AlphaGo vs Lee Sedol: Game 2 – Video Commentary

The comments below are from the discussion during the live stream.

Related Articles


  1. Younggil An says:

    Hello everyone, I hope you’re looking forward to the 2nd game today!

    Unfortunately, real time comments were causing problems for our server with so many visitors, but you can leave a normal comment if you’d like to discuss the game with us. You’ll have to refresh the page from time to time to see new comments.

    • Piroon Rojmahamongkol says:

      We understand that all Go player around the world will come in this web.
      So many Thai Go player looking this real time game at this page.
      Thank you for everything you do about Go.

  2. Matthew Richard says:

    Time to show the machines who made who!

  3. Very exciting! Younggil do you think one day is enough time for Sedol to prepare a new strategy?

    • Younggil An says:

      Hi Wasi. I don’t think so… but let’s see. He might have found AlphaGo’s weaknesses through the game.

  4. Raymond Wan says:

    Is Myungwan Kim and Andrew Jackson doing commentary for the game?

  5. James Fury says:

    I agree with Matt Richard, we need to let these computers know that humans are still superior

  6. Younggil An says:

    Game 2 will soon be started. 🙂

  7. Completely agree!

  8. Younggil An says:

    It looks like Lee didn’t sleep well last night. However, I hope he will play differently today.

    • Guy Incognito says:

      He is nervous. He was blinking a lot before the game. If his nerves are shot, then he has already lost.

  9. Greetings Younggil, i have a question re:the match last night that puzzles me about Lee’s end game strategy. Lee was clearly ahead on time by a significant amount of mins with Deep Mind having just about 5 mins left. Why do you think that Lee chose not to win by letting time run out on the computer?

    • Younggil An says:

      I think Lee was relieved too early when he took the lead at around 93. He must have thought the game was already over, and he didn’t consider that he would lose his won game.

      He was surprised by AlphaGo’s correct moves on the right side and the bottom right, and the game was suddenly decided. There’s nowhere left for Lee to catch up with spending time afterwards.

  10. Zak Smith says:

    I look for Lee Sedol to play a little more conservatively in the opening, and to approach the middlegame fighting as his place to take control of the game. Hopefully he won’t make a severe reading mistake and end up with bad direction like in the lower right corner, which ended the game yesterday.

    I look forward to a good game, and a good match that is by no means over. I still think it is Alphago that has to prove itself.

  11. Anonymous says:

    What do you think of black 13?
    I´ve never seen it before.

    • Younggil An says:

      I’ve never seen that before either.

      Generally, Black plays at 13 without exchanging of 11 for 12. Let’s see what’s going on next…

  12. Younggil An says:

    AlphaGo’s opening is very creative, and Lee plays very carefully because he learned that AlphaGo’s opening sense is quite sensible from game 1.

  13. Zak Smith says:

    I agree, leaving the weak group on the lower side is unusual, and yet it seems difficult to find a way to attack it profitably for now. Perhaps the exchange of the tiger mouth followed by tenuki will become more popular after this match, and AlphaGo will contribute to the opening in go already??

  14. Zak Smith says:

    Oh, after I said it AlphaGo defended his position!! ;-P

  15. Zak Smith says:

    But doesn’t the peep become aji keshi after black defends with the extension on the lower side? White’s territory in the lower right corner becomes more solid, and there are no weak points, so he can plan ahead more easily with the thickness he has after connecting to the peep….

    • Younggil An says:

      Yes, peeping at Black 15 was too early, and is regarded as aji keshi in normal situation.

      By the way, it looks like Lee tries to play normally in this game, and he plays very seriously now.

  16. Zak Smith says:

    I think it’s best now to approach the upper right side and try to develop the right. If black answers without a pincer he might get a more flat position, but with the peep on the lower right corner it seems a pincer might not even be interesting for black.

  17. Jorik Mandemaker says:

    Oh AlphaGo played the diagonal attachment. Isn’t that usually considered bad in situations like this?

    • Younggil An says:

      Right, it was regarded as bad move old days. However, that’s modern style of play and it’s possible.

      Black considers the right side is already limited, so she emphasized the top more than the right side.

  18. Zak Smith says:

    Yes, An Younggil, Lee not only looks more serious, but he is playing very simply and it seems that black is making small mistakes, not just with the peep, but also kicking the white stone on the approach at the top right. I think that white is getting a thicker position, so the middlegame fighting may go better for him.

  19. Zak Smith says:

    Michael Redmond did a triple take when he saw the shoulder hit on the 4th line stone! Hilarious!!!

  20. Anonymous says:

    Black 37?? Wow^^
    Looks so strange.

  21. Anonymous says:

    Ear reddening move No. 2?^^

  22. Zak Smith says:

    Is anyone checking if Lee Sedol’s ears are reddening?

    • Younggil An says:

      Haha, it will be remembered as a very special move if she wins again.

      Lee was surprised to see that, and he spent quite a long time for his next move.

    • Freeman Ng says:

      He did leave the room…

  23. Jorik Mandemaker says:

    Alpha go is really displaying a unique sense of play right now.

  24. Zak Smith says:

    G4 seems forcing, but is it aji keshi again?

    • Younggil An says:

      It looks nice. It forces White to make his left side over concentrated.

      However, Black’s next move (connection) is also interesting.

  25. Zak Smith says:

    An Younggil, you call AlphaGo a “she” like Myungwan Kim?

  26. How would you judge the game so far? Even?

    • Younggil An says:

      It should be better for White, because Black’s moves are unusual. I think White is slightly ahead so far.

  27. Zak Smith says:

    Whoa!!! Suddenly the board is going to erupt in fighting. The middlegame has come upon us!!!

  28. Jorik Mandemaker says:

    What do you think about the result in the lower left. To me it looks like black didn’t really accomplish much.

  29. Anonymous says:

    Lee wins today

  30. Younggil An says:

    The game up to White 60 was better for White, because the fighting from Black 43 looked overplay, and Black’s stones got heavy.

  31. Black did not get much in the lower left. And it conflicts with the black shoulder hit. Cannot see any connection between the fight with the shoulder hit.

  32. Younggil An says:

    Yes, I totally agree with you Jeremy.

  33. Zak Smith says:

    I have the feeling that black is slightly ahead now. Lee seems to be playing in a submissive style.

    • Younggil An says:

      Yes, White 70 looks a bit small and passive. But the game looks still alright for White.

  34. Yes, at least he’s trying, but he’s not very good with words. Redmond seems bored by him.

  35. I wonder how many stones have been dropped so far xD Exciting game!

  36. north korean

  37. Raymond Wan says:

    Lee is using up his time. Seems like this may be a factor later on.

  38. Myungwan is in love.

  39. will lee won this ?

  40. Go Sei Gen says:

    Lee is leading.

  41. far or near guys score…

  42. Younggil An says:

    Black’s last move (81) is very interesting, and it looks nice. Black’s center becomes solid and strong, and the game will be more complicated at the top.

    White should try to settle at the top, but Black’s center is quite strong now…

  43. Zak Smith says:

    I like 81 because the white stone at the top will have difficulty living locally even with another move, so black is simply connecting her thickness and making preparations for the white group to run. I don’t know if white is in trouble, it all depends on the possibility of sabaki. If this group becomes heavy, it will seem to be difficult for white.

    • Younggil An says:

      Yes, Black 81 and 85 were extraordinary, which is difficult for human players to imagine.

  44. Younggil An says:

    White resisted with 86, and it’s Alphago’s turn to attack White’s group at the top.

  45. Younggil An says:

    The game is becoming playable for Black, because AlphaGo’s attack looks gentle and slow, but very well balanced.

  46. Younggil An says:

    It looks as if AlphaGo considers the game is good for her. Which is scary, because we learned that she’s very strong in the endgame and counting from game 1.

  47. Zak Smith says:

    Yeah, with the attack on the center group, black looks really good now.

  48. Younggil An says:

    Black’s cap (101) in the center looks sharp, and it’s not easy to fight against, because White’s group from the top is still weak.

  49. Zak Smith says:

    But white can prepared the cut of the center stones, so the fighting is balanced?

    • Younggil An says:

      Yes, the result in the center seems to be alright for White.

      Black played in the center pretty easily, and I’m not sure if AlphaGo thought Black is already leading or not.

  50. will alphago win again?

  51. who lead now sir?

  52. I believe Alpha uses middle area as the main attack which is almost no professor players does that because it may be too many calculation and too less historical record in professoinal players history, then Lee would not have enough time or experience to play against this new AI style of middle area offense game..

  53. tao hsiao says:

    It seems that AlphaGo will win again.

  54. Younggil An says:

    The game is still very close, and White’s next move is important.

    White jumped towards the center with 110, and the battle in the center will be very important for both.

    The 3-3 point in the top right corner is very big for both.

  55. Zak Smith says:

    I like the flow of moves for the last few moves by white. Nice way to settle while pressuring the lower center black stones. Lee seems to be aiming to take sente and get into the corner.

    • Younggil An says:

      Yes, White’s moves in the center was smooth, but Black started to attack White’s center group with 115, and the following battle will be interesting.

  56. Younggil An says:

    Black 117 on the left side looks strange, because Black fill his own liberties, but let’s see if there’s any special meaning from that exchange.

    • Eric Yoder says:

      When black played this kind of move it’s a signal that alphago is confident that either it is too far ahead or too far behind for the game result to change.

  57. its alpha go lead now guys?

  58. Computer is much stronger towards the end of the game since the possible moves left is much less and the computer can basically caulcate all possible moves and pick the best one. Lee cannot lead in early or middle game then Lee wining chance is low.

  59. Max Goldman says:

    Will Lee lose on time?

    • Younggil An says:

      Lee won’t lose on time, because he’s very good under the time pressure.

      White played bold move at the bottom instead of taking care of his weak group in the center, and it will be exciting to see if Alphago will attack White’s center group or not.

  60. Jorik Mandemaker says:

    Wow is the white center group safe enough to play at the bottom like that?

    • Younggil An says:

      Lee might have thought that. Black also has some weaknesses around the center, so Lee chose an extreme way in the game.

  61. Raymond Wan says:

    Oracle from burying the grave. That is scary.

  62. So guys, who is winning ?

    • Younggil An says:

      White has more territory at the moment, but Black can bully White’s center, and she has sente now.

      If there’s no special move in the center, the game will be slightly better for White.

  63. I do not know how to play Go. Hope someone can tell me who is winning in real time.

  64. go alphago…dont let sedol won game hahaha

  65. Younggil An says:

    The game is still very close, so it’s not easy to say who’s winning now.

  66. Anonymous says:

    Wow, endgame is quite nerve wracking. Now they suddenly move to the corner. What do you think younggil, is Lee corner invasion (and giving up the 4 stones in center) correct ?

    • Younggil An says:

      Hmm, hard to tell. It looks like Lee didn’t find a good move in the center, so he chose a big move in the corner (130). However, the game is very close now, and it doesn’t seem to be good news for Lee.

  67. Max Goldman says:

    Younggil, which one plays better in the endgame from your view?

    • Younggil An says:

      Lee Sedol is specialized in the endgame, but Alphago’s endgame is regarded as perfect…

  68. timidorchid says:

    alpha go has perfect end game so far.

  69. i think bcos a time so close …he cant view seriously …

  70. Anonymous says:

    Younggil, do you think C15 is a mistake of W?

    • Younggil An says:

      Locally, B15 is better, but Lee wanted to leave some bad aji in Black’s center area with 136. I can’t say it’s a mistake.

  71. timidorchid says:

    I think Lee sedol slightly behind now.

  72. Younggil An says:

    White’s jump at the top (142) is Lee’s nice move, and the game will become complicated, which is good news for Lee for now.

  73. Younggil An says:

    White can still save the four stones in the center if White comes back later.

  74. Anonymous says:

    Lose again

  75. who lose ?

  76. Younggil An says:

    It looks like White could play more strongly at the top, but Lee chose an easy move. It seems as if Lee thought the game is alright for him.

  77. Hitting that byo yomi

  78. May be Lee thinks if he doesn’t lose 0-5 is already a win

  79. A complete reversal time wise from the first game

  80. Younggil An says:

    Lee Sedol is in byo-yomi now.

  81. what u mean guys?

  82. I wonder if we’ll see AlphaGo enter into byo yomi as well?

    • Younggil An says:

      He’s 17 minutes left, so we will see AlphaGo’s byoyomi as well.

      By the way, Lee might think he’s not bad, but the game is very close.

  83. alphago will won again this?

  84. timidorchid says:

    what is the komi ?

  85. Anonymous says:

    Who is winning?

  86. In count down mode Lee just doesn’t have enough time to think, I would say the winning chance of Lee is less than 20%

    • Younggil An says:

      Yes, there will be more chances for Lee to make a mistake in byoyomi.

      AlphaGo is playing as if she’s already winning…

  87. timidorchid says:

    0.5 to 1 stone difference if komi is 7.5
    Lee sedol maybe behind by 0.5 to 1?

  88. timidorchid says:

    Hope for miracle, Lee sedol seems behind.
    Alpha go seems perfect so far.

  89. Younggil An says:

    It looks like Black can peep… yes, that’s not easy to respond…

  90. when the game end guys

  91. Younggil An says:

    That was a brilliant endgame tesuji for Black…

  92. Younggil An says:

    Black is winning now, and let’s see if Lee will find a way to catch up.

  93. deep blue 2.0 says:


  94. Younggil An says:

    Oh, it looks like Alphago just made a mistake!

  95. Younggil An says:

    The game can me reversed now.

  96. Go Leeeeeeeeeee!

  97. timidorchid says:

    Black win I think

  98. Anonymous says:

    what´s happening?????

  99. go alphago win this game pls….

  100. Anonymous says:

    Gooo LEE!!! win for human please~!

  101. Younggil An says:

    Lee could play more strongly at the top for White 172, but Lee might think he’s winning now.

  102. Michalson says:

    I hope, that Lee could still win.

  103. Lee Wang Hao says:

    Black is winning. alpha is soon going 2-0 up…..

  104. who win guys?

  105. Younggil An says:

    Black is still leading. Lee should have played more strongly at the top with a ko for 172.

  106. Anonymous says:

    Not close at all. Lee should resign asap.

  107. Anonymous says:

    are you sure ?

  108. Someone should notify them with a light or a beep if a move has been played. Frustrating for the viewer to see the move and watch them try to find what it was.

  109. I found the opening stages interesting. Alphago seemed to be overconcentrating white on the left and right, on a rather large scale. When black cut off the white stone at the top, it seemed that black was securely ahead.

  110. livestream sucks. It’s not possible to check the game’s flow while you’re at work…

    Isn’t it possible to use sgf-applets to follow the game?

    • Younggil An says:

      We’ll update the game records as soon as possible, so you can follow the game later.

  111. Younggil An says:

    Now, Black is ahead, and the game is going to end soon. White lose so many points in the endgame, and although White caught up a bit at the top, it wasn’t enough to reverse the game.

  112. I mean the above, are you sure Lee will resign ?

  113. Diffusion says:

    One question – the rules by which they play (time alotment, etc..) – is this normal, or favoring one of the sides? What I find best about this exchange is that the computer is bringing new moves to the game. It will enrich Go for a long time out.

    • Younggil An says:

      The rule is fair for both sides, and yes, I agree that the new style of moves will enrich Go.

  114. Harald K says:

    The old issue with programs: Once they’re confident they’ll win, they’ll freely throw away points because they don’t care about winning margin. AlphaGo is still a Monte Carlo program at heart!

  115. timidorchid says:

    If alpha go dont make error should be black win now.

  116. Lee would be mentally broken if he lost game 2. He said he would win 5-0 or 4-1!

  117. Michalson says:

    Aghrr, it is so thrilling!

  118. timidorchid says:

    lee sedol needs more time, maybe 8h game is better

  119. game end? who win guys?

  120. raidragon says:

    2 0

    I think Go will be the next Chess

  121. should be more sleep I think

  122. ayy lmao says:

    Pack it up boys, we’re done on this planet, the machines took over.

  123. Looking forward to the game record and your comments Younggil.

  124. Wow, that was really surprising. What a great game though.

  125. I would like to see Alphago tackle Dosetsu’s Masterpiece. That would be an interesting and different test of its skill.

  126. Guy Incognito says:

    He’s a twit.

  127. i liked your comment about “dismissing” 🙂 this is how you should behave against this ignorant arrogancy

  128. !Sleepy Eric.... says:

    OK, we don’t have avengers, no Superman, no Bruce Wayne, but surely the terminators are coming…

  129. AlphaGo (Black) won the second game.

  130. Anonymous says:

    surely the terminators are coming

  131. For many people watching, it seemed that white was leading for a while because nobody saw the tricky moves towards the center. Towards the end, it’s clear that Alpha knew it was winning and even made a suboptimal play just to simplify the end result. Is it possible that Alpha knew it would win from early on, and every move was a simplification because it didn’t try to win as much as it could’ve?

    • I think AlphaGo knows it’s winning from move 1 on 🙂

    • Diffusion says:

      No, I do not think so. Monte Carlo does try to optimize for total win. I think that the moves that are seen as errors were simply “sub-optimal” in terms of domination, but solidified its position. The search tree works by optimizing for a number of moves after the current moves, so its winning position will never be that decisive.

  132. Jem Falor says:

    I think someone needs to let him know.. the only time he can beat the AI is during the opening sequence. Since the first match, I’ve seen the AI made endgame moves during mid game. When I saw that, I know he lost the game. Don’t even need to count. The AI counts for you.

    The game gets less complicated for the AI as it hits near or way before mid game as the number of legal positions on board gets lesser..

    Start off with “perfect” opening sequence and continue the game with perfect, no mistakes, moves; only then you can beat the AI. Otherwise, you can resign in mid game..

    • Alexander Yuan says:

      Sedol seems to get the idea, judging by the response to the last question of the press conference just now.

      (via translator at that press conference) “As I was playing today’s game, I kind of felt that things would get difficult after the mid-game, reaching closer to the end-game. So I feel that for the next game I would have to focus quite hard on the beginning part of the game.”

    • In the post-game press conference, Lee Sedol actually said that the only way he can win is in the first half of the game, and he will focus intensely on it. He believes he can still win at least once. Can’t wait for the next match!

  133. Gil Dogon says:

    So what are the chances of a Ke Jie match next ? I think Ke should hurry to set a time when there may
    still be time for him to have a winning chance, as surely AlphaGo will only become stronger as
    time passes.
    Before this game and after the first he estimated his winning chance as 60%, now after this game
    maybe he can lower this to 50% . In the press conference Lee seemed like he resigned himself to
    be lesser than the machine, and said he felt behind throughout the whole game ….

    • Anonymous says:

      Ke Jie doesn’t stand a chance in even game play against AlphaGo. These two games against Lee are the mark of another level of play than current level of top pros. But I would like to see Ke Jie taken Black in a five game series. We will evaluate better the real strength of AlphaGo.

  134. Jem Falor says:

    there are probably only two person I know of that may have a chance against AI.

    Huang Yuetian/Longshi and Go Seigen..

    the end result for AI against Lee Sedol, imho, would be 5-0.

    • Anonymous says:

      While Longshi was the strongest in his time he could only match low professional nowadays. Same with go seigen.

    • bobiscool says:

      you realize that they are both significantly weaker than current pros?

  135. timidorchid says:

    Go Seigen opening and mid + lee chang ho end game.

  136. Was Go Seigen weaker in the opening than any other pro, living or dead?
    I really doubt it.

  137. timidorchid says:

    It is like Makoto suekuni and Tominaga Kenta both co-operating to draw with logistello (strongest othello progam at that time). That was many years ago. Btw Makoto and Tominaga are world champions.
    And humans need more time to calculate the perfect end game.

  138. Go openings today are degenerate: people rely too much on very closely calculated, specialized lines, as opposed to creating something from scratch. I can’t express how sick I am of seeing YAFCO (Yet another F—— Chinese Opening). Pros play what they know, what is safe, to win games. This is not salutary for the game.
    In chess today, openings that were out of favor for 100 years are seen to be perfectly playable. It is much more interesting for the spectator since computers came along.

  139. Jem Falor says:

    you can’t really compare the past with the present literally.. the concept of whole board strategy right from first move, I believe Longshi has that. There appears to be an inherent pattern that regardless of whether you had prior experience of it, you could still play something new just by basing on how the opponent moves. It’s something synonymous to mirroring. And I believe Longshi has that too.

  140. timidorchid says:

    I think if Lee sedol plays attacking go from the start it might throw alpha go out of book and deep into calculation

  141. Standard openings are for sure analyzed very well by Alpha. Maybe if Lee lost 3rd game he would play more experimental opening.

  142. Ke Jie actually said that Lee Sedol wasn’t the best guy to pit against AlphaGo. He has traditionally been weak in his opening moves; his forte is adopting a fighting strategy in the midgame and precise calculation and closure in the endgame to achieve his many victories. But these mirror AlphaGo’s stregnths and weaknesses too: strategically weak in the opening, strong in the midgame, and infallible in the closing stages. Lee Sedol will never be able to out-calculate AlphaGo in the latter part of the midgame, especially under time pressure, now that he’s a lot slower than before. What we need is a fast young player like Ke Jie who launch into a full-scale press right from the start to catch AlphaGo on the back foot, consolidate territory and minimize mistakes in the midgame, and leave enough time to finesse the endgame.

    • Anonymous says:

      AlphaGo is at least at the same level of play as Ke Jie and other top players. I don’t think the style or the person is the issue, but the handicap. I hope I will prove wrong by the last three games in this series, but so far it seems that AlphaGo is simply stronger than Lee Sedol or Ke Jie. By how much?

      • We won’t really know whether style is the issue until the Ke Jie / AlphaGo match though. It’s definitely true that all the qualities that are usually highlighted as Sedol’s strengths, are also traditionally the AI’s strengths. Even if AlphaGo would rout Ke Jie, the way it would do so would be completely different to its victories over Sedol.

  143. It feels like we are witnessing something much bigger than any of us can imagine.

  144. I don’t think AlphaGo has an opening book, in the sense you mean it; so playing unusual openings is more likely to play into his hands.
    What Lee Sedol needs to do is create some really complicated semeai, that involve lots of kos. Lots of groups on the board where life/death status is really fuzzy. As far as we can tell, AlphaGo hasn’t really solved the semeai problem, it has just got really good at everything else. (Of course, setting up such a game, without creating a losing position, is going to be hard; it may be that Aja and team at DeepMind are now able to steer the game away from its weaknesses.)

  145. Anyone seen or heard about a game where AlphaGo played against AlphaGo?

    • Gil Dogon. says:

      Alphago plays many thousands of games against itself, or earlier version of itself, that is how it improves, by a process called ‘reinforcment learning’ . I guess the day is not far where some of those games can be published and analyzed ….

  146. I don’t think semeai would be a problem for AlphaGo. What we haven’t seen are complicated ko fights and taking advantage of ko threats to eke out close gains., especially under Chinese rules area scoring where move sequence is crucial in filling in the dames. That might trip up AlphaGo.

  147. Anonymous says:

    Does anybody know what actual machine/computer alphaGo is, being present in the room with sedol during the matches? Sure, we know it is some computer involving multiple CPUs etc., but

    – during the games, does it run ‘standalone’? I.e., is it connected to the internet, or to some of the PCs of the alphaGo crew, or ‘offline’? I presume it must be ‘standalone’, but just want to make sure…
    – what power consumption does the computer have? I mean, a human body at full output can generate about 500 W in physical excercise (e.g. cycling). Mentally I don’t know, but maybe similar, maybe say max. 1 kW (?)
    So assume Lee Sedol’s brain uses 1kW at max (an estimated number…) for his calculations. How much is that for alphaGo? Probably a lot more (?)
    – If the power usage of man vs alphago is a lot different, I think we are not talking about a fair contest here. Either, the alphago computer should be limited to a corresponding max electrical power, or its consideration time should be scaled down, so that the products of Calculation Power * Thinking Time are equal for computer and sedol…

    • Anonymous says:

      What nonsense are you talking? By your logic, 15 year old can’t play with 40 year old, since its not fair.

    • Anonymous says:

      AlphaGo has a version for single computers and a distributed version that uses a network of computers. According to comments on another article here, the single machine version uses 48 CPUs and 8 GPU and the networked version uses 1202 CPU and 176 GPU. Presumably, the latter, which is said to be at least 25x stronger than the former, is being used here.

  148. Anonymous says:

    I think it is not about fair or not, win or lost. It is about what extent AI can do now.

  149. Gil Dogon. says:

    Curiously I did not see anywhere that Deepmind described the actaul Hardware resources used in the current match. In the match agains Fang, they had some thing like 1000 CPUs + 700 GPUs which would totaly consume probably at least 300Kilowats. If anything now the scale is even bigger. I am quite sure the computers are in some google/Deepmind data center, and only the moves are relayed by internet to Korea ..
    Contest is not meant to be fair anyway, as man and machine are different, until nowsow men had the edge, and now machine overtakes …..

  150. I am curious what do you think about: What will happen when all 10 top pros play against Alphago? They will be allowed to discuss and play the outcomes on their boards and they will have 10 minutes for each move to answer collectively. Or.. will we ever have a machine who will not be defeated even without time limits for human and where all top players again discuss each other… I am curious about these..

    • NanoThinking says:

      I think this is a very good test to see how strong AlphaGo is. The other question in my mind is, if AlphaGo can improve over time, will AlphaGo learning algorithm can perfect the Go game? i.e. it will never loss?

  151. Come to think of it, no ko fights in all 7 games against Fan and Lee. I really think that could be AlphaGo’s blind spot. There’s talk in Chinese go circles that there might have been a secret pact between Lee and DeepMind not to engage in ko fights, though I have faith Lee would never agree to such a ludicrous condition.

    Would like to see some complicated ko fights in the final 3 games.

    • I also that the lack of ko fights is very strange, especially as Sedol must know of AI problems with ko. My assumption is that AlphaGo is just very good at avoiding positions that might result in ko, pushing the player to risk too much to force a ko. But maybe that’s just what Sedol needs to do, if only to find out how much a ko against AlphaGo is really worth…

  152. Alain Dekker says:

    There are currently several dozen commercial chess programs (search for “computer+chess+rankings”) that play at a higher level than the current World Champion, Magnus Carlsen (search for “Magnus+Carlsen+FIDE”). These programs run on standard (and cheap) hardware. The AlphaGo program uses 100s of CPUs and GPUs and cost $100Ms to develop. The weighting factors used in neural networks are hard to transfer to other programs, so writing another AlphaGo-like program or getting a program of similar strength to play on cheap hardware is some way off.

    While Lee Sedol in some sense is representing humanity, he is already better than 99.9% of human Go players and should in no way feel embarrassed for what looks likely to be a big loss in this match. Lee Sedol is a great ambassador for Go…and humanity! The machine uses 1000s times his calculating ability and energy (not to mention money!). That machines would eventually conquer Go was a given since Deep Blue beat Kasparov. This is really about what AI can do and it really is very impressive. For all the big talk about “helping humanity”, the discussion needs to move quickly to who will control the machines. The huge economic benefit that specialised machines are bringing to humanity needs to be carefully spread so that all humans benefit…not just accumulated in the hands of an economic elite or a few shadowy technological giants.

    Congrats to Google…and good luck humanity! The changes that the likes of Google, Facebook and machine automation will bring in the next few decades will be as exciting and disruptive as the Industrial Revolution 250 years ago.

  153. We’ve just created a question here:
    on whether Sedol will win any of the three remaining matches. Very interested in the probabilities this expert audience would give!

  154. SoySauce says:

    Do you guys have any connection with the guys at Google? I’d be really curious to know what AlphaGo thinks the komi should be (as in, at what komi the win/loss ratio of black switches from >50% to <50% in an AlphaGo vs AlphaGo game).

    • yes, that was my question as well. google really has a chance to answer some vexing questions about go, of which this is one.

      They have a bot that can play at high level professional level, they can play games with opponents exactly the same strength, they can tally how many wins for each.

      So they really could answer close to objectively what a good komi is.

    • The problem is that any particular instance of AlphaGo is trained with a specific komi. It assumes a particular komi and optimizes win percentage, not points. If you train it on a different komi, it very well may be that AlphaGo would learn to use different strategies.

      I think there may be a way to do this (create versions of AlphaGo that minimize/maximize margin instead of probability?) but it is not as simple as one might think.

  155. W. Victor Zhang says:

    At this moment, Lee probably has little chance to match the computer. However, each of card or board games has its own restricted rules. When the rules are modified (not dramatically), most people’s experience could still valid but computer’s may not. Here is my thought to truly test human intelligence vs. artificial intelligence: Change the board size of GO then let Lee matches AlphaGo. The pre-assumption is that neither should not play on that board size before. I believe this would reveal true human intelligence.

    • Anonymous says:

      Victor, that was true for the Monte Carlo approach. With AlphaGo approach, once the board size increases, AlphaGo advantage over top pros increases as well.

  156. I am curious whether altering the game in some way would benefit the human against the machine. For example, would AlphaGo do as well in a one-stone handicap game against a weaker opponent? Or is AlphaGo optimized only for a very specific (i.e. zero handicap) game? This might be one dimension where the human sense of strategy could dominate. Or a game on 21×21 board, for example.

  157. David Tan says:

    I am a bit disappointed that Lee did not take too many risks, especially when the game was so even until the very end. Towards the middle to end game, he was very cautious. Maybe he felt that “creative” moves in the 1st day didn’t pay off, or maybe he was concerned that he did not have time for complicated mid or end game.

    For example, he could be more aggressive against the weak black center-left group in the mid game, or more aggressive on the top right in the end game.

  158. W. Victor Zhang says:

    看来李世石没什么机会了。然而,每个棋局或牌局都有特定的规则。如果改动稍稍规则, 人的经验可能还有用, 机器的就不一定了。为了测试真正的智慧, 我建议李世石和AlphaGo 在非标准的棋盘上比赛 比如18×20或17X17. 前提是他们都没下过这样的棋盘。这样人类的智慧就能体现出来。

  159. Kurta Viktor says:

    The same time (main and byo-yomi) conditions against Alphago?
    Smile on my face.

  160. The komi question can be settled conclusively once and for all – if we get Alphago to play against itself.

    It doesn”t matter if Alphago is optimized for win probability rather than winning margin, because its opponent (Alphago) is also optimized for percentage win, and both start with the same search tree and neural networks and computing power. So we have a self-controlled experiment where the only test variable is order of play, i.e. black or white.

    Start with a komi of say, 4.5 points, and have AlphaGo play one million games against itself. Note win/loss ratio. Then increase the komi stepwise to maybe 8.5 points. At some point the win ratio for black will drop below 50%, and we’ll have the answer.

  161. What about this one? As Almost know well, Lee couldn’t ahead Alphago’s calculation ability over the games. Then Lee should consider of Alphago’s mistakes several times by a unique way. Do not put down a stone sometimes… It might be causing a confuse to AlphaGo’s optimizing calculation.

  162. William Chang says:

    To be fair, Lee was under time pressure, 2hr speed-Go not enough to look ahead tactically. AI has no such real constraint — simply add hardware and electrical current to buy more time. Either limit AI to equal power consumption as the human (see below), or better yet, let humans collaborate to more evenly match the AI’s (purported) 48CPU+8GPU array processors.

    On energy fairness. For kicks I found various bits of info: Average brain uses 20% of caloric intake; Chess players during a tournament burn 125Cal/hr mainly due to stress (Lee probably burned more); Soccer players burn much more during a game; 1KCal=0.00116KWH; each processor under load ~ 200W (double if over-clocked); so it’s very roughly 0.000145KW vs 11KW (48CPU+8GPU) per hour or 1:80000. The networked version of AlphaGo using 1200CPU+170GPU would be 25 times higher at 1:2-million.

    • Let me defend the AI. Before the matches began, people were saying it’s a clean easy 5-0 for human. Now they are whining for time pressure unfairness. If there’s no time factor, let me tell you that computer will definitely outlast human.

      Human brain is actually far better than supercomputer with over 80k processors ( 20% of that would still be more than 1000 CPUs. However, human brain is a lot more efficient (take a lot less energy per computation). So, if you count processing power wise, human has capacity (but may not use all of them) to defeat 1000 CPUs pretty easily.

      So if you measure by energy, computer is unfair. But if you measure by given capacity of processing power, human brain is actually unfair.

      So I think people should stop saying about unfairness. If human have had win, no body would bring up about AI having too less processing power anyway.

  163. William Chang says:

    I came up with a simple mechanism to equalize first-player advantage in general classes of games:

    The player who has made the fewer moves has the option to make a free move subject to game-specific constraints. (Say a token changes hands when a double-move is made.)

    The purpose of the constraint is to preserve the nature of the game; for Go, the two stones in a double-move cannot be adjacent or reduce an enemy group’s liberties from 2 to 0. Komi will of course be eliminated.

    Whether AI will be bothered by the possibility of double-moves, remains to be seen.

  164. Well… I am so excited with game #3!!!
    Now, Lee should have some ideas what he is up to after failing 2 games…
    It seems to me that Alphago is capable of completing tree of probabilities in about later 1/2 of game if Lee gives respect to the machine…

    My disappointment is deepmind team ended up approaching Lee (and Go society) with so-called element of surprise, maybe lies… They shouldn’t have provided how Fan Sui played with Alphago when they planned to train Alphago to much higher level… It is like hinting Alphago would be armed with knife and then showing up the game with machine guns…ha ha ha
    They really deceived me (and probably Go players) in whole world by showing the games result with Fan Sui…

  165. I am shocked by game 2…Looks like the program’s stronger by each day!

  166. Jamee Donoghue says:

    Great suggestions , Just to add my thoughts , people a B22A – Chapter 7 , my colleagues filled a fillable form here