AlphaGo Zero vs. Master with Michael Redmond 9p: Game 5

[@RoryMitchell00]

I have a question I'd like to ask, hopefully for a future Q&A video: of the three major eras/versions of AlphaGo, which one does Michael think has had the most impact on the Go world? Would it be AlphaGo Lee, since it came first and was so unexpectedly strong, had such intense and global media coverage, inspired so many new fans and players, and played new moves like move 37 in game 2? Or would it be AlphaGo Master (including the improved version that beat Ke Jie), since it played a lot of new openings and tried new things in the series of 60 online matches, showed fascinating and unfamiliar styles of play in its series of self-matches, and proved its incredible strength by never losing a single match to a human? Or lastly, now that Michael has had a chance to become familiar with it, would it be AlphaGo Zero, since it plays so much more like a human, has confirmed so much of our knowledge and understanding by rediscovering joseki and popular openings from scratch, and shown such an awe-inspiring strength and demonstrated how much room there still is to improve in Go? Thanks for taking the time to consider this! I'm also hoping this is a subject that Michael will discuss in the upcoming book!

[@hippophile]

Really exceptionally good commentary on this game. Great job guys! So much to learn from it I shall probably watch this two or three times.

[@KilgoreTroutAsf]

nice! 5th game finally out! thank you so much for the review!

[@ignotasanimum]

I love you guys, was missing this soooo much

[@Nox2100]

来たああああ ! After all this time :D *grab popcorn* Thanks for the video.

[@unfixablegop]

Michael Redmond has such a tough job. The fact is that it's not really humanly possible to criticize AGZero. Otoh, you don't just want to be the bot's press secretary.

[@Maharani1991]

It seemed humanly possible to criticize N14, for instance :)

[@decidrophob]

In fact, post-game analyses are essentially different from the blitz play, even at the level of AGZero. MCTS possibly misses vital sequences, not easily approximated through policy or value networks, like Sedol's move 78, even today. Therefore, GIVEN the severe punishment by the opponent AGZero (or Master winning at around 10%), humans may possibly find the risky move causing the significant loss. I do not say that Michael can create those variations on his own, but he may possibly create them very well with the help of AG's game record, possibly better than the programs themselves prior to playing. You can put higher weight around the actually chosen sophisticated plays than what the PN/VN suggest initially in MCTS so to speak. Currently, at least, even AlphaGo Teach or FineArt explanation seem to depend simply on PN/VN, hence unless the criticism algorithm in the above sense is presented, I feel there still is some room for human criticisms. In the very near future, we may expect more sophisticated analyses programs than the current ones, and that is when, using your term, human pros become press secretary, I think. I nevertheless agree with you depending on how you use the term "humanly". As I counter-criticized Michael's criticisms regarding the center N14 cut and the bottom N4 hane, I think it is not easy to "state in an intuitive simple go language" why those moves may be good or bad. The goodness or badness critically depends on the hidden variations only possible by AlphaGo's high skill. In that sense, Hassabis is too correct in saying that go requires tacit intuition, which may be inhuman in a way.

[@MelindaGreen]

Great to see more AG self-play analysis! Could we maybe see some AG0 vs. AG0? I would also love to see more of the Master vs. Master games since it has such a fun and wild style, but I'm just happy to see any AG reviews.

[@decidrophob]

I completely agree. I made a lot of comments in this video. I am pretty sure that Master vs. Master game records are probably stronger than this experiment series since they were played at 2 min / move setting, which is extremely long.

[@mushrooms2261]

I have a dumb theory based purely on wild speculation! Maybe AlphaGo plays bad moves when it thinks it's ahead, but also knows that its *opponent* will do the same, so it tries to play moves that look bad but secretly have some value in order to trick the opponent into overconfidence??? That would explain both the first strange move, and the strange response to it!

[@gatsud4685]

I am sorry but it is not the case at all... When alphago loses he plays basicaly forcing moves wich are bad exhanges because it is the only chance for him to win the game (i.e maybe the oponant won't see the atari) Because when nothing works these moves are the moves with the most wining chances, but of course it is really easy to conter such moves. On the other side when alphago wins all the moves are pretty much working, they are all legit with the same wining chances (close to 100%) . So one of them is chosen almost randomly wich makes the game really fuzzy to look. Even if this move is bad and make him loses some points, in the end the result is the same : he is still wining.

[@paulbloemen7256]

KZbin just presented this video to me, great game! As for the comments, two years after the video was made, I wonder what Leela Zero or KataGo would have to say about it, whether the “crazy” moves like the cut at N14 really are that bad: when replaying the current Leela Zero games, I see these kind of moves quite often, for better or worse.

[@MIKEEMAXER]

" Master likes to play high " .... so this was its secret the whole time ....

[@alekerickson4301]

Can't wait to watch this!

[@robharwood3538]

For anyone interested in having something like AlphaGo Zero available for their own learning and exploration, there's an open source project called Leela Zero which is designed and trained based on AlphaGo Zero's architecture and training protocol (as described in the AGZ paper by Deep Mind). Use it with any Go GUI which supports GTP (e.g. the open source Sabaki project). You can even help train it on your own computer and watch it climb the ranks at OGS and CGS! Cool, eh?

[@robharwood3538]

Leela Zero progress: zero.sjeng.org/ Leela Zero "Read Me" page: github.com/gcp/leela-zero/blob/master/README.md

[@robharwood3538]

Sabaki: sabaki.yichuanshen.de/ Leela Zero on OGS (Royal Zero): online-go.com/player/484523/RoyalZero CGS Rankings (look for names starting with 'LZ'): www.yss-aya.com/cgos/19x19/standings.html

[@omfgacceptmyname]

at 11:00, why does black not push at g17 before f14?

[@robharwood3538]

Hooraaaaaayyyy!!!!! Glad to see this series back! \(^_^)/

[@zielad362]

Thank you very much!

[@decidrophob]

43:30 I feel white's utilizing the aji around the center, probably threatening to save the five stones by capturing two black stones in the center, will either result in a ladder breaker or severe damage on the center for black. Will probably start at L10 attachment. The bad aji Michael mentioned a minute ago probably would have the worst effect if black played aggressive around here.

[@benjaminschooley3108]

It would be interesting if Kasparovs theory that human plus computer would beat the computer alone would hold up with these alphago programs. I think Michael would be able to help zero or master quite a bit in the opening especially, prevent these 3-3 invasions and so on.

[@williamtomlinson85]

I would like to see a game like this. Some of these hard to explain moves look like mistakes, but maybe something deeper is going on here that we just don't see. I get the feeling that the human "correcting" these types of moves would hamper AlphaGo, not make it stronger.

[@HearterSG]

Redmond-sensei!!

[@alekerickson4301]

ok i watched it and it is amazing!

[@decidrophob]

22:25 I personally think that if white double haned at N3 M2 by sente on the right bottom punishing black not playing N4 M5 double hane as he did in the game, and surround the right side moyo utilizing the center aji, then enlarging the left side black moyo does not seem enough. Hence, I feel black was correct in assessing that the left side moyo is not large enough if he foresaw the center above sequence in which white reduced the left top area; which probably motivated the sente N4 M5 double hane sequence on the bottom.

[@decidrophob]

35:30 Probably, white wanted to take sente around here and invade the left side. White is giving pressure to black to continue playing around here, otherwise the peeping stone will be captured big. On the other hand, if white played A, then black would be satisfied to treat the peep as kikashi and defend the left side first, hoping for the future effective yose on the right side. This resembles the center reverse yose that appeared in Match 2 of the Sedol Match, for which Michael and other commentators criticized AlphaGo for not defending the right top.

[@KeithWagner8]

Can we get someone other than Chris Garlock in these videos?

[@luftlayersoroma5105]

17:40 maybe the reason why is the bottom so important is the hane at B3. Than is Wb2 than G2 the followup, and B can get lot of terry without aji

[@decidrophob]

kzbin.info/www/bejne/iandYoGkgriWh7Mm41s Considering the left bottom joseki sequence, the bottom side, at least alone, is quite neatly enclosed. F2 alone therefore does not significantly reduce the value of n4 hane when it is played in sente as it was in the game. The problem with the hane shape, rather, seems to be related to the sharp move J5 and the related aji on created by the sacrificed five stones on the top center. I think Zero probably utilizes tesuji and sees aji deeper than Master. Thus the refutation of M5 looks to me much more subtle to depend essentially on Zero's fighting skill than what Michael explains in the video.

[@Maharani1991]

Another great comment, thank you :)

[@hippophile]

Neatly enclosed, but only with another move added, and of course giving up some points and adding defects on the left side. I believe Michael is right that the bottom area is therefore relatively small and N4 is questionable, using human style analysis (which is not available to the AI).

[@decidrophob]

This is again a quite delicate issue. Master, though probably not as extreme as Zero, is also excellent in demolishing the opponent moyo, as we have seen in the 60 game series. In the old Japanese proverb days, we were taught to "cover from the wide side (for potential huge moyo)". But today pros and Master quite often choose to cover 33 from the narrow side considering this particular joseki in mind. The wall created by covering 33 as we all empirically know today is not necessarily thick and often the moyo is too wide. I clarified this issue in my another comment regarding this video. Simply put, I personally feel, probably as well as Master, that left side moyo is neither huge nor thick enough, particularly considering the aji around the center. Only when it is verified that left side moyo can turn into territory without too much damage elsewhere, and when it is enough, can you say that bottom enclosure in the game is not worth.

[@decidrophob]

Bensondarr and FineArt also look as strong as AlphaGo now, particularly spectacular in creating aji like the center cut at N14 by Zero in this game. kzbin.info/www/bejne/iandYoGkgriWh7Mm27s Michael does not like the cut, but ironically thanks to his detailed analyses, I love the cut very much. It is not explained in the video, but I think white wanted to play F17 atari to defend the left top most effectively. If black responded simply by connecting at g16, then white probably will cut the N line black stones on the right top off via L16 and squeeze by semeai at the very least, which will completely secure the right top. That is why black needed to capture the center stones small instead of connecting at g16. The result looks already very good to me in favor for white at the time of g14 keima and g16 ponnuki, reducing the left side black moyo great deal, instead of little bit commented by Michael. Also, in addition to the squeeze effect on the right top explained in the video, N18 hane-tsugi end game yose became much more huge, half sente considering the severity of M17 cut now. Thus the potential white territory size on the right top has increased significantly. In total, white only seems to have gained a lot to me. It is really extremely amazing that Zero or other AIs can "feel" all these ajis in advance even in blitz plays.

[@Maharani1991]

Thanks for the great comment! :) "Further reading" :)

[@hippophile]

Interesting viewpoint. It is impossible for me to judge, but my feeling is that cut is a bit extreme; it is on the edge of plausible; AGZ is so good with aji but early in the game I have less confidence in AGZ judgement. Black's heterodox move at N4 means we probably don't see the best atempt to refute by black.

[@decidrophob]

I am not entirely agreeing with AGZ either, as I hope I made clear in my other comments. I stress however that today's top AI probably does not make a "simple" blunder type of mistake. The judgment seems only possible via weighing various subtle factors. I really appreciate Michael therefore in his showing concrete ajis. I think we may better stop there and should generally hesitate making the OBJECTIVE nuance judgments aggregating those factors. Thus, I intentionally used the subjective expression such as "I love" or "seems to me". I agree with you regarding the potential severest counter play by black. The main nuance of my comment is that GIVEN the result of the capture in the game, I would be very happy as white.

[@robharwood3538]

Hey AGA admin(s)! (Is it you, Chris? :-) ) I just sent a direct message to this channel about a couple of comments I made that probably got accidentally sent to 'spam' (for links). I know KZbin is notorious for not giving notifications about messages/comments, so I'm hoping sending this comment will increase the chance you notice. Could you 'unspam' the original comment(s)? I think the links will be useful/relevant to you, Michael Redmond, and the rest of the viewers. Thanks! (I'll delete this comment after, thx!)

[@decidrophob]

49:40 I feel that the center smells too fishy for black around here. Michael's ko threat is not huge enough. It neither threatens white's right side group enough nor connects to the top black group at this timing. Letting the right top live alone is not enough. White would ignore, escape the two stones H13 J13 instead of finishing the ko by ponnuki and try to kill the lower left side black or give that level of damage to black around the center top. At the very least, white can finish playing around there in sente and come back to the right top. It was actually Michael himself that explained how important capturing the two white stones in the center is in relation to the life of the left side black group. Black's ko threat R15 in the actual game was, I think, a quite delicate ko threat. First, white cannot ignore because Q15 cut at this timing is too severe. So white has to respond, but she cannot respond by connecting at Q15 because then the half life of black stones is too huge. In the game, therefore, white responded by the brilliant R18 sagari, which inevitably increased the number of black's ko threats a great deal, continuing to threat the cut at Q15. Overall, black's starting the ko itself, instead of letting the left side live with B6 capture by gote, was such a brave/risky sente yose type of move. My understanding is that black's right top sacrifice was a part of black's plan when he started this tenka ko. As the result of the ko fight in the game, black securely captured the two stones, eliminated the ajis during the ko fights, and used the sente to capture the right side three stones of white. I feel black's ko paid off overall.

[@sharpspear8755]

It's a real amazing game

[@Silly.Old.Sisyphus]

i agree with Chris, shaking our heads in amazement at Michael's incredible depth of vision and in admiration of A0's dexterity. King Master is dead, long live Queen Alfie0! Did you know (that in Sumerian calligraphy) "Alpha" means "ox"?! - the first God man invented was the female crescent moon with the horns of an ox, who they believed held the power of childbirth because the length of the lunar cycle is the same as the length of the female menstrual cycle. kzbin.info/www/bejne/a53WgoinbqqdgqM

[@bernardfinucane2061]

Yay! But I have to say the action here is sometimes hard to make any sense of.

[@decidrophob]

kzbin.info/www/bejne/iandYoGkgriWh7Mm41s J2 looks a better threat than K2. White can threaten the bottom black group via J1 later so that m3 escape works. No ko is required. Kosumi is better in creating the living shape on the center in the actual game.

[@Maharani1991]

+

[@goinrio]

30:24 wow

[@tracykeithwall]

Intro music works but no voices the rest of the video.

[@benjaminschooley3108]

First!

[@ignotasanimum]

Benjamin Schooley bitch!

[@benjaminschooley3108]

Alexandre Herkenhoff Gama better luck next time 😜

[@ignotasanimum]

😂😂😂

[@ignotasanimum]

Benjamin Schooley thanks! Good game!