I don't think it's correct to describe both Handol and Alphago's responses to move 78 against Lee Sedol as a "bug". I only have very limited experience with deep learning algorithms, but it seems to me that Handol just made the best move it could find with the information it had. This ended up as a mistake, but calling every AI mistake a bug kinda takes away from the achievement of the other player. It feels more like a blind spot than an actual bug in the code. I've fed some of my games through Leela Zero and even I have seen cases where it just couldn't read out a simple ladder. But anyways, great game and great video!
I think you're correct insomuch as calling it a "bug" is a little misleading. It's more like Lee Sedol "played a specific kind of variation that lead to the AI overlooking something". It's actually fascinating the Lee Sedol has managed to do this twice. I wonder if something about his playing style / tesujis he looks for somehow leads to situations that confuse the AI. Or, it's just a coincidence :) On "'I've fed some of my games through Leela Zero and even I have seen cases where it just couldn't read out a simple ladder" -- that probably has more to do with your computing power. Leela (and most AI) have trouble reading ladders if the computer has insufficient processing power. If you let your computer run variation for long enough, Leela will eventually realize it's a ladder. I think that kind of error is very different from the exploit Lee Sedol found in this game. Handol was probably powered with a super computer. Handol simply couldn't foresee the net, irrespective of how many variations it read out.
Go Pro Yeonwoo, this isn't a net bug, the bot didn't read the ladder for move 78, all ladders work for black after and because of it the bot will bug out from reading anything else. I know this because it is easy to trick LeelaZero or Katago with multiple ladders in one game.
I understand that Handol might claim to be stronger than AlphaGo-Master (Ke Jie), but this game demonstrates that might not be true. It was kind of Lee Sedol to demonstrate this.
Just a mere 10 years ago it would have been considered absolutely ludicrous to try to play against one of the top human players in the world... with the human player getting two handicap stones.
Haha no.. just a quote of what you said. ;) What a great game to end a career! I love how Lee Sedol never shies away from complications, even when facing a superhuman opponent.
you are wrong, the version ke jie played was stronger than the version lee sedol played. this was alpha go masters, which later also played online, and this version was much stronger than the version lee sedol played. you find the elo ratings online. if ke jie has played the version lee sedol played, he would had have good chances. even kie jie said that this version was much much stronger and that he wanted to play against alpha go because, he looked at the game of lee sedol and thought this version of alpha go was beatable for him. and the latest version of alpha go zero, was over 5500
i can't really comprehend what pros thinking but this time, i am surely don't know what Lee Sedol had in mind about being retires, i mean his Go Isn't "dead" yet -is it had something to do with association and such?-
I wonder why AIs are so bad at ladders, maybe because it's a very iterative process and it has to calculate a whole bunch of other meaningless variations at each step of the ladder ?
Lee Sedol: "It's the only move that makes sense for Go pros"
Go Pros: "huh?"
When the robots rise up against us Lee Sedol will lead the resistence.
Lee Sedol = John Connor
I will remember this moment at the battle against korean speaking robots.
He will save us with move 78
@@sethsun true . Well that ai was made by a human too
It’s quite the coincidence that both “divine moves” were the 78th move each time.
*xfiles theme plays*
Awesome commentary! Also didn't know the result so thanks for keeping suspense until the end!!
@gino betti omg I'm so sorry if this happened I will change it now!!
I love ur Pro games analysis. U make them fun and easy to understand.
Amazing! Thanks for making such great videos!
Thank you for your wonderful commentary!
I used it in my kids class, because this game has many ladders and nets, which they can understand.
You are an amazing story teller and even the IA got a personality with you ! i love to watch i learn a lot
I don't think it's correct to describe both Handol and Alphago's responses to move 78 against Lee Sedol as a "bug". I only have very limited experience with deep learning algorithms, but it seems to me that Handol just made the best move it could find with the information it had. This ended up as a mistake, but calling every AI mistake a bug kinda takes away from the achievement of the other player.
It feels more like a blind spot than an actual bug in the code. I've fed some of my games through Leela Zero and even I have seen cases where it just couldn't read out a simple ladder.
But anyways, great game and great video!
I think you're correct insomuch as calling it a "bug" is a little misleading. It's more like Lee Sedol "played a specific kind of variation that lead to the AI overlooking something". It's actually fascinating the Lee Sedol has managed to do this twice. I wonder if something about his playing style / tesujis he looks for somehow leads to situations that confuse the AI. Or, it's just a coincidence :)
On "'I've fed some of my games through Leela Zero and even I have seen cases where it just couldn't read out a simple ladder" -- that probably has more to do with your computing power. Leela (and most AI) have trouble reading ladders if the computer has insufficient processing power. If you let your computer run variation for long enough, Leela will eventually realize it's a ladder. I think that kind of error is very different from the exploit Lee Sedol found in this game. Handol was probably powered with a super computer. Handol simply couldn't foresee the net, irrespective of how many variations it read out.
Finally😆 I couldn't find anywhere to watch live!
BadukTv
Live replays on Fox Go Server, OGS, and KGS
maybt Tygem?
Lee Sabom plays the move all professionals must play... Handol walks into swinging door. Super-excited for next game!
You must be working really hard, for now you need two videos for any event :)
Yes i spent my whole day for making videos!
Go Pro Yeonwoo, this isn't a net bug, the bot didn't read the ladder for move 78, all ladders work for black after and because of it the bot will bug out from reading anything else. I know this because it is easy to trick LeelaZero or Katago with multiple ladders in one game.
For the record, it's easy to trick this DDK with multiple ladders as well. ;-)
This position was checked by Leela zero, fun that it is considering whole bunch of moves beside only this one)
@@petribalanceisnice Without a doubt. I agree.
Thanks!! Best Go commentary!
"Handol thinks very long at first move" yeah because it copies Lee Sedol's brain and it knows what he thinks.
I understand that Handol might claim to be stronger than AlphaGo-Master (Ke Jie), but this game demonstrates that might not be true. It was kind of Lee Sedol to demonstrate this.
Do you have any reccomend baduk app for andriod? thx
Great game! and your comments are so funny! Keep it up!
"My dog is also shocked"
X-D
Very good review , thank you !
Hello, can u share what time the second match will start? Thx
Now
Just a mere 10 years ago it would have been considered absolutely ludicrous to try to play against one of the top human players in the world... with the human player getting two handicap stones.
I was watching this game on TV today and I couldn't believe this happened xD it was so funny
My dog is also shocked. :D
really? haha
Haha no.. just a quote of what you said. ;)
What a great game to end a career! I love how Lee Sedol never shies away from complications, even when facing a superhuman opponent.
Just to show Lee Sedol artistry not just strength. Not the only AI he bugged out LMAO
Developers spending the night trying to make it see the net earlier. :-)
Interestingly, Lee Sedol also bugged out Alphago game 4 with move 78.
Yes exactly!
12:02 Wow I am a noob, but even I can see that Handol is crazy efficient with every move that it plays.Scary
So is Handol available for the public to play ?
Yes on hangamebaduk but not a version against Leesedol
Tomorrow he will play version 3.0.1. :-)
Amazing bug, I'm impressed!
you are wrong, the version ke jie played was stronger than the version lee sedol played. this was alpha go masters, which later also played online, and this version was much stronger than the version lee sedol played. you find the elo ratings online. if ke jie has played the version lee sedol played, he would had have good chances. even kie jie said that this version was much much stronger and that he wanted to play against alpha go because, he looked at the game of lee sedol and thought this version of alpha go was beatable for him. and the latest version of alpha go zero, was over 5500
Lee Sedol "the bot killer" ;)
i can't really comprehend what pros thinking but this time, i am surely don't know what Lee Sedol had in mind about being retires, i mean his Go Isn't "dead" yet
-is it had something to do with association and such?-
Lee Sedol the Great!
Mr Seedol is the best :)
🙏🙏 Thank you
영어로 바둑에 대한 비디들을 만들어서 찐자 감사합니다
I wonder why AIs are so bad at ladders, maybe because it's a very iterative process and it has to calculate a whole bunch of other meaningless variations at each step of the ladder ?
So beautiful! (you)