00:45:34,186 --> 00:45:36,319 between man and machine. 638 00:45:36,321 --> 00:45:40,222 Game one, the machine takes down a man. 639 00:45:40,224 --> 00:45:41,823 Huge shock. 640 00:45:41,825 --> 00:45:43,391 Headlines around the world. 641 00:45:43,393 --> 00:45:45,159 [Man] The reactions on the ground here 642 00:45:45,161 --> 00:45:47,494 from the folks in Korea were just stunned. 643 00:45:47,496 --> 00:45:49,095 They estimate 60 million people 644 00:45:49,097 --> 00:45:51,163 watched the game in China alone. 645 00:45:51,165 --> 00:45:53,765 It's probably bringing up to maybe 80 million people 646 00:45:53,767 --> 00:45:55,333 watched this game worldwide. 647 00:45:55,335 --> 00:45:57,201 It was just incredible and, if anything, 648 00:45:57,203 --> 00:46:00,971 today is probably even more of a madhouse. 649 00:46:00,973 --> 00:46:03,506 So Lee Sedol knows today, he knows that this is just 650 00:46:03,508 --> 00:46:05,241 a really important game, right. 651 00:46:05,243 --> 00:46:06,375 He's gotta win. 652 00:46:25,327 --> 00:46:27,895 [Michael] Lee Sedol on white. 653 00:46:27,897 --> 00:46:31,265 I think probably looking for a little payback. 654 00:46:31,267 --> 00:46:33,000 [Chris] My impression is that maybe 655 00:46:33,002 --> 00:46:36,903 he underestimated AlphaGo and he's gonna change his tactics. 656 00:46:41,308 --> 00:46:42,207 Yeah. 657 00:46:44,410 --> 00:46:46,177 He's playing at half his speed actually. 658 00:46:59,390 --> 00:47:01,257 [Michael] Well, if I say anything about AlphaGo 659 00:47:01,259 --> 00:47:04,360 that is not normal, is maybe the way it handles a game 660 00:47:04,362 --> 00:47:06,395 when it thinks its ahead. 661 00:47:06,397 --> 00:47:07,495 We're actually gonna have a visit 662 00:47:07,497 --> 00:47:10,265 from one of the team and we'll talk about exactly 663 00:47:10,267 --> 00:47:11,899 that point from the inside. 664 00:47:11,901 --> 00:47:13,233 Thanks so much for coming by. 665 00:47:13,235 --> 00:47:14,333 I really appreciate it. 666 00:47:14,335 --> 00:47:19,238 Can you sort of share a bit of what is going on in AlphaGo. 667 00:47:19,240 --> 00:47:22,574 So AlphaGo has these three main components. 668 00:47:22,576 --> 00:47:25,577 There's the Policy Network which was trained 669 00:47:25,579 --> 00:47:30,147 on high level games to imitate those players. 670 00:47:30,149 --> 00:47:32,082 And then we have a second component, 671 00:47:32,084 --> 00:47:35,251 we call this the Value Net and it can evaluate 672 00:47:35,253 --> 00:47:38,220 the board position and say what is the probability 673 00:47:38,222 --> 00:47:41,190 of winning in this particular position. 674 00:47:41,192 --> 00:47:44,059 And the third component is the Tree Search 675 00:47:44,061 --> 00:47:47,729 where it would look through different variations of the game 676 00:47:47,731 --> 00:47:52,166 and try to figure out what will happen in the future. 677 00:47:53,668 --> 00:47:56,403 So, if we now take a position like this, 678 00:47:56,405 --> 00:48:00,173 first the Policy Network will scan the position 679 00:48:00,175 --> 00:48:04,343 and come up with what would be the interesting spots to play 680 00:48:05,445 --> 00:48:08,246 and it builds up a tree of variations. 681 00:48:08,248 --> 00:48:12,616 And then employs this Value Net that tells it how promising 682 00:48:12,618 --> 00:48:16,552 is the outcome of this particular variation. 683 00:48:18,521 --> 00:48:22,556 So AlphaGo tries to maximize its probability of winning, 684 00:48:23,658 --> 00:48:26,126 but it doesn't care at all about the margin 685 00:48:26,128 --> 00:48:27,560 by which it wins. 686 00:48:27,562 --> 00:48:29,461 Okay, so when you see a slow-looking move 687 00:48:29,463 --> 00:48:31,429 that's maybe an indication that AlphaGo 688 00:48:31,431 --> 00:48:33,464 thinks it has a good chance to win. 689 00:48:33,466 --> 00:48:36,300 Yeah, that is a little giveaway. 690 00:48:36,302 --> 00:48:38,068 A little tell, we're looking for a tell. 691 00:49:08,131 --> 00:49:10,098 [Man] Oh, looks like Lee is taking 692 00:49:10,100 --> 00:49:12,166 a little bit of a break. 693 00:49:14,269 --> 00:49:16,536 (soft music) 694 00:49:45,498 --> 00:49:48,466 (speaking in foreign language) 695 00:49:51,837 --> 00:49:53,603 The value, uh... 696 00:49:55,672 --> 00:49:58,640 That's a very surprising move. 697 00:49:58,642 --> 00:50:01,776 [Chris] I thought it was a mistake. 698 00:50:21,562 --> 00:50:25,631 So coming on top of a fourth line zone is really unusual. 699 00:50:25,633 --> 00:50:27,332 [Chris] Yeah, that's an exciting move. 700 00:50:27,334 --> 00:50:30,501 I think we're seeing an original move here. 701 00:50:30,503 --> 00:50:34,538 That's the kind of move that you play Go for. 702 00:50:36,541 --> 00:50:37,607 [Man] Hey. 703 00:50:39,643 --> 00:50:41,310 Interesting stuff. 704 00:50:41,312 --> 00:50:43,712 This fifth line shoulder hit is uh... 705 00:50:43,714 --> 00:50:46,314 [Michael] I wasn't expecting that. 706 00:50:46,316 --> 00:50:47,515 I don't really know if it's a good 707 00:50:47,517 --> 00:50:49,283 or bad move at this point. 708 00:50:49,285 --> 00:50:52,419 [David] The professional commentators almost unanimously 709 00:50:52,421 --> 00:50:54,187 said that not a single human player 710 00:50:54,189 --> 00:50:56,489 would have chosen move 37. 711 00:50:56,491 --> 00:50:58,657 So I actually had a poke around in AlphaGo 712 00:50:58,659 --> 00:51:00,692 to see what AlphaGo thought. 713 00:51:00,694 --> 00:51:03,895 And AlphaGo actually agreed with that assessment. 714 00:51:03,897 --> 00:51:07,565 AlphaGo said it was a one in 10,000 probability 715 00:51:07,567 --> 00:51:11,135 that move 37 would have been played by a human player. 716 00:51:11,137 --> 00:51:13,704 So it knew that this was an extremely unlikely move. 717 00:51:13,706 --> 00:51:17,474 It went beyond its human guide and it came up 718 00:51:17,476 --> 00:51:20,777 with something new and creative and different. 719 00:51:20,779 --> 00:51:23,613 I am very much watching the game 720 00:51:23,615 --> 00:51:25,781 through these commentators. 721 00:51:25,783 --> 00:51:27,716 That's the way it works, so when they're confused 722 00:51:27,718 --> 00:51:29,350 I'm certainly confused. 723 00:51:29,352 --> 00:51:31,251 At the same time, I'm latching onto the fact 724 00:51:31,253 --> 00:51:32,752 that they are confused, right. 725 00:51:32,754 --> 00:51:34,787 That is an interesting moment. 726 00:51:34,789 --> 00:51:36,221 When everyone else is confused, 727 00:51:36,223 --> 00:51:39,324 who's not confused besides the machine? 728 00:51:51,870 --> 00:51:54,705 [Chris] He's back, Lee is back. 729 00:51:58,876 --> 00:52:01,811 (dramatic music) 730 00:53:07,576 --> 00:53:08,943 Go is like geopolitics. 731 00:53:08,945 --> 00:53:10,811 Like something small that happens here 732 00:53:10,813 --> 00:53:14,780 can have a ripple effect hours down the road 733 00:53:14,782 --> 00:53:16,815 in a different part of the board. 734 00:53:16,817 --> 00:53:20,852 The game kind of turned on its axis at that moment. 735 00:53:51,916 --> 00:53:54,284 [Michael] This is a tough game for Lee Sedol. 736 00:53:54,286 --> 00:53:58,487 AlphaGo is just not letting Lee Sedol do what he wants. 737 00:53:58,489 --> 00:54:00,322 Black has almost 60 points. 738 00:54:00,324 --> 00:54:01,556 That's a lot. 739 00:54:01,558 --> 00:54:04,825 [Andrew] That's not a good sign. 740 00:54:04,827 --> 00:54:08,694 Lee Sedol just slapped himself on the side of the head. 741 00:54:08,696 --> 00:54:10,829 [Myungwan] Oh wow. 742 00:54:10,831 --> 00:54:13,832 I think black's ahead at this point. 743 00:54:13,834 --> 00:54:14,766 It's looking good, ain't it? 744 00:54:14,768 --> 00:54:17,535 We're on that steady path now. 745 00:54:33,751 --> 00:54:36,018 (soft music) 746 00:54:48,431 --> 00:54:50,365 Oh, he resigned. 747 00:54:50,367 --> 00:54:53,935 It looks like Lee Sedol has just resigned. 748 00:55:21,396 --> 00:55:25,632 There was this heavy sadness over that whole floor. 749 00:55:25,634 --> 00:55:26,766 You could feel it during the game. 750 00:55:26,768 --> 00:55:28,634 I felt it during the game. 751 00:55:28,636 --> 00:55:31,436 And I'm leaving the commentary room to go 752 00:55:31,438 --> 00:55:33,938 to the press conference and I was stopped by someone, 753 00:55:33,940 --> 00:55:35,606 another technology reporter. 754 00:55:35,608 --> 00:55:38,075 At first, all he wanted to talk about was the technology 755 00:55:38,077 --> 00:55:39,543 and how great this was. 756 00:55:39,545 --> 00:55:41,978 But then even he kind of slipped into this moment 757 00:55:41,980 --> 00:55:45,648 of melancholy where he was upset as well. 758 00:55:52,889 --> 00:55:54,656 I am quite speechless. 759 00:55:54,658 --> 00:55:58,092 I admit that it was a very clear loss on my part. 760 00:55:58,094 --> 00:55:59,960 From the very beginning of the game, 761 00:55:59,962 --> 00:56:01,962 there was not a moment in time that I felt 762 00:56:01,964 --> 00:56:04,464 that I was leading the game. 763 00:56:04,466 --> 00:56:08,034 You feel elated and you feel a little bit scared. 764 00:56:08,036 --> 00:56:11,103 There is something, I think, frightening to people 765 00:56:11,105 --> 00:56:14,039 about a machine that learns on its own. 766 00:56:14,041 --> 00:56:17,809 For us, AlphaGo was obviously just some computer program, 767 00:56:17,811 --> 00:56:20,678 but looking at the commentary on the internet, 768 00:56:20,680 --> 00:56:22,946 I already saw the commentators call AlphaGo 769 00:56:22,948 --> 00:56:26,983 like he and she during the games, completely unconsciously. 770 00:56:28,986 --> 00:56:31,520 AlphaGo is really a very, very simple program. 771 00:56:31,522 --> 00:56:34,055 It's not anywhere close to full AI and we already 772 00:56:34,057 --> 00:56:38,492 see that happening, so I found that very interesting. 773 00:56:41,562 --> 00:56:43,996 The tendency to anthropomorphize AI systems 774 00:56:43,998 --> 00:56:47,666 is one of the big obstacles in the way of actually trying 775 00:56:47,668 --> 00:56:51,102 to understand how AI might impact the world in the future. 776 00:56:51,104 --> 00:56:53,571 For example, the conversation is about what could go wrong, 777 00:56:53,573 --> 00:56:55,573 like what the risks are and invariably 778 00:56:55,575 --> 00:56:57,007 you see this Terminator picture. 779 00:56:57,009 --> 00:56:59,843 Every single time, there are these red, glowing eyes, right? 780 00:56:59,845 --> 00:57:01,778 We're really closer to a smart 781 00:57:01,780 --> 00:57:03,946 washing machine than Terminator. 782 00:57:03,948 --> 00:57:09,184 If you look at today's AI, we are really very nascent. 783 00:57:09,952 --> 00:57:13,987 I'm extremely excited and passionate about AI's potential, 784 00:57:15,155 --> 00:57:19,157 but AI is still very limited in its power. 785 00:57:19,159 --> 00:57:21,993 I think that people are right to think 786 00:57:21,995 --> 00:57:24,962 that there is a danger that, as we continue to improve 787 00:57:24,964 --> 00:57:28,165 these systems that we might misstep that threshold 788 00:57:28,167 --> 00:57:31,034 where we do cross over into danger. 789 00:57:31,036 --> 00:57:32,935 But the good news is there are already people 790 00:57:32,937 --> 00:57:34,469 thinking about those dangers. 791 00:57:34,471 --> 00:57:35,803 You know, there's a lot of talk now 792 00:57:35,805 --> 00:57:37,203 and we're leading the discussion on this, 793 00:57:37,205 --> 00:57:39,573 that maybe there should be a kind of cross industry, 794 00:57:39,575 --> 00:57:41,975 best practices working group or something 795 00:57:41,977 --> 00:57:44,811 where the leaders of the research teams 796 00:57:44,813 --> 00:57:47,046 in those organizations, you know, the big ones 797 00:57:47,048 --> 00:57:50,482 that are working on AI, IBM, Microsoft, so on, come together 798 00:57:50,484 --> 00:57:54,252 and make sure that AI is used ethically and responsibly. 799 00:57:54,254 --> 00:57:57,555 I think what is important is that there is this community 800 00:57:57,557 --> 00:58:01,658 of people who are leading the cutting edge of AI, 801 00:58:03,027 --> 00:58:05,795 who are interacting with academics and already are thinking 802 00:58:05,797 --> 00:58:09,598 about the longterm and how we can ensure 803 00:58:09,600 --> 00:58:12,033 that innovation is responsible as the power 804 00:58:12,035 --> 00:58:15,102 of these machines gets even greater. 805 00:58:16,037 --> 00:58:18,071 This is it, folks. 806 00:58:18,073 --> 00:58:20,206 Day three, game three. 807 00:58:20,208 --> 00:58:24,142 Lee Sedol, Go master, back to the wall. 808 00:58:24,144 --> 00:58:26,544 He's down two-zero. 809 00:58:26,546 --> 00:58:29,413 He's got to win today to keep hope alive. 810 00:58:44,028 --> 00:58:47,963 I heard that four pros went to visit him 811 00:58:47,965 --> 00:58:50,799 to console him into a review after-- 812 00:58:50,801 --> 00:58:51,633 Console him? 813 00:58:51,635 --> 00:58:53,534 You think he's upset? 814 00:58:53,536 --> 00:58:54,935 I mean Lee Sedol is upset. 815 00:58:54,937 --> 00:58:56,603 Yeah. 816 00:58:56,605 --> 00:58:59,105 In the beginning, there was a forceful fight. 817 00:58:59,107 --> 00:59:03,809 And AlphaGo play very well, so he secure a very early lead. 818 00:59:03,811 --> 00:59:07,179 From (mumbles) the window was very high already. 819 00:59:07,181 --> 00:59:10,949 It was climbing towards a hundred percent. 820 00:59:55,761 --> 00:59:57,895 It's looking good for us. 821 00:59:57,897 --> 01:00:00,764 That black ruby is huge and it's got nowhere to go 822 01:00:00,766 --> 01:00:03,599 and it's gonna be running around. 823 01:00:04,667 --> 01:00:07,135 (soft music) 824 01:00:12,175 --> 01:00:15,243 [Commentator] I don't know how to describe this situation. 825 01:00:15,245 --> 01:00:17,979 If I were black, I will resign. 826 01:00:17,981 --> 01:00:21,282 He should admit that we are facing the strongest 827 01:00:21,284 --> 01:00:23,984 existence ever in the Go history. 828 01:00:25,086 --> 01:00:26,853 There's no point in playing out the end game 829 01:00:26,855 --> 01:00:29,221 when you're gonna lose, right. 830 01:00:32,091 --> 01:00:34,258 [Chris] Even if black can live there, oh, he resigned. 831 01:00:34,260 --> 01:00:35,259 [Michael] It's done, it's done. 832 01:00:35,261 --> 01:00:37,060 - Okay. - Wow. 833 01:00:37,062 --> 01:00:37,960 Wow. 834 01:00:39,329 --> 01:00:42,364 You saw history made here tonight. 835 01:00:42,366 --> 01:00:44,366 [Chris] AlphaGo has won again, three straight wins. 836 01:00:44,368 --> 01:00:46,801 [Michael] Three straight wins, has won the match. 837 01:01:09,957 --> 01:01:11,757 [Demis] You know, certainly, I feel a bit ambivalent 838 01:01:11,759 --> 01:01:13,425 about it given I'm a games player 839 01:01:13,427 --> 01:01:16,394 and Go is the pinnacle of board games. 840 01:01:16,396 --> 01:01:17,828 But I really like the statement 841 01:01:17,830 --> 01:01:20,363 one of the top Chinese professionals said, 842 01:01:20,365 --> 01:01:24,066 "If AlphaGo wins, maybe we'll really start 843 01:01:24,068 --> 01:01:27,802 to get to see what this game is about." 844 01:01:27,804 --> 01:01:29,770 [Maddy] I couldn't celebrate. 845 01:01:29,772 --> 01:01:32,305 It was fantastic that we had won, 846 01:01:32,307 --> 01:01:36,208 but there was such a big part of me that saw this man 847 01:01:36,210 --> 01:01:39,878 trying so hard and being so disappointed. 848 01:03:58,416 --> 01:04:01,184 (dramatic music) 849 01:04:08,025 --> 01:04:10,026 [Cade] You can see that he was more relaxed 850 01:04:10,028 --> 01:04:12,395 after he had lost three games in a row. 851 01:04:12,397 --> 01:04:14,697 That said, the stakes were still high. 852 01:04:14,699 --> 01:04:18,534 You know, in the end, it isn't about pride. 853 01:04:23,605 --> 01:04:27,607 [Chris] Can Lee Sedol find AlphaGo's weakness? 854 01:04:27,609 --> 01:04:29,675 Is there, in fact, a weakness? 855 01:04:29,677 --> 01:04:31,977 I was thinking I'd just pull the plug. 856 01:04:31,979 --> 01:04:33,178 - You'd just pull the plug? - Yeah. 857 01:04:33,180 --> 01:04:35,346 Everyone's still cheering for Lee Sedol 858 01:04:35,348 --> 01:04:38,415 and, yeah, it's not going very well. 859 01:04:49,993 --> 01:04:53,228 It feels pretty good for black at this point. 860 01:04:53,230 --> 01:04:54,862 It feels pretty good for black, yes. 861 01:05:37,471 --> 01:05:39,405 [Michael] It's developing into a very, 862 01:05:39,407 --> 01:05:41,073 very dangerous fight. 863 01:05:41,075 --> 01:05:43,508 This is really Lee Sedol's type of game. 864 01:05:43,510 --> 01:05:45,343 He likes this kind of fight. 865 01:05:45,345 --> 01:05:47,511 White has to find something inside black's territory. 866 01:05:47,513 --> 01:05:49,446 [Commentator] I think he is already planning 867 01:05:49,448 --> 01:05:50,513 on trying something. 868 01:05:56,186 --> 01:05:57,452 Yes, yes. 869 01:05:57,454 --> 01:05:59,520 There is a little potential there. 870 01:05:59,522 --> 01:06:02,089 And I think maybe he's gonna try to do something. 871 01:06:02,091 --> 01:06:03,923 [Andrew] Lee Sedol magic. 872 01:06:06,459 --> 01:06:08,526 Lee Sedol, he's running short on time, 873 01:06:08,528 --> 01:06:11,262 but he's gonna have to use up all his time. 874 01:06:11,264 --> 01:06:13,497 [Commentator] He just burned like seven or eight minutes 875 01:06:13,499 --> 01:06:14,931 just on this move already. 876 01:06:18,035 --> 01:06:19,502 We can't find anything. 877 01:06:23,273 --> 01:06:25,140 [Chris] What is Lee Sedol up to here? 878 01:06:25,142 --> 01:06:26,808 [Michael] Yeah, he's really concentrating. 879 01:06:26,810 --> 01:06:28,509 [Chris] He really is, look at that. 880 01:06:52,667 --> 01:06:55,301 [Chris] This is the hinge of the game. 881 01:06:55,303 --> 01:06:57,169 Oh, look at that move. 882 01:06:57,171 --> 01:06:58,703 That's an exciting move. 883 01:06:58,705 --> 01:07:01,205 - Look. - Oh, he found the wedge. 884 01:07:01,207 --> 01:07:02,406 [Chris] Whoa. 885 01:07:02,408 --> 01:07:04,641 [Michael] It's gonna change the equation 886 01:07:04,643 --> 01:07:06,542 because now black cannot escape. 887 01:07:06,544 --> 01:07:09,577 That would be so cool if that works. 888 01:07:12,380 --> 01:07:13,279 Ooh! 889 01:07:15,415 --> 01:07:16,782 [Commentator] AlphaGo has just played 890 01:07:16,784 --> 01:07:19,251 something maybe unusual. 891 01:07:19,253 --> 01:07:20,585 [Michael] You know, I'm not actually sure 892 01:07:20,587 --> 01:07:22,720 what AlphaGo is trying to do here. 893 01:07:22,722 --> 01:07:24,154 [Chris] What's that about? 894 01:07:24,156 --> 01:07:25,588 [Michael] I don't really understand it. 895 01:07:25,590 --> 01:07:26,322 Well, well. 896 01:07:36,832 --> 01:07:40,267 This could be that it actually can't find a way through. 897 01:07:40,269 --> 01:07:43,270 I think this is it's looked far enough ahead to see 898 01:07:43,272 --> 01:07:45,338 that it doesn't work and now maybe it's on tilt? 899 01:07:45,340 --> 01:07:46,672 I don't know. 900 01:07:46,674 --> 01:07:48,340 There it comes. 901 01:07:50,109 --> 01:07:53,511 It's like it's fallen off a cliff. 902 01:07:53,513 --> 01:07:55,613 Yeah, it's made a mistake. 903 01:07:55,615 --> 01:07:57,848 Did anything strange happen in the... 904 01:07:57,850 --> 01:07:59,516 [Man] No, it all looked normal. 905 01:07:59,518 --> 01:08:02,685 Well, we can definitely say, well, there's a weakness. 906 01:08:02,687 --> 01:08:05,654 Well, we definitely say it's a mistake. 907 01:08:05,656 --> 01:08:09,391 [David] I felt this mixture of a sinking feeling 908 01:08:09,393 --> 01:08:11,559 in my stomach where I was wondering if AlphaGo 909 01:08:11,561 --> 01:08:13,894 is becoming delusional in this situation where I could see 910 01:08:13,896 --> 01:08:16,296 that it was starting to play strangely. 911 01:08:16,298 --> 01:08:18,197 And, at the same time, relief that Lee Sedol, 912 01:08:18,199 --> 01:08:21,767 that he was actually in with a chance now. 913 01:08:26,873 --> 01:08:31,576 I knew after move 78, after like 10 or 20 moves, 914 01:08:31,578 --> 01:08:34,178 I saw AlphaGo strange moves. 915 01:08:35,380 --> 01:08:37,614 I know AlphaGo somehow became crazy, 916 01:08:37,616 --> 01:08:39,415 but I didn't realize why. 917 01:08:39,417 --> 01:08:40,315 Whoa! 918 01:08:41,150 --> 01:08:43,317 We slipped from 95 ahead. 919 01:08:44,419 --> 01:08:46,553 At the point, where we made the mistake? 920 01:08:46,555 --> 01:08:48,388 I think that something went wrong. 921 01:08:48,390 --> 01:08:51,357 That's the longest we've searched this entire game. 922 01:08:51,359 --> 01:08:56,227 I think it's searched so deeply it's confused itself. 923 01:08:56,229 --> 01:08:57,661 [Michael] I do get the impression that AlphaGo 924 01:08:57,663 --> 01:08:59,829 is sort of gone off on a tangent. 925 01:08:59,831 --> 01:09:01,397 What are you doing now? 926 01:09:01,399 --> 01:09:02,731 Maybe it has a master plan? 927 01:09:02,733 --> 01:09:05,600 No, it doesn't even think it has, does it? 928 01:09:05,602 --> 01:09:06,934 So it knows it's made a mistake. 929 01:09:06,936 --> 01:09:09,603 It starts evaluating it the other way. 930 01:09:09,605 --> 01:09:11,371 Look, look, Lee's so confused. 931 01:09:11,373 --> 01:09:12,372 He's like what is it doing? 932 01:09:12,374 --> 01:09:13,673 That's not a I'm scared confused. 933 01:09:13,675 --> 01:09:15,074 That's a what is it doing? 934 01:09:16,609 --> 01:09:18,576 You know, you asked if it was a bug. 935 01:09:18,578 --> 01:09:22,746 I've said before that if DeepMind has figured out 936 01:09:24,248 --> 01:09:26,916 how to write code that doesn't have bugs 937 01:09:26,918 --> 01:09:29,551 that is a bigger news story than AlphaGo. 938 01:09:29,553 --> 01:09:31,586 Oh, you're kidding me. 939 01:09:31,588 --> 01:09:34,521 This next move we're gonna play, 940 01:09:35,789 --> 01:09:37,723 I think they're gonna laugh. 941 01:09:37,725 --> 01:09:39,791 Everybody's gonna laugh. 942 01:09:47,365 --> 01:09:48,431 [Commentator] Oh, oh, oh! 943 01:09:54,471 --> 01:09:55,704 [Man] I don't really know 944 01:09:55,706 --> 01:09:57,939 what AlphaGo is trying to do here. 945 01:09:57,941 --> 01:09:59,707 That's the understatement of the year. 946 01:10:01,643 --> 01:10:02,843 [Andrew] No, no, that's the move. 947 01:10:02,845 --> 01:10:05,645 Aja Huang makes no misclicks. 948 01:10:05,647 --> 01:10:07,713 Lee Sedol is very confused. 949 01:10:07,715 --> 01:10:09,648 These are not human moves. 950 01:10:09,650 --> 01:10:12,884 This move, also, sort of inexplicable. 951 01:10:12,886 --> 01:10:16,687 [Chris] I mean you can clearly call those the saves. 952 01:10:16,689 --> 01:10:17,888 [Michael] Yes, of course. 953 01:10:17,890 --> 01:10:19,756 But it's the first time in the four matches 954 01:10:19,758 --> 01:10:21,758 that we've seen moves like that. 955 01:10:21,760 --> 01:10:24,460 Oh, the value dropped even more. 956 01:10:24,462 --> 01:10:25,794 That's weird. 957 01:10:32,968 --> 01:10:34,969 [Andrew] Come on, Lee Sedol. 958 01:10:34,971 --> 01:10:36,804 This unbelievable. 959 01:10:36,806 --> 01:10:37,704 All right! 960 01:10:37,706 --> 01:10:40,707 I was so confident that this black mire 961 01:10:40,709 --> 01:10:42,775 would just be consolidated by black 962 01:10:42,777 --> 01:10:43,775 and there's nothing there. 963 01:10:43,777 --> 01:10:47,279 And, somehow, he's just erased it all, like it's gone. 964 01:10:47,281 --> 01:10:48,713 So he's found his weakness? 965 01:10:48,715 --> 01:10:51,148 That wedge move probably surprised it. 966 01:10:52,650 --> 01:10:54,484 AlphaGo seemed to have such a good game 967 01:10:54,486 --> 01:10:57,687 and then this whole sequence in the center 968 01:10:57,689 --> 01:10:58,821 sort of changed that. 969 01:10:58,823 --> 01:11:00,889 He pulls off a miracle. 970 01:11:00,891 --> 01:11:02,790 He managed to make it so complicated 971 01:11:02,792 --> 01:11:06,360 that the artificial intelligence doesn't evaluate it. 972 01:11:06,362 --> 01:11:08,830 I think it looks like Aja Huang maybe sitting there 973 01:11:08,832 --> 01:11:11,365 also knows that the game is over. 974 01:11:15,936 --> 01:11:17,802 [Andrew] Now, I haven't seen him smile yet. 975 01:11:20,105 --> 01:11:21,738 [Andrew] He's so serious. 976 01:11:21,740 --> 01:11:23,706 He wants to be careful. 977 01:11:40,956 --> 01:11:43,757 (dramatic music) 978 01:11:47,828 --> 01:11:51,130 [Man] There you go, AlphaGo resigned. 979 01:11:52,866 --> 01:11:54,800 Looks like AlphaGo has resigned. 980 01:11:54,802 --> 01:11:56,568 [Chris] Wow. 981 01:11:56,570 --> 01:11:58,636 The most amazing game, 982 01:12:00,005 --> 01:12:01,438 I'm almost gonna tear up, 983 01:12:01,440 --> 01:12:04,974 was game four where he comes back and wins. 984 01:12:07,610 --> 01:12:10,411 (crowd cheering) 985 01:12:54,188 --> 01:12:55,888 I heard from so many people 986 01:12:55,890 --> 01:12:59,590 saying they were running out in the street. 987 01:12:59,592 --> 01:13:01,158 They were so happy. 988 01:13:01,160 --> 01:13:02,993 They were chanting, they were celebrating. 989 01:13:02,995 --> 01:13:07,197 Especially after a zero-three down, at the time, 990 01:13:07,199 --> 01:13:11,534 seems to be hopeless that the end of the world is coming, 991 01:13:11,536 --> 01:13:13,002 but we see the light. 992 01:13:13,004 --> 01:13:18,006 Usually, I'm happy when I win not like my colleague wins. 993 01:13:18,974 --> 01:13:22,076 But this time it felt like my win. 994 01:13:37,091 --> 01:13:39,959 (upbeat music) 995 01:13:46,899 --> 01:13:49,734 (crowd clapping) 996 01:14:34,178 --> 01:14:37,680 My question is about number 78 move, 997 01:14:37,682 --> 01:14:41,083 Chinese top Go player (mumbles) it was a God play, 998 01:14:41,085 --> 01:14:43,018 what were you thinking when you made that play? 999 01:14:43,020 --> 01:14:43,985 Thank you. 1000 01:14:59,667 --> 01:15:01,134 [Demis] It does leave me a little bit in awe 1001 01:15:01,136 --> 01:15:02,935 of the human brain's power, in particular, 1002 01:15:02,937 --> 01:15:06,338 Lee's amazing ability to cause AlphaGo problems 1003 01:15:06,340 --> 01:15:09,140 and find something seemingly out of nothing. 1004 01:15:09,142 --> 01:15:12,243 And so we really want to understand what had happened. 1005 01:16:05,630 --> 01:16:07,297 (upbeat music) 1006 01:16:07,299 --> 01:16:09,832 [Announcer] Welcome back to the Four Seasons. 1007 01:16:09,834 --> 01:16:12,801 World champion, Lee Sedol, went looking 1008 01:16:12,803 --> 01:16:17,172 for AlphaGo's weakness in game four and he found it. 1009 01:16:17,174 --> 01:16:20,108 Today, last round, he's looking to see 1010 01:16:20,110 --> 01:16:21,809 if he can repeat that. 1011 01:16:21,811 --> 01:16:22,809 We'll see what happens. 1012 01:16:22,811 --> 01:16:24,878 [Commentator] The place is a madhouse downstairs. 1013 01:16:24,880 --> 01:16:27,714 There may be as many, if nor more, journalists here today 1014 01:16:27,716 --> 01:16:30,316 than there were for game one. 1015 01:16:30,318 --> 01:16:33,118 [Maddy] We were really excited for the fifth match 1016 01:16:33,120 --> 01:16:34,352 to see what would happen. 1017 01:16:34,354 --> 01:16:36,087 Was this going to be a three-two thing 1018 01:16:36,089 --> 01:16:38,089 or was it going to be a four-one thing? 1019 01:16:38,091 --> 01:16:39,390 And, you know, there's a different message 1020 01:16:39,392 --> 01:16:41,225 with both of those. 1021 01:17:08,786 --> 01:17:12,021 You weren't crazy about the timing on this move. 1022 01:17:12,023 --> 01:17:13,222 No, yeah. 1023 01:17:13,224 --> 01:17:15,724 And now, you don't approve of this? 1024 01:17:15,726 --> 01:17:18,193 Yeah, I'm sort of thinking that maybe AlphaGo 1025 01:17:18,195 --> 01:17:20,261 hasn't recovered from game four yet. 1026 01:17:24,699 --> 01:17:25,432 Oh no, I hope it doesn't... 1027 01:17:25,434 --> 01:17:28,201 Oh no, it needs to play the... 1028 01:17:37,977 --> 01:17:40,845 [Chris] Are we seeing another short circuit 1029 01:17:40,847 --> 01:17:42,913 or is there something, what's a-- 1030 01:17:42,915 --> 01:17:45,916 [Commentator] I think it could be a kind of a misreading. 1031 01:17:45,918 --> 01:17:47,984 [Michael] And were you pretty comfortable 1032 01:17:47,986 --> 01:17:48,884 with saying that-- 1033 01:17:48,886 --> 01:17:50,886 [Chris] That it's looking good for Lee Sedol? 1034 01:17:50,888 --> 01:17:53,354 [Michael] Good for Lee Sedol. 1035 01:17:55,223 --> 01:17:56,422 Why is it playing in there for? 1036 01:17:56,424 --> 01:17:59,159 There's no reason for white to be playing that move. 1037 01:17:59,161 --> 01:18:01,194 It's a bad move and, in some cases, 1038 01:18:01,196 --> 01:18:03,829 it's going to lose a point too. 1039 01:18:12,872 --> 01:18:15,106 The whole game we thought that AlphaGo was wrong 1040 01:18:15,108 --> 01:18:17,274 about the board position and we were super worried 1041 01:18:17,276 --> 01:18:19,175 that it's gonna play garbage. 1042 01:18:19,177 --> 01:18:21,210 It's going to be like losing in a very embarrassing way. 1043 01:18:21,212 --> 01:18:23,445 And this continued for the whole game. 1044 01:18:29,184 --> 01:18:30,951 Well, that we don't know, but we're not good enough 1045 01:18:30,953 --> 01:18:32,018 Go players to know that, right? 1046 01:18:32,020 --> 01:18:32,818 Yeah. 1047 01:18:32,820 --> 01:18:35,354 As it turned out, none of us know Go well enough 1048 01:18:35,356 --> 01:18:37,356 to accurately judge what AlphaGo is doing. 1049 01:18:39,292 --> 01:18:42,260 This one looks like white is winning. 1050 01:18:42,262 --> 01:18:43,861 Aah! 1051 01:18:43,863 --> 01:18:46,930 We all say some of AlphaGo's moves are so weird 1052 01:18:46,932 --> 01:18:49,465 and strange and maybe mistakes. 1053 01:18:49,467 --> 01:18:52,234 But, after a game is finished, 1054 01:18:53,236 --> 01:18:57,172 we have to doubt ourselves, our judgment. 1055 01:18:57,174 --> 01:19:00,975 AlphaGo making another kind of nonsensical throw in. 1056 01:19:00,977 --> 01:19:02,375 We're not really sure what that's about. 1057 01:19:02,377 --> 01:19:04,912 This is what 10 or maybe 11 dan play looks like. 1058 01:19:04,914 --> 01:19:07,514 It looks weird and we don't quite understand it. 1059 01:19:07,516 --> 01:19:11,050 I think it is important to study more 1060 01:19:11,052 --> 01:19:14,553 about AlphaGo's mistake-like moves. 1061 01:19:14,555 --> 01:19:18,856 And maybe we can adjust our knowledge of AlphaGo. 1062 01:19:18,858 --> 01:19:20,357 To me the most amazing thing to come out 1063 01:19:20,359 --> 01:19:23,193 of my understanding of Go, as a result of watching 1064 01:19:23,195 --> 01:19:26,996 AlphaGo play, are the infamous slack moves. 1065 01:19:26,998 --> 01:19:29,231 Well, there's something strange about the way 1066 01:19:29,233 --> 01:19:31,032 it's playing because it's playing some moves 1067 01:19:31,034 --> 01:19:32,967 that are not really necessary. 1068 01:19:32,969 --> 01:19:34,201 Right. 1069 01:19:34,203 --> 01:19:37,103 A slack move is a move that looks lazy. 1070 01:19:37,105 --> 01:19:39,171 You can see these other better moves 1071 01:19:39,173 --> 01:19:40,872 and AlphaGo is rejecting them. 1072 01:19:40,874 --> 01:19:44,141 But what I think AlphaGo is teaching us 1073 01:19:44,143 --> 01:19:47,444 is that we've been using score as a proxy 1074 01:19:48,345 --> 01:19:50,245 for chance of winning. 1075 01:19:50,247 --> 01:19:53,214 So the bigger my margin of territory, 1076 01:19:53,216 --> 01:19:55,382 the more confident I am that I'm gonna win. 1077 01:19:55,384 --> 01:19:57,283 And AlphaGo is saying no, no, no. 1078 01:19:57,285 --> 01:19:59,118 It shouldn't matter how much you win by. 1079 01:19:59,120 --> 01:20:01,220 You only need to win by a single point. 1080 01:20:01,222 --> 01:20:03,388 Why should I be seizing all this extra territory 1081 01:20:03,390 --> 01:20:04,588 when I don't need it. 1082 01:20:04,590 --> 01:20:06,924 The lessons that AlphaGo is teaching us 1083 01:20:06,926 --> 01:20:10,827 are going to influence how Go is played 1084 01:20:10,829 --> 01:20:12,728 for the next thousand years. 1085 01:20:16,199 --> 01:20:18,266 By how much are we talking? 1086 01:20:18,268 --> 01:20:19,333 Two points? One and a half. 1087 01:20:19,335 --> 01:20:20,400 One and a half points. 1088 01:20:23,237 --> 01:20:24,503 Of course, Go is just a game. 1089 01:20:24,505 --> 01:20:26,571 But we could learn important lessons 1090 01:20:26,573 --> 01:20:29,307 from a computer being so successful at Go. 1091 01:20:29,309 --> 01:20:31,375 Machines will have the capability, 1092 01:20:31,377 --> 01:20:34,611 not only to crunch through huge amount of data, 1093 01:20:34,613 --> 01:20:37,347 but also to analyze it intelligently. 1094 01:20:37,349 --> 01:20:40,116 Just as in the case of the Go games, 1095 01:20:40,118 --> 01:20:43,552 the machine made moves that surprised even experts. 1096 01:20:43,554 --> 01:20:46,421 And, eventually, the machines will gain our confidence 1097 01:20:46,423 --> 01:20:48,589 because we will see that very, very often 1098 01:20:48,591 --> 01:20:53,026 they make a better guess than we could have made as humans. 1099 01:21:23,357 --> 01:21:26,525 Unfortunately for Lee Sedol, I think white 1100 01:21:26,527 --> 01:21:29,494 might have a slight advantage here. 1101 01:21:36,935 --> 01:21:38,101 Who told us that? 1102 01:21:38,103 --> 01:21:39,168 Someone, I don't know. 1103 01:21:52,381 --> 01:21:54,382 Winning, okay, not won. 1104 01:21:54,384 --> 01:21:56,550 That's what she said. 1105 01:22:16,170 --> 01:22:17,436 No, not really. 1106 01:22:24,476 --> 01:22:25,642 We don't know the score. 1107 01:22:25,644 --> 01:22:27,677 We have no idea what's going on, right. 1108 01:23:17,126 --> 01:23:19,727 (crowd clapping) 1109 01:23:24,499 --> 01:23:28,602 I think it just really is a once-in-a-lifetime thing. 1110 01:23:28,604 --> 01:23:33,273 I would say it's the most amazing thing I've experienced. 1111 01:23:33,275 --> 01:23:36,442 You know, for us, it's the culmination of a 20-year dream. 1112 01:23:36,444 --> 01:23:39,845 It started as a pure research endeavor. 1113 01:23:39,847 --> 01:23:41,313 We just wanted to understand 1114 01:23:41,315 --> 01:23:43,381 can neural networks play the game of Go? 1115 01:23:43,383 --> 01:23:44,715 And, from there, it went on 1116 01:23:44,717 --> 01:23:47,717 to a level that I never expected. 1117 01:23:49,386 --> 01:23:52,687 And I'm unbelievably proud of the team. 1118 01:23:53,755 --> 01:23:56,423 When I started doing artificial intelligence, 1119 01:23:56,425 --> 01:23:58,591 it was four or five years ago 1120 01:23:58,593 --> 01:24:00,693 and I was really interested in that. 1121 01:24:00,695 --> 01:24:02,695 But many people would discourage me. 1122 01:24:02,697 --> 01:24:04,563 They would say there is no future. 1123 01:24:04,565 --> 01:24:08,166 So actually seeing that within five years 1124 01:24:08,168 --> 01:24:11,702 we are at this stage right now it's amazing by itself. 1125 01:24:11,704 --> 01:24:16,139 Everybody say, (speaking in foreign language) 1126 01:24:17,407 --> 01:24:19,441 There are so many possible application domains 1127 01:24:19,443 --> 01:24:22,744 where creativity in a different dimension to what humans 1128 01:24:22,746 --> 01:24:25,613 could do could be immensely valuable to us. 1129 01:24:25,615 --> 01:24:27,581 And I'd just love to have more of those moments 1130 01:24:27,583 --> 01:24:29,182 where we look back and say, "Yeah, 1131 01:24:29,184 --> 01:24:30,616 that was just like move 37." 1132 01:24:30,618 --> 01:24:32,918 Something beautiful occurred there. 1133 01:24:32,920 --> 01:24:36,588 At least in a broad sense, move 37 begat move 78 1134 01:24:36,590 --> 01:24:39,857 begat a new attitude in Lee Sedol, 1135 01:24:39,859 --> 01:24:42,626 a new way of seeing the game. 1136 01:24:42,628 --> 01:24:45,628 He improved through this machine. 1137 01:24:46,696 --> 01:24:49,431 His humanness was expanded after playing 1138 01:24:49,433 --> 01:24:52,233 this inanimate creation. 1139 01:24:52,235 --> 01:24:55,603 And the hope is that that machine, in particular, 1140 01:24:55,605 --> 01:24:58,672 the technology behind it, can have the same effect 1141 01:24:58,674 --> 01:25:00,206 with all of us. 1142 01:25:01,308 --> 01:25:03,575 I remember hearing a talk by Kasparov 1143 01:25:03,577 --> 01:25:07,645 who says that a good human plus a machine 1144 01:25:07,647 --> 01:25:09,813 is the best combination. 1145 01:27:02,058 --> 01:27:05,126 (melodic music)