AlphaGo vs Lee Sedol: Game 3 – Video Commentary

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

Related Articles

Comments

  1. Younggil An says:

    Welcome everyone! Today we’re watching game 3 of the match between Lee Sedol and AlphaGo.

    AlphaGo is winning 2–0 so far, so Lee has to win today to have any chance of winning the match.

    Also, for Lee Sedol fans, I’m proud to announce our new book Relentless: Lee Sedol vs Gu Li is now available for pre-order. You can download two free chapters here.

    If you have any questions about the game (or the book) just let me know!

  2. The book looks good. 78 pages is quite the generous preview. I am definitely going to pick up a copy. Personally I’m pessimistic about Lee Sedol’s chances tonight, but we shall have to see!

  3. AnthonyC says:

    I hope Lee can win! Actually so far no one other than DeepMind team and Fan Hui has seen how to defeat AlphaGo (the previous version) so we don’t know what weakness AG has.

    • Younggil An says:

      Lee’s researched to find AlphaGo’s weaknesses, and let’s see if he’s found anything from last two games.

  4. Maybe trying out the Kobayashi opening?

  5. Younggil An says:

    It looks like Lee is going to play a moyo game today.

  6. Any thoughts on the opening so far?

  7. Anonymous says:

    Niceee

    We never saw Alpha Go manages a heavily attacked group
    How exciting!

  8. I wonder if Sedol focused on cutting edge opening theory, if he could extract some benefit

  9. White looks heavy

  10. Paintcan says:

    A fight at the start is the way to go! History unfolds!

  11. flameblade says:

    Note tht we have yet to see any ko fights so far…

  12. Paintcan says:

    G5 press? F7 bend now?

  13. Younggil An says:

    White’s group looks heavy, but blocking at White 26 looks right haengma.

  14. Alphago not afraid of hane at the head of 2 stones.I thought it would extend up

    • Younggil An says:

      Yes, right. If Black hanes at the head of 2 stones, White can double hane underneath, so Black extended on the 3rd line.

  15. Younggil An says:

    The opening so far is even, but it’s more likely Lee Sedol’s style of play compared to last two games.

  16. Jorik Mandemaker says:

    Interesting shape by AlphaGo. I wonder how this will turn out.

  17. Younggil An says:

    White 32 looks very interesting and sharp. It’s not easy to respond… so Lee will take some time for next move.

  18. Younggil An says:

    White 34 was another strong move, and it will probably become a fighting game from the beginning today.

  19. Anonymous says:

    W36 was unexpected though
    Probably this continuation was Alphago’s favorite?

    • Younggil An says:

      Yes, you’re right that W36 was unexpected.

      AlphaGo might have concluded that was the safest option compared to fighting with cutting at E10.

  20. timidorchid says:

    I feel human should be given more time like 8h vs computer 8h?

    • Younggil An says:

      Yes, I think so too. When Lee was in byoyomi in game 2, he had difficulties for counting under the time trouble.

      3 hours might have been better for human, because computer wouldn’t be better with 3 hours.

      • AI can search deeper, given more hrs would mean AI coming up with a lot more optimum move, if not best move.

      • In fact I am not sure whether it is true that ‘computer wouldn’t be better with more time’.
        I don’t know how to play Go, but I love Chinese Chess and I know a thing or two about computer programs.

        Alpha Go needs time to evaluate each situation, it takes time to do the search (for ‘optimal’ move) no matter how much it has ‘remembered’. It takes longer time when there are more possibilities and that’s why it gets stronger and stronger towards the end game.

        It will also use the time in opponent’s turn to keep evaluating different possibilities, so the longer Lee think at the start game, the bigger advantage Alpha Go has. On the flip side, if it runs out of time, it can only choose sub-optimal that can be found within time limit.

        So I believe if we cut the match to 1 hr maybe things can be different. But actually it maybe more fair to let human has 2 hr vs machine 1 hr, after all it’s a parallelized machine that has 1202 CPU and 176 GPU, according to a source online…

  21. Younggil An says:

    Lee’s play in this game was very strong so far.

    • Jorik Mandemaker says:

      Do you think the fighting is favorable to black? Or is white managing well?

      • Younggil An says:

        It looks like White is not bad, because Black’s area’s destroyed if Black doesn’t capture any of White’s groups.

        Black 47 was unusual, but another strong move to aim White’s weaknesses in connection.

  22. Younggil An says:

    White 48 was calm, and it’s time for Black to attack White’s group from the top. However, it’s not easy to seal White in, so Lee’s taking his time.

  23. Ryan Hamilton says:

    Is Lee in trouble with the clock?

  24. Mr.An, considering that Black had the first move, and had many stones in that area, is this still equal result? Who do you favor in this fight?

  25. David Ormerod says:

    I’m worried about Black’s push in the bottom left. It feels so heavy! Does Lee have some special plan in mind?

    • Younggil An says:

      The flow of the game is already favorable for White, so Lee tried to find a complicated fighting by pushing up.

      If Black can’t do anything special here, the game will still be good for White…

      • David Ormerod says:

        I thought attaching at White’s bamboo jump vital point instead was practically the only move. I guess Lee felt it wasn’t good enough.

  26. Uber​dude says:

    I think AlphaGo wins this one again, white groups are too resilient and black didn’t get any profit from the attack (just some top side, lost a lot on left side).

    • Uber​dude says:

      The calm inevitability of the deterioration of Lee’s position is scary.

    • David Ormerod says:

      I agree. I have this weird feeling in my stomach watching this game. To me it feels like playing an even game against Younggil…

      It always amazes me how a stronger player can gradually accumulate small efficiencies.

  27. Anonymous says:

    That black look thick and bad shape.

  28. Did Lee need to move out with the clump of stones? This game looks like it’s going to be a wipeout for white.

  29. Son Pham says:

    I’m amateur so I might be wrong but did Lee played too conservative in the last two moves? They don’t pressure white so much and White F13 seems to cut the two black groups

    • Younggil An says:

      Since the situation wasn’t easy for Black, Lee was trying to play more strongly than he does normally.

  30. M Irvan G says:

    Lee’s position in the battle looks not so good.

  31. My kyu level commentary – Lee Sedol is getting schooled by the computer sensei. This match will be decided soon and American spectators like me can go to bed early.

  32. Is that unusual to enter byoyomi since the mid game, at the pro level?
    How much impact on the play it is going to be to Mr.Lee?

    • Younggil An says:

      It’s not unusual to spend too much time at the pro level, but it’s not a good idea to be in byoyomi early against the machine. Black’s play isn’t going well.

  33. Younggil An says:

    Black’s moves are unnatural and struggling, that’s because the fighting is already bad for Black…
    White 54 was also strong, and it’s hard for Black to continue.

    If Black has to run away, White will be happy to befriend with its top group.

  34. Anonymous says:

    I feel like black cannot do much, and white has big potential at the bottom…

  35. With move G11 looks like Lee is resigning the fight 🙁

    • Younggil An says:

      Yes, right, but White’s left group isn’t that strong yet, so Black 57 was the vital point to still White’s eye shape.

  36. More kyu level commentary. F8 is a sad looking stone.

  37. Bottom and right corner scare me.

  38. Anonymous says:

    Isn´t 59, 60 a bad exchange for black?

  39. Son Pham says:

    This is so high-level stuff, F15 turns out to be a slacking move for White, and Black was able to save the thick group, now White has a bit of the liability on the left side.

    Still, the play has influence the left corner and bottom. Still seems to be a White game

  40. Anonymous says:

    Now problem is white bottom.

  41. Younggil An says:

    White 64 was calm, but Black 65 to 67 are relentless. Black has to fight because the game is still slightly better for White.

  42. It’s amazing to think that “slack moves” might just be “acceptable” from the point of view of AlphaGo.

  43. Younggil An says:

    Yes, if the sequence in the game is already in AlphaGo’s plan from 58, that’s amazing.

  44. timidorchid says:

    anyway can anyone feedback to team alpha go for more time like 8h for lee sedol instead of the current time?

    • David Ormerod says:

      The only way to do it would be to allow Lee to play the game for a few hours at a time over many days. If you get any top pro to play an eight hour game against AlphaGo, the pro will get tired and the computer won’t. That will just make things easier for the computer.

  45. How urgent is for Black the aji of a White move at B12 to start the ko? Do you think Lee will comeback there?

    • Younggil An says:

      That’s a good question. I don’t think Black has time to come back, because the game is still not good.

  46. Do you agree that F15 is slack ? I feel like the bottom is becoming huge.

  47. David Ormerod says:

    The computer wasn’t pushed hard enough before now to show its real strength, but in the beginning of this game after invading Black’s top left it seemed to play even more powerfully than before.

    Now that it has the advantage, I think it will play its relaxed (high probability of winning) moves again, and people will feel like it’s a close game, but it’s not. At this point, it seems like it’s just toying with Lee and has already won the game.

    That’s just my feeling about the game so far, but it’s quite scary if true. What if pros need to take handicaps to play AlphaGo?

    • I was also thinking about this, AlphaGo seems to be just at least a level stronger then Lee Sedol or Ke Jie – I mean after he is showing his power and gets ahead, he can afford to play such safe and slow moves and still win.

    • Younggil An says:

      Yes, that will be very scary if that’s true that it plays relaxed move because it’s already winning.

  48. Son Pham says:

    AlphaGo seems highly optimized. I don’t think it aimed to “humiliate” Lee by aggressively pushing on the black groups. Like F15 turns out to be quite slack, but F15 with G13 makes a wall for the bottom territory. G17 seals it.

  49. timidorchid says:

    Sadly lee sedol will lose today due to time pressure and end game errors.

    • Younggil An says:

      Yes, White has only one weak group (from the top), and it’s not easy to exploit.

      Black’s attachment at 77 looks interesting. Lee also knows that he won’t be able to win with normal moves.

  50. That would be tremendously sad, Lee Sedol loosing because of time.

  51. Anonymous says:

    Final fight start!!

  52. P13 is so exciting 🙂

  53. Alex Zhou says:

    I’m afraid I have to agree. It feels like Alpha Go has an advantage on the board, and that’s not even considering the time pressure Lee will be in.

  54. Son Pham says:

    AlphaGo has a “policy network” to determine which part to favor a move. Can Lee possibily “confuse” Go because what he really is doing is to attack the White group in the center?

  55. what about Gu li against this demoniac machine? I am thinking a very strong fuseki player…

    • Alex Zhou says:

      Given the impressive performance against Lee Sedol so far, I think Gu Li will surely lose to AlphaGo, as Gu Li is prone to making mistakes in endgame, whereas AlphaGo has virtually perfect endgame.

    • David Ormerod says:

      That’s a good idea jangalf. Catching up in the middle game and endgame (like Lee often does) just doesn’t seem to work against AlphaGo.

      Maybe they need a team of pros to challenge AlphaGo in a multi-day game though? 🙂

  56. David, I hope if Lee Sedol loses this game he will take a one stone handicap for the next two games. I want to see how strong this machine is!

    • David Ormerod says:

      Hey Andy! It would be interesting, but I’m not sure that it will happen this time. Lee might feel like it was humiliating with the whole world watching, and even if he didn’t other people would get upset about it…

  57. guo chen says:

    li’s strategy is to separate white into many small groups. Alphago is proved to be stronger in overall board position strategies. Local fight could be AG’s weakness. So today Li tried to split white

    • Younggil An says:

      I agree. White’s moves so far were fine, and let’s see what White will in the bottom right.

      White 86 was calm and solid, and Black was going to attack White’s center group after preparation.

  58. Where are you guys getting the move numbers? They aren’t shown on youtube.

  59. Hi what are the compensation points for white for making the second move ?.. Komi ?

  60. Son Pham says:

    Lee is setting up an attack on the center board? 4 stones are suiciding so that black can put pressure on the center and the top right. (I just repeat Mike here :p). Has AlphaGo felt for it?

  61. This fight is really exciting 🙂

  62. timidorchid says:

    aw I dont want to see Lee sedol lose by time pressure:(

  63. timidorchid says:

    gl Lee Sedol kill big dragon

  64. Younggil An says:

    White H3 (90) was the vital point of the eyeshape, and it’s not easy for Black to keep chasing.

    • Son Pham says:

      Mike has pointed out that move for so long. White is always ahead by that move. If it felt any threatened it can always just play that move. If black does so it gives white a chance simply extend the white group in the center. Also, Black has to save the bottom right group now.

    • Uber​dude says:

      Why didn’t white play there earlier (e.g. before that attach on right side). Looked like a good exchange to destroy shape in sente to me. Maybe because white would not simply connect but hane and hanging connection to the corner first?

  65. Hi, Younggil what you think if we have longer time controls , would they favour machine or human pro’s. ?
    thanks.

    • Younggil An says:

      Hi Ahsan. I heard that longer time (3 hours) would be beneficial for human pros. However, much longer time (more than 5 hours) is too much, because human will become tired.

  66. Mr. An, I know that this thought would have been preposterous just one week ago, but if AlphaGo wins all five games the way she won the first two games, do you think someone like Ke Jie would accept to take handicap in a series against AlphaGo? Or do you think the professional would not want to lose face, and I will just ignore and dismiss AlphaGo?

  67. Younggil An says:

    Hmm, I don’t know, but pros won’t take handicap against AlphaGo at least now. If Ke Jie wants to challenge, he will surely want to play as even.

  68. Younggil An says:

    Black lived in the bottom right corner from 91 to 95, and White’s turn to complete its big territory at the bottom.

  69. timidorchid says:

    hope lee sedol can kill some groups,he is behind 🙁

  70. Thanks for you answer An…. Ok just curious are you alos pro… can i ask your rank…
    thanks agains..

  71. *also

  72. timidorchid says:

    @Ahsan, I think in internet go, the gap between pro and amateur is very small,
    there are many inseis who beat 9 dan pro too, and there are quite alot of people playing online, so it is hard to tell but An has pro skill and his analysis is good:) I like it.

  73. But if Ke Jie will want to play even after this event, this will look very humiliating to a legend like Lee Sedol… I mean, if Ke Jie says he can play even with AlphaGo, he will really imply that he can also crush Lee Sedol in the same manner as AlphaGo, which I think is very far from the truth as we have seen in the Mlily finals…

  74. hello, Mr younggil.
    do you think that some other pro with different style of playing than Lee Sedol could get a better result?

  75. Younggil An says:

    The game is still better for White, and it doesn’t look easy for Lee to catch up after watching AlphaGo’s last two games.

    White’s weakness is the center group, but since White blocked at H3 (90) that’s not easy to attack.

  76. ok… i just googled you and it turned out that you are 8d pro …. that is very high indeed….. its nice to have a conversation with you..
    thanks.

  77. ok, tounament finished. Alphago beat the Federer in Go, it should be prepared to play with the Djokovic in Go.

  78. timidorchid says:

    I think Lee Sedol may resign now 🙁

  79. Can Lee get much by attacking center ? It feels like, with AlphaGo’s near perfect yose, this game is almost over for Lee.

  80. Isn’t white easily winning now? The “weak” group is really thick and strong, and white’s territory on the bottom is huge.

  81. Son Pham says:

    Looking at the board situation right now, I actually think F15 turns out to be very strong. Remember back then there was a long two black group so that move threatened both black group. Lee was able to save it, but F15 and F13 set up a wall nicely for the territory in the bottom.

  82. Younggil An says:

    White’s atari in the top right (102) is generally a bad exchange, but AlphaGo might have thought that’s the best way to win.

  83. B2 ???

  84. timidorchid says:

    Maybe Lee changho wins alpha go in 24h game

  85. Does AlphaGo think it is behind?

  86. Younggil An says:

    White’s placement in the top left corner (104) is interesting and let’s see how strong AlphaGo is at life and death.

  87. Thanks for comments again An…….There is some 9d pro Cho Hyeyeon who is saying on live feed that if he is black he may already resigned here. is black’s position that bad ?

  88. Younggil An says:

    That’s tricky for Black to respond, because Black doesn’t want White to live at the top easily.

  89. If white lives easily, he won, right ?

  90. Younggil An says:

    Yes, it will be difficult for Black to catch up then.

  91. It seems like a move like B2 can increase the time pressure on black. But I think I heard them say that AlphaGo doesn’t pay attention to how much time its opponent has. That seems like a weakness in its game that a human opponent can exploit.

  92. I feel like Lee did not get much from attacking the top left group, which is surprising.

    • Younggil An says:

      Yes, right, that’s because White’s H3 was nice. Black should have turned before White played there.

      By the way, it’s Black’s turn to harass White’s right side group, because he’s still behind.

  93. I am afraid of the 3-0 coming…that would make me a bit sad :/

  94. Younggil An says:

    This fighting on the right side would be Black’s last hope, and the attachment at O11 (113) is unusual way of play.

    Let’s see what’s going on.

  95. This feels desperate…I’m gonna light a candle for Lee 🙂

  96. Younggil An says:

    Black L15 (115) seems to be a tricky play, but White’s response at M14 is calm and cold.

  97. Son Pham says:

    Mr. An, I would like to thank you so much for translating move to the coordinate format, makes it much easier to amateur like me to understand, even though I don’t think that’s the format that the go server follow. (Does it tho?)

  98. Jacob Tsimerman says:

    Thanks Younggil for your commentary. It makes the game much more enjoyable to watch 🙂

  99. Vu Duc Thong says:

    I’m surprised at L15, but at my amateur level I cannot see anyway for Lee to struggle for the lower territory. Time is running out also… such misery.

  100. Yes, a big thanks Younggil 🙂

  101. Son Pham says:

    Midgame doesn’t even end yet and Lee only has 6 minutes. He will be under tremendous time pressure now

  102. Younggil An says:

    Thanks guys for your cheerful message.

    By the way, the game is getting more difficult for Black…

  103. Yes, I think Lee is desperate to make something happen…

  104. I wonder if they can train this machine by making it play against another version of itself. Machine learning should create different development versions depending on its playing history. This contest is making the game of Go very interesting to a wider audience I think. Thanks for comments

    • That is how they have trained AlphaGo between its matches against Fan Hui in October in preparation for this match. Hundreds upon hundreds of games at a high level to help enforce its learning habits.

  105. It’s amazing that Alphago let Lee live on the left a long time ago forseeing that it would eventually get the bottom. Such confidence!

  106. Younggil An says:

    Black’s atari at C18 aims for a ko fight later. That’s very big, and make his own group safer as well.

  107. Son Pham says:

    The way AlphaGo takes 1 minute to make a connecting move is just cold. I would be so scared fighting against this cold inhuman optimizing game.

  108. they’re already doing that with AlphaGo’s reinforcement learning phase

  109. timidorchid says:

    Think Lee sedol has clearly lost but just experimenting moves to see alpha go’s response

  110. Ayan Nandy says:

    Can someone tell me what is the area of AlphaGo and Lee at this momemnt? I am a novice and don’t know how to calculate

    • Son Pham says:

      White has about 50-60 at the bottom and about 10 on the left side.

      Black has about 25-30 on top, 10 at the bottom left and 7 points at the bottom right.

      White is already winning about 20 points and the top black territory can be invaded.

      • Ayan Nandy says:

        When would you say an empty point is in white’s area? Or black’s area?

        • Son Pham says:

          Haha, it’s quite tough to say. I usually say visually. I usually imagine how the border eventually play out and then the area that the border surrounds will be come Black or White territory. We can only estimate because it is always unclear how a player makes moves

          • Ayan Nandy says:

            Is there any webpage where I can see the final territory of black and white marked for the Match1 and 2?

            • Since Lee resigns before it totally plays out, it’s hard to give a concrete answer. Only an estimate.

  111. time for a3

  112. Younggil An says:

    If Black doesn’t make something happening inside of White bottom territory, there won’t be any chances…

    • Ayan Nandy says:

      Younggil, how do I identify the territory of white or black? Is an empty cell in the territory of the nearest stone’s color? What if the two or more nearest with different colors?

      • Younggil An says:

        Oh, it’s not easy for me to answer Ayan… I hope somebody will explain to your question.

      • Son Pham says:

        You must surround the territory for the it to be counted. For example, P15 is neither territory because neither Black and White surrounds it fully. S1 is quite for sure Black territory because there are lots of Black stone arround

  113. Do you think it is possible ?

  114. He doesn’t even have time to read…

  115. Younggil An says:

    It doesn’t look possible to live, but it can be a help to make the game more complicated. If White responds accurately, it will still be hard…

  116. I think Lee will resign soon.

  117. Younggil An says:

    Black L17 (127) is the same meaning, and White’s response at J17 is accurate. Black should keep on playing here since there’s nowhere to fight.

  118. Younggil An says:

    Lee Sedol (Black) is in byoyomi at 129 (J15).

  119. I would never have imagined 3-0 as a result…Not it my dreams

  120. Younggil An says:

    Lee Sedol is in last byoyomi, and White’s responses so far is perfect.

  121. Even if black lives small is that enough to win?

  122. Younggil An says:

    Black H18 (131) is another tricky move to respond well, and it will be the last fighting in this game.

  123. Younggil An says:

    That’s not easy for White to play accurately, because Lee’s moves are flexible and resisting.

  124. Son Pham says:

    J18 only serves to stall time

  125. Younggil An says:

    It looks like White made a mistake with J17 (138), and it will be a ko. It will be hopeful news for Black!

  126. Younggil An says:

    Lee didn’t accept the ko fight, but if White reinforces its weakness, what Black can do?

  127. Go Lee 🙂
    Too bad he is in his last byoyomi :/

  128. Younggil An says:

    It will still be a ko, but it doesn’t look good enough for Black.

  129. Son Pham says:

    If Lee won this fight it will become one of his masterpiece

  130. Son Pham says:

    Will AlphaGo be able to handle 2 ko at the same time? That will be interesting.

  131. Younggil An says:

    Wow, White I2 (148) looks strange, but let’s see what’s AlphaGo’s plan is.

    It seems as if AlphaGo hate fighting ko, and that’s interesting.

  132. Maybe he thinks the ko is not worth the top ?

    • Younggil An says:

      I’m not sure. It might have thought even if Black plays one more move, that’s still possible to fight.

      White has more ko threats, so it’s not easy for Black to fight the ko.

  133. Black could have resigned 50 moves ago.

  134. AlphaGo seems to be taking longer to think now.

  135. Seems like this is the end.

  136. Son Pham says:

    Jeez, connecting D17 is the kind of ko threat that AlphaGo makes. That’s scary

  137. Son Pham says:

    I feel like AlphaGo isn’t even afraid of playing neck-and-neck with one point in the hardest way possible.

  138. T.T

  139. Requiescat in pace, Homo sapiens

  140. Anonymous says:

    It will be easy 5-0 for AlphaGo. If Kie Ji plays AlphaGo, it will be another easy 5-0.

  141. Anonymous says:

    Relatively Strong Amateur here. (7/8 Dan Tygem) I can’t say much more about what just happened.
    It’s astounding.
    Alpha Go is probably the best player ever.

  142. I feel a little bad that Lee Sedol can’t review the game with AlphaGo as he is used to doing with other pros.

  143. Anonymous says:

    Wow…

  144. Paintcan says:

    History is made. Hopefully AlphaGo will be opened up so humans can study with it in the near future!

  145. Quite a few comments on the three matches so far have been questioning certain moves Lee Sedol made, and that suggests an idea. Now that Alphago has demonstrated it can beat top individual Go players, how about pitting it against a ‘Human Cloud’? That is, a group of humans analysing the game and selecting moves, not just one. It could be an invited group of the highest ranked pros, or even a real-time internet challenge against ‘the world’: any online Go player anywhere on the globe could be part of the Human Cloud team.

    The human moves would be selected in much the same manner as Alphago does it, by a ranking system. The trending proposed moves can be seen by all and the move that gets the most votes is chosen. There should be some time element so the choice is made before a deadline, but the time period would be flexible depending on the progress of the game (obviously it would be much shorter under time pressure in the endgame). The time period would allow discussion and analysis of possible outcomes of various moves before the deadline arrives and the most popular move is selected.

    This would be a kind of Human AI! One person is fallible, having a ‘style’, certain propensities, and prone to errors that could be stress related, but having the Human Cloud make the choice of moves by consensus should eliminate obvious mistakes, throw up audacious brilliant moves and generally make a far more formidable foe for Alphago than any single individual. It would advance the game exponentially. Plus it would be an enthralling and inclusive spectacle that could turn into a series, especially if the Human Cloud turns out to be equal or even superior to Alphago.

    To me it seems the obvious next challenge after this. If Alphago still beats all humans working together then we know we really are f****d!

    • I don’t think this would ever happen. There’s a reason why they didn’t do the same thing in Chess.

      The actual obvious challenge is to give pros a handicap, which I don’t think will happen anytime soon.

  146. I would say that Alphago could give any human 3 stones, if it were programmed differently with regards to komi.

  147. Anonymous says:

    But where were Lee Sedol mistakes?? They both played a very strong game, but Lee Sedol was somehow clearly behind from the beginning..
    What just happened?

    • Son Pham says:

      I don’t think Lee made any serious mistakes in this game. He attacked very heavily on the left side, but the black stones got dense and became a liability for Lee, which backfires him the entire match.

      I think the only serious mistakes Lee made were a conservative move during that attack, which makes the Black dense. That’s one mistake and cost him the entire match.

      AlphaGo did make a “mistake” of letting those black stones go with F15 move, but it turned out to set up nicely for the bottom territory, which is 50-60 points.

      • That actually worked out very nicely. Maybe the computer sees the “slack” move as not so slack?

        • Son Pham says:

          Looking back, I think AlphaGo clearly knows that it wasn’t a slack move. But the thing that’s ultra scary is that AlphaGo is so comfortable making that kind of big and uncertain trade off.

  148. Jem Falor says:

    any chance for Lee Changho against AI?

  149. Jem Falor says:

    im particularly interested to see his expression throughout the game..

  150. William Chang says:

    It is conventional wisdom that to beat AI one must win the strategic contest from the very beginning, and not get drawn into tactical dog-fight. But here it is the AI that’s winning opening variations, by virtue of having explored brute-force more opening variations than the sum of human knowledge. _That_ is AlphaGo’s real strength.

  151. Alexander Valkovich says:

    Since it’s general purpose AI (more or less), people should try to teach it all the 4 disciplines supposed to be taught to the elites of ancient China: calligraphy (image recognition), martial arts (robotics), poetry (natural language processing) and Go. All in one. Some sort of synergy is to be anticipated I guess!

  152. William Chang says:

    Ha ha.

    Seriously though, AI should excel in all of the above arts, they all involve design i.e. optimization: Art is harmony from disharmony, equilibrium of in-equilibrium.

    Downstroke and across, world lines and ways scribing/ Cosmic time and space, in spirit and mind crossing/ The day’s warmth a gentle wind brings/ By orchid’s quiet scent contemplating.

  153. Needs a pro who is solid in the opening to have a chance – the AI is too strong in the middle and end game. Time is also an issue – complicating matters favours the AI

  154. Anonymous says:

    David Ormerod: yes, that’s right =)

    medlar: exactly. i doubt there’s anyone atm. imho, only Huang Longshi and/or Go Seigen may have a chance at it.

  155. And now we just need to give AlphaGo the hability to feel the pride of victory to balance the pain of defeat that feels Lee Sedol. I feel equally happy and sad about this development of IA

  156. Unbelievable yet believable, well, I think prolly someone like Go Seigen could revenge

  157. Gil Dogon. says:

    I really feel for Lee today. Somehow from his words and his voice in the press conference I felt that he had a really bad day today. He opened by apologizing, I guess to his fans, and he probably felt that he did not play well today, and was harsh upon himself. He essentially admitted to cracking under the pressure that he put himself under. Still he was really too harsh with himself, and I really hope he can play better and the remaining two games ….

  158. I feel that as a chess player also, I have some perspective different from go players.
    Go —even professional go at the highest levels—is quite obviously not a game of accurate play.
    The great players have an incredible sense of possible moves, but they can’t possibly evaluate all choices, as there are too many. The program is not handicapped in terms of move selection. This was very clear with the shoulder hit in game 2, for example. I believe the gap between pro and computer will be several stones before long, solely due to more accuracy of play.
    I didn’t get the sense that Lee played badly at all today. If he did make a serious mistake, I think it was very early, before more 30. That opening sequence on the left simply wasn’t good for black.

  159. I wonder what would happen if AlphaGo were to play on a 23×23 board without having any previous experience or analysis of it? A human would be able to adapt without much difficulty but the computer would find it very difficult. This would be an even greater test of AI, to be able to transfer skills to something it’s never done before.

    • Zaw Lwin you don;t seem to understand how AlphaGo was made …
      You would only need to put AlphaGo training in [a]×[a] board size even crazy [a]×[b] by playing millions of millions of games against itself end you will have the same results…
      also keep in mind because Top pros have train or there life in 19×19 opening theory the my had even bigger trouble playing on 23×23 boards or larger.

      • Anonymous says:

        I don’t think I made my point clearly enough. What I meant was for AlphaGo to play against a top pro on a 23*23 board, but *without* letting AlphaGo analyse any games against itself on this new sized board. I realise this will require considerable changes in the code but it is interesting to think to what extent an AI could be asked to transfer its self-taught skills on a 19×19 board to a different sized board it’s never played on before.

        • Zaw Lwin says:

          Above post from ‘Anonymous’ is from me. What I’m pointing out is that if AlphaGo is not able to play well on a 23×23 board, without letting it play against itself millions of times on it first, then that is a clear limitation on it’s intelligence whereas a human pro would still play a pretty decent game on a new board size despite never having played on it before. Humans still have the advantage in transferring learned skills to new situations.

          • David Stone says:

            Transfer learning is still a very hard problem in AI. This might be a relatively easy test case for transfer learning, but I suspect you are right, that humans would adapt much better compared to today’s AI.

  160. Furthermore, the AI can only get stronger – the day will come where it will be able to analyse every move from move one and play the perfect game. Game over for humans sad to say.

    • David Stone says:

      That is not possible. Mathematically, there are more board positions than atoms in the universe, so you would need a computer as large as the universe to analyze them all. But it is still getting stronger, definitely.

  161. I forgot to add that Lee Sedol now needs to take a handicap stone

  162. nhat minh vo says:

    oh no !
    lee se dol lose

  163. Anonymous says:

    Haha here we are
    Look at all those weak kyu rated amateurs in the world underestimating top professionals

  164. William Chang says:

    This from a non-player, beg your pardon.

    I thought move 23 Lee should have taken G13; with move 30.G13 AlphaGo effectively blocked him off for good. Is human Go out of balance favoring edge-play tactics?

    Perhaps Go players should also learn the connection game Hex (yet unconquered by AI), to re-emphasize connecting to the center. I invented a hexagonal Go/WeiQi variant called Weilian (“surround-connect”) that mixed Go surround edge-play and Hex connect center-play, and in these 3 games I really felt an affinity to AlphaGo which clearly put more emphasis on this than Lee did.

  165. So flameblade, you’ve had your ko fights, satisfied ? 😉
    Strange idea to think that a ko fight may has been a weakness for AlphaGo.

  166. Younggil, is there any chance that Lee Sedol will take a handicap or is that something for the future do you think? I’m worried that one of my favourite players has been well beaten and doesn’t look like winning one game.

  167. Paul Webster says:

    Though it might be hard to accept – I’d like Lee Sedol to see it as a great learning exercise.
    The euro player who
    Was beaten in October has continued to work with the AlphaGo team and says that as a result his ranking has jumped 300.

Speak your mind