AlphaGo vs Lee Sedol, who will win?

[Simba]

Lol - I don't mean his hands were shaking before he played when he was counting, but that he was shaking when he actually played the stones. He certainly wasn't doing this all the way through.

One of the key moments for me was Lee's bad timing of the s13 atari. If there's one thing bots are good at it is deciding if your move is really sente or a bluff and playing something else more important first. I hope he can put this loss behind him and not be so careless in the next game. He needs to maintain concentration for every move, the bot is merciless at punishing fake sentes.

Do you feel that Lee wouldn't have played S13 if he'd known it to be (temporarily) gote? That move is huge even if gote and has potential life and death applications for both groups. I feel personally that R4 instead of Q5 would be more sensible, since then black would make some territory in the corner rather than white, but it's possible I'm just thinking too naively with this.

What do you think of black's R17 attachment? It's 'bad' in this joseki, but do you think maybe the pros are just missing something by claiming this? It'd be really cool if some new joseki were discovered with AlphaGo's help.

[Uberdude]

Do you feel that Lee wouldn't have played S13 if he'd known it to be (temporarily) gote? That move is huge even if gote and has potential life and death applications for both groups.

Yes, it's big (maybe there is even a chance those 3 captured white stones make a dead shape?). I can only speculate how seriously Lee thought about a white teunuki, it would be interesting to hear his comments on it.

I feel personally that R4 instead of Q5 would be more sensible, since then black would make some territory in the corner rather than white, but it's possible I'm just thinking too naively with this.

q5 makes some sense following s13 in that it aims to separate the big white wall group from the corner and ask where its eyes are (as s13 threatens q15 which takes away those eyes and possibly kills the whole thing). But the problem was white settled in the corner in sente and then came back to answer s13 at the top right to make eyes. But yes r4 to make corner points was probably better as most commentaries say. Given that white answered q5 at q6 (rather than r4) I think p6 at r4 to get the corner (and leave white with more cuts that direct r4) would have been better. Commentary also said p6 was bad (the losing move says An Younggil) and should block the corner, but the problem is white has the fine tesuji at s5 to both defend the r6 cut and threaten to kill the corner with s3 in the variation following move 128 r4 r3 p6 p2 s5 s3 so can then return to the top right at t14 or whatever. Still should be better than the game though.

What do you think of black's R17 attachment? It's 'bad' in this joseki, but do you think maybe the pros are just missing something by claiming this? It'd be really cool if some new joseki were discovered with AlphaGo's help.

Locally speaking I still think it is bad, but the problem is it is not as bad for white as white getting the j16 pincer is for black. Younggil says move 15 q13 (the standard punishment) was questionable. So I guess black should have played j16 himself and endured the suffering of white r14? In fact I think this opening shows that Lee's unusual move 7 at r8 (presumably attempting to make a fuseki AlphaGo hadn't seen in the training) was bad and mini-Chinese or something normal would have been better (r8 also ended up inefficient to my eyes following the pushes, and then the brialliant r10 much later embarassed it). Kasparov's comment about humans playing suboptimal moves attempting to exploit or trick AIs is spookily prescient here:

Early chess machines had blind spots and exploitable weaknesses and the temptation is to target these instead of playing a normal game. I could not resist doing so against Deep Blue. Mind sports like chess and Go require intense concentration and when focus is disrupted by trying to trick a computer you can end up tricking yourself into making objectively dubious moves. As machines get stronger, these are punished.

[handa711]

What are the pros and fans saying about this loss?

[jeromie]

I wasn't suggesting to never resign (indeed I often give up and resign rather early in my games). If he played out and lost this game by 5 points, and the next by 3, then resigning the 3rd when he is losing by 5 would be more warranted as AlphaGo would have demonstrated it can successfully finish a game (it did weird things at the end of Fan Hui game 1). Zen19X 7d lost a game on KGS by 0.5 because it filled in its own territory too many times (presumably due to some bug in the scoring or komi logic) and whilst I expect DeepMind have tested this thoroughly I still think it could be worth verifying. Winning in such a way would certainly be cheap, but by treating the AI with the respect he would a human he could be hiding its crucial weakness. Also didn't Sakata often play out big losses, I recall someone posting anecdotes here about how he wanted to know how many points his mistakes were worth.

I agree for this reason. But then the one aspect of go I might be better at than Lee Sedol is exploiting bots around my own level.

[Charles Matthews]

Also didn't Sakata often play out big losses, I recall someone posting anecdotes here about how he wanted to know how many points his mistakes were worth.

He did. He could do pretty much what he liked, once he had held the top seven Japanese titles at once. He also had a victory ritual, sending up a smoke signal; so I suppose if he was playing out the endgame against you and wasn't smoking, you might have had the consolation that he knew he was losing and was dying for a cigarette.

But in this case I don't imagine Lee was thinking that debugging AlphaGo's endgame was part of his contract. In the live feed, he put a white stone on the board quite quietly (for Chinese rules he would have to take from the opponent's bowl?); and started to discuss the game with someone behind him. Which is presumably what he wanted to do most, absent an opponent to relate to.

[Charles Matthews]

In fact I think this opening shows that Lee's unusual move 7 at r8 (presumably attempting to make a fuseki AlphaGo hadn't seen in the training) was bad and mini-Chinese or something normal would have been better [...]

Yes, this seems to me to be the story of the game: Lee played to take AlphaGo completely ex-book, and the plan failed. If he has a team/seconds to support him, and if this plan emerged from discussion with them, then it was naive about the likely strengths of the software. But maybe he just wanted to try something new.

Anyway, hindsight is usually a great help. Will be interesting to see what AlphaGo does with Black. I'd have thought Lee would be well off introducing complex corner variations. But perhaps, with White, set up a game of competing frameworks.

[Uberdude]

Commentary also said p6 was bad (the losing move says An Younggil) and should block the corner, but the problem is white has the fine tesuji at s5 to both defend the r6 cut and threaten to kill the corner with s3 in the variation following move 128 r4 r3 p6 p2 s5 s3 so can then return to the top right at t14 or whatever. Still should be better than the game though.

Actually, I wonder about the following bizarre move to prevent the s5 tesuji:
128 r4 r3 p6 r6! s6 p2 r7 q15.
By putting in the r6 cut, which at that point white will atari from underneath, before defending at p2 black forces white to defend the cut with the r7 capture (which is a loss of about 2 points compared to if white defended at s6 directly) and that way black can take sente to play q15. There is still the s3 s2 t2 ko which is a pain. This kind of cutting and not extending feels very unprofessional and wacky, but if he can live (sort of) in sente to get q15 isn't it better for black? And that ko is not easy to fight for white: because black gets q15 white is short of eyes and it looks like the whole white dragon could die if black wins the ko.

[pookpooi]

Nature editor write that knowing AlphaGo confident rate while the match is still playing make him uncomfortable to look at Lee Sedol struggle from midgame to the end... while DeepMind team is smiling all the way through the match and amused when professional commentator voiced their different confused opinions... (I add more spice from the original news just for fun)

http://www.nature.com/news/the-go-files ... er-1.19544

[Kirby]

Apparently this is more of a matter of whether one knows Korean.

I disagree, but don't have anything to add. Like I said, I am not trying to convince anybody.

[mhlepore]

Nature editor write that knowing AlphaGo confident rate while the match is still playing make him uncomfortable to look at Lee Sedol struggle from midgame to the end... while DeepMind team is smiling all the way through...

That's why I voted for AlphaGo in the poll (3/2 margin). If you play a bot online, you get this suffocating feeling from the middle game on, because you know the bot is unlikely to make mistakes. If AlphaGo has a slight lead heading into macro-endgame, it probably isn't correct to say "things are close."

[pookpooi]

That's why I voted for AlphaGo in the poll (3/2 margin). If you play a bot online, you get this suffocating feeling from the middle game on, because you know the bot is unlikely to make mistakes. If AlphaGo has a slight lead heading into macro-endgame, it probably isn't correct to say "things are close."

I'm more aggressive than you, I vote for 5-0 for AlphaGo (would have bet if I've money). I rely on Lee Sedol's personality that will keep him to continue underestimating AlphaGo and lost all of his five game. But I think he realized now that AlphaGo is not a cheap trick so the result may vary but I still favor AlphaGo to win at least 3 matches.

[pwaldron]

Is there a copy of the game record posted anywhere?

[DrStraw]

Is there a copy of the game record posted anywhere?

https://gogameguru.com/alphago-defeats- ... ol-game-1/

[Spaceman-Spiff]

Also: https://online-go.com/review/114161

[belikewater]

In a news article I saw Lee quoted as saying that AlphaGo played a move "I thought was almost impossible to play". Does anyone know which move he was referring to?