An interesting thing to observe in the long run would be if humans playing against AI are able to improve their moves and beat the AI.
Google’s AlphaGo AI defeats team of five leading Go players
51–57 of 57 posts
Re: Google’s AlphaGo AI defeats team of five leading Go players
#52Earlier quoted context omitted.
Lee Sedol was near the top (#7). He wasn't the very top, that was Ke Jie. And now Ke Jie has been beaten 2/3, maybe 3/3 times. Sedol was beaten soundly -- 4/5 games. So this is an improvement, but not an exponential one. Go is definitely conquered like chess was 1-2 decades ago. https://www.goratings.org/en/
According to some information that the AlphaGo team released earlier in the week, the current version of AlphaGo can give a three-stone handicap to the version that played Lee Sedol: http://www.usgo.org/news/2017/05/new-version-of-alphago-self... . I don't know if that qualifies as "exponential", but that's a massive difference in skill at the game. If it's accurate at all, I think there's basically zero chance that…
And the more interesting note - it gained a three stone handicap of skill with 1 year of practice.
If they were to continue at it should we expect a 3 stone improvement in skill every year? Maybe they made some major leaps this time and future changes aren't as large. So a 3 stone improvement every two years.
It doesn't take long until the Go being played is fundamentally a different game than what we play. Even if it's on the same board with same rules.
What does it look like when Go still isn't solved, but AI is 20 stones better than the best humans? At that point we can no longer even tangibly measure its brilliance. Only ask it to compare its brilliance to other brilliant AIs. What will that even look like through out eyes? What will like look like when an AI is 20 stones better than us at war games?
I assume next step for AlphaGo is a real time strategy game like Starcraft, or some other game with incomplete information. 3-5 years from now when mastered it'll be hard to persuade people not to use it in the real world.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#53Have they done a human and AlphaGo vs AlphaGo yet? As I understand it centaur teams of human and computer in chess are superior to computers or humans playing alone. It would be interesting to see if this holds with go as well.
Where did you find this information for chess? It sounds unintuitive that a superior player would benefit from a weak player's help.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#54This is going to make all the Go players out of job pretty soon. :)
Cars have been faster than people for ages, but marathons and sprints are still popular sports.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#55Have they done a human and AlphaGo vs AlphaGo yet? As I understand it centaur teams of human and computer in chess are superior to computers or humans playing alone. It would be interesting to see if this holds with go as well.
Where did you find this information for chess? It sounds unintuitive that a superior player would benefit from a weak player's help.
I can't give you the source I got it from, as it is something I have read multiple times over the years from different sources. A quick search turns up a lot of articles on the subject, with the claim that centaur teams are better.[0]
I cannot follow you on the unintuitive part. If you have a chess program playing against itself, as long as it isn't capable of playing a perfect game every time, the program being supervised by a human should better, as the human will have a non-zero chance of spotting a mistake (or a better move) with enough plays. With low of the risk of mistakes from the weaker human, as she can confer with the judgement of the program. If the risk of human error is lower than the chance of human insight, the centaur team should have the edge on the lone computer.
After checking up on the subject again, it seems like the effect is no longer strong enough to topple the best chess computers though.[1]
[0]https://www.bloomreach.com/en/resources/blogs/2014/12/centau...; http://www.huffingtonpost.com/mike-cassidy/centaur-chess-sho...; http://www.glassbeans.com/blog/why-centaurs-will-dominate-th...
[1]https://chess.stackexchange.com/questions/15772/how-do-the-b...
Re: Google’s AlphaGo AI defeats team of five leading Go players
#56Earlier quoted context omitted.
Where did you find this information for chess? It sounds unintuitive that a superior player would benefit from a weak player's help.
I can't give you the source I got it from, as it is something I have read multiple times over the years from different sources. A quick search turns up a lot of articles on the subject, with the claim that centaur teams are better.[0] I cannot follow you on the unintuitive part. If you have a chess program playing against itself, as long as it isn't capable of playing a perfect game every time, the program being supe…
I find it unintuitive when the strength difference between the two players is too high, and I think matter-of-fact statements like in the stackoverflow answer ("Of course, a human assisting a computer player will be stronger than the same computer player alone") need to be questioned.
Imagine a beginner receiving suggested moves from a grandmaster. I can't see how this beginner could be stronger than the grandmaster alone, and they would almost certainly be weaker if they override just one of the grandmaster's suggestions.
From what I read, chess AIs are this far above humans.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#57Have they done a human and AlphaGo vs AlphaGo yet? As I understand it centaur teams of human and computer in chess are superior to computers or humans playing alone. It would be interesting to see if this holds with go as well.
Where did you find this information for chess? It sounds unintuitive that a superior player would benefit from a weak player's help.
There are centaur championships in chess, for example ICCF. Mainly, people try do strategy and choose directions (which engines are not that good at) , while computer does the calculations.