AlphaGo vs. AlphaGo with Michael Redmond 9p: Game 10, Part 2

[@RoryMitchell00]

Wow. That was amazing. Well worth the time spent, though I think I need to sleep for a few days just to recover from all the effort needed to follow along. White appeared to be in a winning position, threw the bottom left side away to gain some opportunity in the middle. After that, things were balanced on a knife's edge with a lot of complexity and possible variations, until black slowly turned things around, and appeared to have a winning move (as Michael found). However, Black chose a different move (I would love to see what the policy/value networks thought of Michael's move) and White then went on to win by half a point. This is strategy and reading at an incredible level. Thanks again Michael for sharing your wonderful insights with everyone!

[@Rubrickety]

"It's not true when I say it's not going to die... 'cause it is going to die." Definitely one of the best Redmond quotes ever.

[@srstaley1018]

Michael's commentary is exceptional. I just learned even how to play go after watching AlphaGo, but I can listen to him all day explain the moves.

[@cellardoor70]

WOW I'm speechless. What a spectacular saga of a game. Not to mention Michael's outstanding analysis.

[@leeresglas5115]

I love this! It's so entertaining to watch Michael and Chris analyse really complex games while learning a lot about the game. Thank you for doing the series and keep up the amazing work!

[@creatspirit]

Thank you Chris Garlock and Michael Redmond. It was a delight as allways to watch and listen to your analysis of the Alphago selfgames. Keep it up!

[@MichaelFJ1969]

Thank you so much! I really appreciate watching Michael's commentary, and I'm grateful for (and impressed by) the amount of work Michael has put into this commentary.

[@MarillionBD]

Thanks for this commented games. Every time I hear a Michael Redmond commentary, I feel I'm learning.

[@trucid2]

This is what 21st century go looks like.

[@unfixablegop]

Yeah. Unplayable. Move over, humans.

[@ConsciousBreaks]

This is what 31st century Go looks like.

[@PocketCalculatorNL]

Exciting, detailed and rich commentary! Thank you and I look forward to next week

[@outcast12345]

What a crazy and beautiful game. Im so amazed by michaels commentary and Alphago´s play at the same time!

[@KilgoreTroutAsf]

Thanks again for this wonderful series of reviews.

[@chardonnay5767]

That center fight was something else. What a great game, thanks for doing the review

[@furr_]

You guys and these games make me so happy.

[@bernardfinucane2061]

A human player simplifies the game by splitting the board into local areas and remembering the status of each area. That way they can focus on the local position. Strong players can shock weaker ones by re-imagining where the boundaries go, or predicting where they will go, or with tenuki. "Ear reddening" style moves transcend this completely by addressing the entire board. Alphago has no concept of local positions and doesn't remember between moves which groups are dead and which are alive. It simply re-evaluates the entire board each move, as far as I know. So this constant shifting of individual groups from dead to alive to seki to ko and back again aren't any additional problem for it. For the same reason, it's also very good at tenuki and reddening ears.

[@mitchumsport]

wow! really appreciate michael lending his intellect to this project and the rest of the team for putting it together. I may have to watch it again to get more out of it, but this was amazing!

[@Keldor314]

I can only conclude that white decided it was going to loose, and so threw the corner away in order to make a ridiculously complex game where victory could go either way.

[@unfixablegop]

Complex and simple mean nothing to AlphaGo. All it sees is the winrate of it's value-net assisted playouts. The winrate of the chaos move was higher than that the winrate of the "natural" move. This doesn't really translate to anything meaningful in human terms, because it doesn't mean anything to Alphago either. Michael Redmond is painstakingly untangling AlphaGos plans. He is performing a great service for us, but the irony is, that AlphaGo has no plans.

[@Silly.Old.Sisyphus]

I can only conclude that white decided it was not going to lose, and so threw the corner away in order to make a win safer.

[@Danielmoen88]

I get what you're saying but I think you're missing a point. In a simple position an AI would be able to read more accurately and a larger number of the variations should point to a win for the player being ahead, while a complex position would bring more uncertainty and perhaps even the win percentage since a larger number of variations it reads out doesn't spell clear (it being 0,5 points or 30 points) victory for that same player. So playing a good! move that greatly complicates the position for the player behind might bring up the win percentage with that move, since it will open up more possible winning variations compared to the simple move with more "certain" losing variations. Complex and simple mean nothing to Alpha Go, yes, but it doesn't mean it won't simplify or make the game more complex as long as it increases it's win percentage.

[@Keldor314]

Might be good to look at it like this: "If I play the natural game, 60-40 chance I loose. But if I go into this horribly complex and unreadable set of variations, all bets are off and I don't know what will happen, so 50-50!" So if you know you're in a bad position, throw the board into an unreadable mess and hope for the best. Variations leading into the great unknown are better than variations leading to a known loss, though surely not as good as variations leading to a known win.

[@trucid2]

AG's motto in this game is "Cut all over and let God sort 'em out."

[@ANSIcode]

Reminds me of the folklore advice on how to get to 1d on KGS: "Cut everything, then read" :D

[@roylangston4305]

Thank you so much, Michael and Chris. This game really gives a clear impression of Alphago's superhuman powers of tactical analysis. The questions on my mind are, can Alphago give handicaps, and when will we be able to buy it? IMO Deepmind is sitting on a potential billion-dollar product.

[@grolich]

It was a beautiful game and analysis to watch two years ago. A couple of years later, I decided I needed to test that, so the interesting thing is that putting the black attachment in the bottom left (where all hell broke loose and the huge exchange happened) - into several versions of LeelaZ, (and also KataGo), there seems to be a consensus that white is a bit better after any of the moves c5, d5, or c7... With the only disagreement being about how much are white's winning chances. All gave him above 50%... But left a very close game. These AIs don't treat that exchange as complicated apparently... Just as another way to play. Also interesting is that KataGo - which shows predicted game score results - seems to think it's a half point game indeed :)

[@seventus]

Thanks, this is gold.

[@andremouss2536]

7:30 the only following move that could cause a real problem later for white if he retreats at C5 is black E5. Usually Alphago isn't very fond of building so straightforwardly the center, he much prefer to do this in combination with an attack or a threat, but that may be the exception.

[@infinitysalinity7981]

We need to crowd-fund Google to lock Michael up for a whole day in a small room with a Go board, stones and Alphago.

[@TomValedro]

This made perfect sense to me. I play exactly like white does in most of my games.

[@seishinkan5838]

welp my brain just exploded midway into this game.

[@andrewgr144]

This was very entertaining and I enjoyed watching it. Without the commentary, it would have been little more than an almost random series of moves to me. I don't think I'm capable of actually learning anything from this game, unlike some of the more "human" seeming games they've covered; but I'm perfectly happy just being amused for a couple of hours.

[@Silly.Old.Sisyphus]

i'm exhausted just watching!

[@unfixablegop]

Do it like me. Switch off your brain before it overloads.

[@grolich]

about Michael's claim as to the black win he thought he spotted - Apparently, it is very difficult for the AIs to figure out what's happening there and they change their minds slightly all the time... it appears it wants to connect at C7 to force black to connect on the first line, after which it tries many things - pushing at F9 and claiming the later atari black gives (in sente...in the video) is not sente :P offering moves like G12 (the almost cut... just to get that extra center move) or O18... Sometimes it switches briefly to top right mid variation without the push and cut... But it keeps changing its mind for a bit, suddenly agreeing with Michael a bit more, then not wanting to push and disagreeing again... even after given a whole lot of time it appears It is not sure anymore about the result. So I'm afraid without a more powerful analysis tool/computer, and having just a lowly amateur dan endgame capabilities, I am unable to resolve that question even with these powerful tools :P Even with a bit of tinkering with different continuations and using the AIs to help guide me. I am always impressed by the depth of analysis pros can perform on their own :) Just in case people are interested :) That was still a wonderful analysis, and I enjoyed it a lot :)

[@whendrik]

As a mere human, it is difficult to imagine how a more intelligent species would think, like super intelligent aliens. If they exist, they would play Go like this.

[@unfixablegop]

The levels at which go can be played are almost infinite. Maybe AlphaGo has already reached the level of winning with black, no komi, against perfect play. But the last points against perfect play are like trying to scale a vertical cliff without fingerholds. Aliens with even just a few thousand years of a head start on us would look at AlphaGo like we look at a dog fetching a stick.

[@aresistar8285]

ty

[@alekerickson4301]

Fan hui still just has alpha go on his laptop right?

[@TysonJensen]

Alek Erickson Its a laptop that can connect back to Deepmind. They don't let copies of critical software exist on laptops unless there is no other choice.

[@billyswong]

AlphaGo is not efficient enough to live on a laptop yet; or one may say a laptop is not powerful enough to host AlphaGo yet.

[@benjaminschooley3108]

This is the thing with computer players, they have no problem whatsoever walking out onto the tightrope.

[@PiratDunkelbart]

Well, WHO would have thought, that Go is all about reading at the very top. ^^ No wonder why Im still stuck at low dan with soso reading skills trying to make it up with strategy and direction of play. ^^

[@derekstanyer]

I would more say that at the top reading is just AS important as direction of play and strategy. Since reading can enhance your strategy and your direction of play. Reading isn't so good if you aren't doing the right things with it. But yes.... I need to improve my reading too :(

[@kfm1242]

Sickest stuff I have ever saw

[@ArmouredLord]

Just feedback, I love the amount of effort put into this by Michael, but I feel that Chris sometimes sidetracked the commentary by leading to a repeat of what was already said, or was simply slightly awkward by smiling / looking confused at points in the video throughout. I don't wish to be anymore specific, but it made me a little uncomfortable and for me detracted from what is a very brilliant analysis by Redmond. I guess personally for me it's a bit of the body language / facial expressions, if we have just redmond on cam, and Chris on voice, it just feels a lot better!