AlphaGo and AI progress
March 8, 2016 3:32 PM Subscribe
The first game between AlphaGo and Lee Sedol is scheduled to begin at 23:00 EST tonight, in Seoul. There will be a livestream with commentary in english. Since the Deepmind Go-playing computer's previous victory against Fan Hui, Go professionals and AI researchers alike have had time to consider what it means.
Myungwan Kim 9p, with Andrew Jackson, provides a video review of the earlier games. British Go Journal has some comments (full article pdf). Everyone seems to agree that when AlphaGo played Fan Hui, it would not have been strong enough to defeat Lee Sedol. Nobody knows how much stronger it's gotten since then. Both sides seem confident of victory.
Demis Hassabis, CEO of DeepMind, gives the 2016 Strachey Lecture explaining further his company's approach to artificial intelligence. The second half of it, starting at 25 minutes, centres on AlphaGo.
Miles Brundage attempts to estimate how much of AlphaGo's strength over other Go programs is due to it being able to use more computing power than has previously been thrown at the problem, and makes the distinction between general AI systems and general approaches to narrow AI.
If you want to learn to play Go before the livestream begins, Interactive Way To Go is one place to start. If you prefer 1990's video based learning, GoeBasics3.mov is another.
Myungwan Kim 9p, with Andrew Jackson, provides a video review of the earlier games. British Go Journal has some comments (full article pdf). Everyone seems to agree that when AlphaGo played Fan Hui, it would not have been strong enough to defeat Lee Sedol. Nobody knows how much stronger it's gotten since then. Both sides seem confident of victory.
Demis Hassabis, CEO of DeepMind, gives the 2016 Strachey Lecture explaining further his company's approach to artificial intelligence. The second half of it, starting at 25 minutes, centres on AlphaGo.
Miles Brundage attempts to estimate how much of AlphaGo's strength over other Go programs is due to it being able to use more computing power than has previously been thrown at the problem, and makes the distinction between general AI systems and general approaches to narrow AI.
If you want to learn to play Go before the livestream begins, Interactive Way To Go is one place to start. If you prefer 1990's video based learning, GoeBasics3.mov is another.
So, after Go, what's the next test that we'll surely need Real Artificial Intelligence to beat, only to find out years later you can actually beat it with a few tricks and the slow but steady march of Moore's Law?
posted by ckape at 3:53 PM on March 8, 2016
posted by ckape at 3:53 PM on March 8, 2016
Top Trumps?
posted by dng at 3:55 PM on March 8, 2016 [1 favorite]
posted by dng at 3:55 PM on March 8, 2016 [1 favorite]
Inspired by the ballad of John Henry
Well Lee Sodol was a little baby
Sittin' on his daddy's knee
He picked up a stone and
a laid down an Atari
And cried, "Go's gonna
be death of me, Lord, Lord
Go's gonna be the death of me"
Now the Google he
said to Lee Sodol
"I'm gonna bring that
AlphaGo around
I'm gonna bring that
alohaGo out to play the game
I'm gonna knock that
Atari on down, God, God
I'm gonna knock that
Atari on down"
John Henry told the Google
"Lord a man ain't noth' but a man
But before I let that AlphaGo
beat me down
I'm gonna die with a stone
in my hand, Lord, Lord
I'll die with a stone in my hand"
posted by humanfont at 4:01 PM on March 8, 2016 [6 favorites]
Well Lee Sodol was a little baby
Sittin' on his daddy's knee
He picked up a stone and
a laid down an Atari
And cried, "Go's gonna
be death of me, Lord, Lord
Go's gonna be the death of me"
Now the Google he
said to Lee Sodol
"I'm gonna bring that
AlphaGo around
I'm gonna bring that
alohaGo out to play the game
I'm gonna knock that
Atari on down, God, God
I'm gonna knock that
Atari on down"
John Henry told the Google
"Lord a man ain't noth' but a man
But before I let that AlphaGo
beat me down
I'm gonna die with a stone
in my hand, Lord, Lord
I'll die with a stone in my hand"
posted by humanfont at 4:01 PM on March 8, 2016 [6 favorites]
So, after Go, what's the next test that we'll surely need Real Artificial Intelligence to beat, only to find out years later you can actually beat it with a few tricks and the slow but steady march of Moore's Law?
The republican nomination?
posted by lumpenprole at 4:31 PM on March 8, 2016
The republican nomination?
posted by lumpenprole at 4:31 PM on March 8, 2016
So, after Go, what's the next test that we'll surely need Real Artificial Intelligence to beat, only to find out years later you can actually beat it with a few tricks and the slow but steady march of Moore's Law?
The turing test?
In the year 2026- "I mean, it's not a *real* AI, sure, but it has a really innovative language simulator, and the way they're doing neural networking for goal selection is getting so much better, and you just have too feed in enough data sets and making a real-enough presentation of emotions is easy, and..."
posted by DGStieber at 4:54 PM on March 8, 2016
The turing test?
In the year 2026- "I mean, it's not a *real* AI, sure, but it has a really innovative language simulator, and the way they're doing neural networking for goal selection is getting so much better, and you just have too feed in enough data sets and making a real-enough presentation of emotions is easy, and..."
posted by DGStieber at 4:54 PM on March 8, 2016
There is a chess-like game called Arimaa that was designed to not be playable with proficiency by a computer.
posted by Windopaene at 5:06 PM on March 8, 2016
posted by Windopaene at 5:06 PM on March 8, 2016
The turing test?
3/4 of people on IM can't plausibly pass as human.
posted by eriko at 5:17 PM on March 8, 2016 [8 favorites]
There is a chess-like game called Arimaa that was designed to not be playable with proficiency by a computer.
Checking Wikipedia:
Checking Wikipedia:
Every year since 2004, the Arimaa community has held three tournaments: a World Championship (humans only), a Computer Championship (computers only), and the Arimaa Challenge (human vs. computer). In 2015, the challenge was won decisively by the computer (Sharp by David Wu), with top players agreeing it was no fluke but proof that computers are currently superior to humans.posted by fredludd at 5:19 PM on March 8, 2016 [4 favorites]
.. except that in 2015, [the AI] Sharp clinched victory in each of the three mini-matches by winning the first six games, finishing 7-2 overall and winning the Arimaa challenge. More background on why the game was hoped not to b easy for computers is over at the page Computer Arimaa, basically a high branching factor and too many opening positions for computers to have opening books. Arimaa's branching factor of 17,000 is higher even than go's 250.
In contrast to chess and go, humans didn't have a hundreds-of-years head start on computers in Arimaa. I suspect that is part of why the Arimaa challenge was won so quickly.
posted by the antecedent of that pronoun at 5:21 PM on March 8, 2016 [1 favorite]
In contrast to chess and go, humans didn't have a hundreds-of-years head start on computers in Arimaa. I suspect that is part of why the Arimaa challenge was won so quickly.
posted by the antecedent of that pronoun at 5:21 PM on March 8, 2016 [1 favorite]
In contrast to chess and go, humans didn't have a hundreds-of-years head start on computers in Arimaa
Yeah, seems like there's a danger of just making a game so complicated that neither computers nor humans can play it effectively...
posted by thefoxgod at 5:29 PM on March 8, 2016
Yeah, seems like there's a danger of just making a game so complicated that neither computers nor humans can play it effectively...
posted by thefoxgod at 5:29 PM on March 8, 2016
Isn't the Turing Test inextricably entwined with a behaviorist theory of mind? If so, why does anyone still take it seriously?
posted by thelonius at 5:30 PM on March 8, 2016
posted by thelonius at 5:30 PM on March 8, 2016
Nobody should, thelonius.
posted by Joseph Gurl at 7:40 PM on March 8, 2016
posted by Joseph Gurl at 7:40 PM on March 8, 2016
Isn't the Turing Test inextricably entwined with a behaviorist theory of mind? If so, why does anyone still take it seriously?
You should absolutely take the Turing test seriously -- not as a practical test, but as a thought experiment: a reductio ad absurdum against the viewpoint that machine intelligence is inherently impossible. The point of the test is that if you're happy to ascribe sentience to a human you've only interacted with via text, and you can't tell the difference between that human and a computer, then denying that the computer is sentient is special pleading.
I don't really understand what you mean by "inextricably entwined with a behaviorist theory of mind". Accepting the Turing test argument doesn't mean that you can't talk about an AI's internal mental processes, or experience of the world. It just means that you shouldn't need to understand those processes before you're willing to call it sentient.
posted by teraflop at 8:51 PM on March 8, 2016 [15 favorites]
You should absolutely take the Turing test seriously -- not as a practical test, but as a thought experiment: a reductio ad absurdum against the viewpoint that machine intelligence is inherently impossible. The point of the test is that if you're happy to ascribe sentience to a human you've only interacted with via text, and you can't tell the difference between that human and a computer, then denying that the computer is sentient is special pleading.
I don't really understand what you mean by "inextricably entwined with a behaviorist theory of mind". Accepting the Turing test argument doesn't mean that you can't talk about an AI's internal mental processes, or experience of the world. It just means that you shouldn't need to understand those processes before you're willing to call it sentient.
posted by teraflop at 8:51 PM on March 8, 2016 [15 favorites]
I know just enough about Go to be a Lee Sedol fan, and I'm honestly astonished to see a computer doing this well against him. What suspense! How unfortunate that I'm up way past my bedtime!
posted by Jeanne at 9:19 PM on March 8, 2016
posted by Jeanne at 9:19 PM on March 8, 2016
Mr. Kim said a bit ago that Lee Sedol looks to be ahead; he's a bit concerned about the aggressive move by white at R10.
posted by isthmus at 10:14 PM on March 8, 2016
posted by isthmus at 10:14 PM on March 8, 2016
Looks like Lee Sedol resigned. No official announcement yet. It was very close right until the end.
posted by isthmus at 11:33 PM on March 8, 2016
posted by isthmus at 11:33 PM on March 8, 2016
"Top human players rely heavily on intuition and feelings to choose among a near-infinite number of board positions in Go, making the game extremely challenging for artificial intelligence."
In other words, humans were better at go all this time because they felt better than AIs. I don't know whom to feel more insulted for, our dumb emotional human go masters or our cold, calculating robot soon-to-be overlords.
posted by J.K. Seazer at 1:49 AM on March 9, 2016 [1 favorite]
In other words, humans were better at go all this time because they felt better than AIs. I don't know whom to feel more insulted for, our dumb emotional human go masters or our cold, calculating robot soon-to-be overlords.
posted by J.K. Seazer at 1:49 AM on March 9, 2016 [1 favorite]
I don't really understand what you mean by "inextricably entwined with a behaviorist theory of mind".
You really see no resemblance?
posted by thelonius at 1:54 AM on March 9, 2016
You really see no resemblance?
posted by thelonius at 1:54 AM on March 9, 2016
Argh, what is up with the audio/video of the livestream? This is painful to watch with it skipping and cutting all the time.
posted by vernondalhart at 1:57 AM on March 9, 2016
posted by vernondalhart at 1:57 AM on March 9, 2016
In Player of Games, Iain M Banks says that purely deterministic games, (solvable games? if you like) like Chess and Go, which we currently hold as the highest status games are very much looked down on by the Culture, because Minds could of course out play anyone. Truly interesting games needed elements of randomness, chance and luck and it's how the player reacts to these elements that makes the game interesting.
That's a pretty prescient thought for 1988.
posted by Just this guy, y'know at 2:43 AM on March 9, 2016 [2 favorites]
That's a pretty prescient thought for 1988.
posted by Just this guy, y'know at 2:43 AM on March 9, 2016 [2 favorites]
Interestingly, reinforcement learning hit human-level performance at backgammon four years later.
posted by tss at 4:03 AM on March 9, 2016
posted by tss at 4:03 AM on March 9, 2016
Lee Sedol concedes here, saying: “I think the failure at the beginning continued throughout the game until the end. [...] I didn’t know AlphaGo would play such a perfect game. [...] Later I was thinking that it was actually a difficult game for both players, and AlphaGo made a winning move that humans can never make".
posted by progosk at 5:48 AM on March 9, 2016 [1 favorite]
posted by progosk at 5:48 AM on March 9, 2016 [1 favorite]
In Player of Games, Iain M Banks says that purely deterministic games, (solvable games? if you like) like Chess and Go, which we currently hold as the highest status games are very much looked down on by the Culture, because Minds could of course out play anyone. Truly interesting games needed elements of randomness, chance and luck and it's how the player reacts to these elements that makes the game interesting.
Right, but we're not very good at those either. A computer can easily beat you at Rock Paper Scissors, for example.
posted by leotrotsky at 6:29 AM on March 9, 2016 [1 favorite]
Right, but we're not very good at those either. A computer can easily beat you at Rock Paper Scissors, for example.
posted by leotrotsky at 6:29 AM on March 9, 2016 [1 favorite]
Top human players rely heavily on intuition and feelings to choose among a near-infinite number of board positions in Go
As Michael Redmond said during the commentary, intuition sometimes comes from an out-of-focus memory. Whether it's put there by personal experience or by genetic inheritance, it's something in our convoluted neural networks that suggests a good move, but the reasons for it being there aren't accessible to our rational conscious mind. Since AlphaGo has no rational self-conscious side to speak of, in that way all of its moves are based on intuition.
posted by sfenders at 6:32 AM on March 9, 2016
As Michael Redmond said during the commentary, intuition sometimes comes from an out-of-focus memory. Whether it's put there by personal experience or by genetic inheritance, it's something in our convoluted neural networks that suggests a good move, but the reasons for it being there aren't accessible to our rational conscious mind. Since AlphaGo has no rational self-conscious side to speak of, in that way all of its moves are based on intuition.
posted by sfenders at 6:32 AM on March 9, 2016
In other news, humans remain undefeated against computers in the game of Rugby.
...but give it time.
posted by leotrotsky at 6:37 AM on March 9, 2016
...but give it time.
posted by leotrotsky at 6:37 AM on March 9, 2016
Right, but we're not very good at those either. A computer can easily beat you at Rock Paper Scissors, for example.
Speak for yourself, human.
posted by Jairus at 6:41 AM on March 9, 2016
Speak for yourself, human.
posted by Jairus at 6:41 AM on March 9, 2016
Hunh. I really thought computer dominance on a 19x19 board was years and years away.
posted by Zed at 8:40 AM on March 9, 2016
posted by Zed at 8:40 AM on March 9, 2016
Wow. This is a seismic result. Clearly all that sniffing about "Oh he's just the European champion" by the Go cognoscenti (not at MeFi) last time was misplaced.
posted by RedOrGreen at 9:37 AM on March 9, 2016
posted by RedOrGreen at 9:37 AM on March 9, 2016
Feng-Hsiung Hsu, one of the scientists involved in the Deep Blue project, suggested that a world-champion computer Go player was about ten years out ... in 2007.
posted by Skorgu at 3:57 PM on March 9, 2016 [2 favorites]
posted by Skorgu at 3:57 PM on March 9, 2016 [2 favorites]
This one's not "world champion" just yet you know. I mean it probably will be, but in the mean time it's got some work to do. First it's got to win some more games with Lee, and then Ke Jie among others will probably want to have a go. Maybe people will figure out its weaknesses. Or maybe it will teach us whole new concepts beyond anything imagined before. It's up there with the best human players, but not yet proven to be superhuman.
posted by sfenders at 6:30 PM on March 9, 2016
posted by sfenders at 6:30 PM on March 9, 2016
I seem to have stumbled into tonight's live feed with no commentary if anyone wants it. Although I'm probably going to switch back to the other one as it gets into the part of the game I have no hope of understanding without professional help.
posted by sfenders at 8:47 PM on March 9, 2016
posted by sfenders at 8:47 PM on March 9, 2016
I have this dilemma -- thank you sfenders -- of being compelled to watch this for hours each night (and strangely for me, remembering that it's on) despite understanding the game at the level of a five year old. Last night I didn't even know how to interpret the board at the end. "Why is the upper left largely empty, how can all that area not be worth playing??" Etc. The stream is a soothing alternate universe of analysis of things I don't understand.
posted by sylvanshine at 10:05 PM on March 9, 2016 [1 favorite]
posted by sylvanshine at 10:05 PM on March 9, 2016 [1 favorite]
For anyone watching these games live, @AskAKorean is running a live translation of Korean-language commentary.
posted by isthmus at 11:05 PM on March 9, 2016
posted by isthmus at 11:05 PM on March 9, 2016
"Sometimes I have persistent delusions that I can't fix." -- Michael Redmond, Match 2 commentary
posted by isthmus at 12:07 AM on March 10, 2016 [1 favorite]
posted by isthmus at 12:07 AM on March 10, 2016 [1 favorite]
Watching Michael Redmond's commentary is like watching the game stuck in a short time loop. He makes a possible play, and bam -- it's on the board either from Alphago or Lee Sedol.
posted by TemporaryTurtle at 12:12 AM on March 10, 2016
posted by TemporaryTurtle at 12:12 AM on March 10, 2016
I think Redmond said that it's a clear win for AlphaGo at this point.
posted by isthmus at 12:15 AM on March 10, 2016
posted by isthmus at 12:15 AM on March 10, 2016
Oh FFS. "The next match is on, Friday, is that right? I don't know." You don't even know when the next match is?? Is there anything positive that Chris Garlock offers to the commentary?
posted by J.K. Seazer at 12:31 AM on March 10, 2016
posted by J.K. Seazer at 12:31 AM on March 10, 2016
I switched between various commentators, and it seemed like all of them kept changing their minds as to who was winning, even though they could not identify any real mistakes from either side until very near the end. Looks increasingly like Lee Sedol may have some chance to become the last human to win a single game from the strongest computer.
posted by sfenders at 12:39 AM on March 10, 2016 [1 favorite]
posted by sfenders at 12:39 AM on March 10, 2016 [1 favorite]
Ke Jie does indeed want a go. I believe that article was published before match 2 but I'm not 100% confident.
posted by Skorgu at 5:38 AM on March 10, 2016
posted by Skorgu at 5:38 AM on March 10, 2016
This is not un-reminicent of the accounts of the matches between Chinook and the terrifyingly superlative checkers master Marion Tinsley between 1990-94. My memory is fuzzy, but I don't think there was any machine learning thrown at Chinook, except possibly an attempt at improving its position evaluation function. It was otherwise just a straightforward program with a large opening book database, a very efficient mid-game min-max search function, and a massive end-game solving database (which eventually extended all the way to the start of the game by 2007).
Nonetheless, Tinsley apparently relished the competitiveness of Chinook. It was one of the only opponents that ever challenged him (he was reputed to have only lost 7 games in a 45-year career, one of them a casual game with a friend while drunk). Chinook, unlike any human player he faced, wasn't scared of him. Almost everyone he played against, in formal competitition or not, would play to attempt a draw, and count themselves very lucky if they could manage it. Chinook, not knowing who it was playing against, always played to win, and this really tickled him.
Chinook's creator, Jonathan Schaeffer, wrote a very readable book about the program's development and the matches against Tinsley. It mixes layman-friendly explanations of the algorithms involved with the personal story of his meetings with Tinsley, and I'm quite fond of it. Incidentally, David Silver and Aja Huang, both very involved in AlphaGo's development, apparently came up through Schaeffer's games group at the University of Alberta's computer science department.
posted by figurant at 9:12 PM on March 10, 2016 [3 favorites]
Nonetheless, Tinsley apparently relished the competitiveness of Chinook. It was one of the only opponents that ever challenged him (he was reputed to have only lost 7 games in a 45-year career, one of them a casual game with a friend while drunk). Chinook, unlike any human player he faced, wasn't scared of him. Almost everyone he played against, in formal competitition or not, would play to attempt a draw, and count themselves very lucky if they could manage it. Chinook, not knowing who it was playing against, always played to win, and this really tickled him.
Chinook's creator, Jonathan Schaeffer, wrote a very readable book about the program's development and the matches against Tinsley. It mixes layman-friendly explanations of the algorithms involved with the personal story of his meetings with Tinsley, and I'm quite fond of it. Incidentally, David Silver and Aja Huang, both very involved in AlphaGo's development, apparently came up through Schaeffer's games group at the University of Alberta's computer science department.
posted by figurant at 9:12 PM on March 10, 2016 [3 favorites]
Kottke posted his impressions of the games so far.
Nah, I think "creative move" will be the day AlphaGo spells out "I'm trapped!" on the game board.
posted by polymodus at 4:43 PM on March 11, 2016
Nah, I think "creative move" will be the day AlphaGo spells out "I'm trapped!" on the game board.
posted by polymodus at 4:43 PM on March 11, 2016
New Yorker piece on the games so far.
posted by crazy with stars at 6:00 PM on March 11, 2016 [1 favorite]
posted by crazy with stars at 6:00 PM on March 11, 2016 [1 favorite]
3 wins for AlphaGo now. I for one welcome...yadda yadda yadda.
Anyone know what kind of watch Lee Sedol is wearing?
posted by Literaryhero at 12:23 AM on March 12, 2016
Anyone know what kind of watch Lee Sedol is wearing?
posted by Literaryhero at 12:23 AM on March 12, 2016
Looks like 1:00 pm Seoul = 8:00 PST, on:
Sunday, March 13 Seoul = Saturday 12th PST
Tuesday, March 15 Seoul = Monday 15th PST source
Is there a photo of his watch? The commentator mentioned it but I'm curious but couldn't find an image.
posted by polymodus at 1:05 AM on March 12, 2016
Sunday, March 13 Seoul = Saturday 12th PST
Tuesday, March 15 Seoul = Monday 15th PST source
Is there a photo of his watch? The commentator mentioned it but I'm curious but couldn't find an image.
posted by polymodus at 1:05 AM on March 12, 2016
Hey, he took one! GO SEDOL GO!
Very pleased about this.
posted by Joseph Gurl at 1:51 AM on March 13, 2016
Very pleased about this.
posted by Joseph Gurl at 1:51 AM on March 13, 2016
Some of interesting points from skimming the gogameguru comment thread:
- the computer tends to be referred to in the feminine (is this common elsewhere?)
- it's at risk of being redubbed "Avocado" (after a Michael Redmond stumble)
- Hassabis clarified his tweets about the computer's "thought" and "realisation" about moves 79 and 87 in this match: “When I say ‘thought’ and ‘realisation’ I just mean the output of #AlphaGo value net. It was around 70% at move 79 and then dived on move 87"
posted by progosk at 3:26 AM on March 13, 2016
- the computer tends to be referred to in the feminine (is this common elsewhere?)
- it's at risk of being redubbed "Avocado" (after a Michael Redmond stumble)
- Hassabis clarified his tweets about the computer's "thought" and "realisation" about moves 79 and 87 in this match: “When I say ‘thought’ and ‘realisation’ I just mean the output of #AlphaGo value net. It was around 70% at move 79 and then dived on move 87"
posted by progosk at 3:26 AM on March 13, 2016
"The result “W+Resign” was added to the game information."
posted by progosk at 3:30 AM on March 13, 2016 [2 favorites]
posted by progosk at 3:30 AM on March 13, 2016 [2 favorites]
Haha that is rad, progosk.
posted by Joseph Gurl at 4:17 AM on March 13, 2016
posted by Joseph Gurl at 4:17 AM on March 13, 2016
After a bad mistake early in the last game, AlphaGo makes a comeback to win the series 4-1.
posted by progosk at 6:11 AM on March 15, 2016
posted by progosk at 6:11 AM on March 15, 2016
Where can I review these games and step through them in my browser? Preferably with commentary notes, but I"ll take anything.
posted by Theta States at 10:06 AM on March 23, 2016
posted by Theta States at 10:06 AM on March 23, 2016
You can download the SGF files or view the games interactively through a Flash interface here. No commentary, sorry.
posted by J.K. Seazer at 11:12 AM on March 23, 2016
posted by J.K. Seazer at 11:12 AM on March 23, 2016
« Older Cyclists chased by an ostrich | depressed binge HoC rinse repeat Newer »
This thread has been archived and is closed to new comments
If you want to be able to appreciate the game that will be livestreamed, learn Go several years ago.
posted by fredludd at 3:42 PM on March 8, 2016 [16 favorites]