PredictionBook is now read-only ( read more ).

AlphaGo defeats Lee Sedol in 3/5 or more games in March

Created by Jach on 2016-01-28; known on 2016-04-01; judged right by Jach on 2016-03-12.

  • Jach estimated 5% on 2016-01-28
  • davidmanheim estimated 70% and said “Given that they won 5/5 against a (much weaker) European champion in October, and since they have had 6 months to upgrade, this seems very reasonable.” on 2016-01-28
  • daniel74f estimated 40% on 2016-01-29
  • ChristianKl estimated 80% and said “I don’t think the European champion is much weaker. The difference between 9 dan pro and 2 dan pro isn’t 7 stones. ” on 2016-01-29
  • Raahul_Kumar estimated 65% on 2016-01-29
  • Raahul_Kumar said “@David: Difference between even 1 dan pro and 7 dan amateur is huge. But difference between 2 dan and 9 dan pros is tiny.https://en.wikipedia.org/wiki/Professional_go_handicaps” on 2016-01-29
  • ChristianKl said “It’s worth nothing that AlphaGo not only won all five games in the match against Fan Hui, Fan Hui resigned in 4 of them.” on 2016-01-29
  • PseudonymousUser estimated 65% on 2016-01-29
  • JoshuaZ estimated 72% on 2016-01-29
  • themusicgod1 estimated 51% on 2016-01-29
  • PlacidPlatypus estimated 60% on 2016-01-29
  • Jach said “Ke Jie 9p predicts 5% or less. Elo calculations taking into account the 2/5 preliminary round wins by Fan Hui put the p single game at 13%, 3/5 games at 1.8%…” on 2016-01-29
  • mortehu estimated 80% and said “The informal games were run with shorter time limits, and DeepMind is putting $1M on the table” on 2016-01-29
  • gwern estimated 33% and said “plot NN dan vs time on chart, looks adequate. 30 days to train alphago entirely, ~140 days between Hui & Sedol matches leaves a few cycles of improvement and more if Google’s throwing in at least $1m of GPU time having gone so very public” on 2016-01-30
  • gwern said “There’s also the Hui games being kept secret and DeepMind apparently deciding on this very aggressive timeline on their own, which implies they think they can win or at least make a credible showing based on their improvement curves” on 2016-01-30
  • ioannes estimated 60% on 2016-01-31
  • m1el estimated 80% and said “Lee Sedol may use an anti-bot tactic, but AlphaGo is likely to imporove in a few months.” on 2016-01-31
  • PseudonymousUser estimated 24% on 2016-01-31
  • PseudonymousUser estimated 40% and said “Each new game AlphaGo plays against itself only adds a marginal benefit at this point. ” on 2016-01-31
  • JoshuaZ said “ChristianKI, how common is resignation in upper pro games? ” on 2016-01-31
  • ChristianKl estimated 40% and said “Update based on https://www.youtube.com/watch?v=NHRHUHW6HQE” on 2016-01-31
  • ChristianKl said “@JoshuaZ I don’t have got data for that. On the other hand if the game would have been played out at full strength that could tell us something about the strength of AlphaGo.” on 2016-01-31
  • realitygrill estimated 10% and said “Upon review of AlphaGo-FanHui games, current AlphaGo win. What matters is the 5 months before the match with Sedol. But I don’t see a clear manner in which AlphaGo can improve upon its minor weaknesses (with my limited understanding of it).” on 2016-02-01
  • realitygrill estimated 10% and said “I meant currently, AlphaGo cannot win against Sedol (unless he blunders hard).” on 2016-02-01
  • PseudonymousUser said “Realitygrill, AlphaGo doesn’t just study existing games, it also plays itself. Each month of study is comparable to 1000 years for a human. So it could get a lot better.” on 2016-02-01
  • PseudonymousUser said “JoshuaZ, resignation is very common.” on 2016-02-01
  • realitygrill said “@Tapetum-Lucidum I took that into account. I think self-play has inherent limitations in the current approach. My estimate would change if AlphaGo regularly played pros, but even then I’m not completely sure.” on 2016-02-01
  • danielfilan estimated 75% on 2016-02-01
  • Waring estimated 60% on 2016-02-01
  • davidmanheim estimated 60% and said “Given the analyses of the matches, and people’s reactions to the relative strength of Fan Hui and Lee Sedol, I was probably initially overconfident.” on 2016-02-02
  • lavalamp estimated 88% and said “Obviously they won’t send October’s version off to the slaughter. How do you think the team convinced some google exec to cough up 1M$ if they’re wrong?” on 2016-02-02
  • WilliamKiely estimated 25% and said “https://www.reddit.com/r/baduk/comments/43xret/how_likely_is_alphago_to_win_50_against_lee_sedol/” on 2016-02-03
  • SatvikBeri estimated 75% on 2016-02-03
  • molochhasme estimated 60% on 2016-02-04
  • hipp0 estimated 75% on 2016-02-04
  • hipp0 estimated 84% and said “DeepMind team said they were quietly confident, and arguably, their opinion is all that matters. I’d say their elo is probably around 100-200 points higher (70% win rate = 0.84 probability over 5 game series).” on 2016-02-04
  • RobbBB estimated 35% on 2016-02-04
  • Houshalter estimated 25% and said “I’m very optimistic about the future, but AlphaGo specifically seems kind of crude and has known limitations/weaknesses that a human player could exploit.” on 2016-02-04
  • EloiseRosen estimated 40% on 2016-02-04
  • aarongertler estimated 65% and said “I feel similarly to Gwern, but in this case I use the same words to mean a higher-than-50% chance. If AlphaGo really had crude, exploitable weaknesses, I think Fan Hui would have taken a game.” on 2016-02-06
  • gwern estimated 40% on 2016-02-07
  • WilliamKiely said “https://www.gjopen.com/questions/133-will-google-s-alphago-beat-world-champion-lee-sedol-in-the-five-game-go-match-planned-for-march-2016” on 2016-02-07
  • Laurent Bossavit estimated 30% on 2016-02-08
  • WilliamKiely estimated 40% and said “I have a $500 bet at even odds that Lee Sedol will win the match. https://www.gjopen.com/comments/comments/126917” on 2016-02-08
  • CarlShulman estimated 54% and said “My best guess model of DeepMind says they would not have announced this game unless they think they are more likely to win than not, and there is a lot of room to improve. Deep Blue lost before winning, Watson didn’t. I have a 1:1 bet on this.” on 2016-02-09
  • WilliamKiely estimated 45% on 2016-02-10
  • Daniel Reeves estimated 60% and said “Bet with Eliezer Yudkowsky that gives me 0 expected utility if this is my probability.” on 2016-02-10
  • znort987 estimated 65% and said “The AlphaGo that has played FanHui is not the AlphaGo that will play Lee Sedol. It has had a full 4 month to play against itself. Also, it is unclear how much more computing power Google will use this time.” on 2016-02-12
  • znort987 said “There’s a money-backed bet on bitbet for this:https://bitbet.us/bet/1249/alphago-will-defeat-lee-sedol-overall-in-march/” on 2016-02-12
  • CarlShulman estimated 62% and said “New statements by DeepMind engineers.” on 2016-02-19
  • Raahul_Kumar estimated 85% and said “AlphaGo distributed is 22-24x more powerful than single core version which beat a 2dan pro.1760 CPUs,240 GPUs.” on 2016-02-19
  • WilliamKiely estimated 65% and said ““Park Chi-Moon, the vice-chairman of Korea’s Go association, injected a note of caution for humanity. “DeepMind said they believe AlphaGo has a 50% chance of winning … we believe they are in fact more confident,” he said.”” on 2016-02-22
  • WilliamKiely said “http://www.theguardian.com/technology/2016/feb/22/google-deepmind-go-alphago” on 2016-02-22
  • nshepperd estimated 75% on 2016-03-05
  • Josh Holland estimated 70% on 2016-03-05
  • Josh Holland estimated 30% and said “I really should read which way round predictions are more carefully >_<” on 2016-03-05
  • Laurent Bossavit estimated 45% and said “Revising based on analysis at https://gogameguru.com/can-alphago-defeat-lee-sedol/” on 2016-03-07
  • themusicgod1 estimated 38% on 2016-03-07
  • CarlShulman estimated 72% on 2016-03-08
  • Temeraire estimated 82% on 2016-03-08
  • leo_gri estimated 63% on 2016-03-08
  • nshepperd estimated 90% on 2016-03-09
  • artir estimated 40% on 2016-03-09
  • artir estimated 67% on 2016-03-09
  • WilliamKiely estimated 95% on 2016-03-09
  • gwern estimated 75% on 2016-03-09
  • EloiseRosen estimated 80% on 2016-03-09
  • realitygrill estimated 65% and said “Having taken game one, it’s undeniable that the rest of the series will be at least competitive. Sedol must be shocked, though, which will probably factor into the subsequent games.” on 2016-03-09
  • CarlShulman estimated 83% on 2016-03-09
  • Temeraire estimated 94% and said “I was pretty confident AlphaGo would win all its games and today’s result is both supportive evidence and required Lee to win 3/4 from now on. Main concern is model error + Carl’s substantially lower estimate.” on 2016-03-10
  • JoshuaZ estimated 52% on 2016-03-10
  • JoshuaZ estimated 72% on 2016-03-10
  • PseudonymousUser estimated 80% on 2016-03-10
  • PseudonymousUser estimated 80% on 2016-03-10
  • PseudonymousUser estimated 80% on 2016-03-10
  • rhaps0dy estimated 75% on 2016-03-10
  • kuudes estimated 85% on 2016-03-10
  • CarlShulman estimated 95% on 2016-03-10
  • CarlShulman estimated 96% on 2016-03-10
  • Laurent Bossavit estimated 80% and said “Time to update, after the first two games went to AlphaGo.” on 2016-03-10
  • realitygrill estimated 98% on 2016-03-10
  • JoshuaZ estimated 92% on 2016-03-10
  • JoshuaZ estimated 92% on 2016-03-10
  • EloiseRosen estimated 93% on 2016-03-10
  • leo_gri estimated 96% on 2016-03-10
  • sdr estimated 95% on 2016-03-11
  • PseudonymousUser estimated 93% on 2016-03-11
  • nshepperd estimated 99% on 2016-03-12
  • Jach   judged this prediction right on 2016-03-12.