when do you think computers 9 dan?

[phillip1882]

so, whats everyone's prediction?

[DrStraw]

Depends which server you are talking about.

[yoyoma]

Depends which server you are talking about.

That makes me think it would be fun to see one of the top bots on Tygem, which has much weaker dan ranks than KGS. If a bot can get 6d on KGS maybe it can already get 9d on Tygem!

[EdLee]

Singularity essay (2001).

[Mike Novack]

First of all, what kind of 9 dan was meant? We don't usually consider ranks above 8 dan amateur. And the current strength of the strongest programs is around 6 dan amateur on rather strong hardware.

Based on how go playing software has progressed, I would say "unpredictable". We have not seen the move from 6 kyu to 6 dan as an incremental process of improvement but jumps because of a conceptual breakthrough. First we had "rule based AI's" which didn't need a lot of processing power playing at about 6 kyu. Then came the MCTS insight, a entirely different approach, jumping to about 1 kyu immediately and only then incrementally improving to about 6 dan by tuning compromises for the algorithm using limited time. This however, very much computationally expensive and we are probably at or close to the point of diminishing returns so greatly increasing crunch power probably not much help.

Another breakthrough? It is too early to tell what a neural net approach or neural net combined with MCTS might yield. The first go playing neural nets are just trained to "predict the move an expert would make" (given games between experts). There may be different/additional training that would be more effective. And there is the possibility of having a neural net (fairly quickly) produce a limited set of candidate moves with the decision between these by MCTS.

The point I am trying to make is that it's simply too early to tell. But getting from 6 dan amateur to 9 dan pro isn't going to be a matter of incremental improvement.

[hyperpape]

I mostly agree with you, that it's probably not a matter of incremental improvement. However, I was struck by the recent win over Cho Chikun at 3 stones--it's the best result so far that I know of.

Based on that, I said 2030, but wouldn't be confident that it won't be either after 2035 with slow progress, or before then with a new breakthrough.

[yoyoma]

I mostly agree with you, that it's probably not a matter of incremental improvement. However, I was struck by the recent win over Cho Chikun at 3 stones--it's the best result so far that I know of.

Based on that, I said 2030, but wouldn't be confident that it won't be either after 2035 with slow progress, or before then with a new breakthrough.

I didn't hear of a bot beating a pro on 3 stones. CrazyStone lost on 3 stones to Cho Chikun on 2015-03-17:
http://www.computer-go.info/h-c/

[hyperpape]

Oy vey. You're right. What's worse, I looked at the sgf and the L19 thread. No idea how I managed to screw that up.

[Pippen]

Are pros still unbeaten with 3HC? I am sure that no pro ever lost a 2HC game to a program. I think when the first program wins over a pro with 1HC then it will be interesting to look into the future. Because with 2 or more HC stones the game is just different...it's like if a chess player would play without two pawns...it's not just the pure quantity of the two pawns, but the whole game becomes different structurely because of that.

As it is now I'd think 2030-2100 is the time frame. 2030 if there's some breakthrough soon, 2100 if there isn't, but I guess in 60/70 years cpu's will be so much more powerful that the sheer quantity of calculations could beat the human down. Imagine a program with 1.000.000.000.000 MC-simulations per second. I think that would help even if CS or Zen don't develop further.

[Kirby]

Computers are already stronger than me, so it doesn't make much difference to me.

[zedmango]

Are pros still unbeaten with 3HC? I am sure that no pro ever lost a 2HC game to a program. I think when the first program wins over a pro with 1HC then it will be interesting to look into the future. Because with 2 or more HC stones the game is just different...it's like if a chess player would play without two pawns...it's not just the pure quantity of the two pawns, but the whole game becomes different structurely because of that.

As it is now I'd think 2030-2100 is the time frame. 2030 if there's some breakthrough soon, 2100 if there isn't, but I guess in 60/70 years cpu's will be so much more powerful that the sheer quantity of calculations could beat the human down. Imagine a program with 1.000.000.000.000 MC-simulations per second. I think that would help even if CS or Zen don't develop further.

Yes, pros are still unbeaten with 3 handicap stones but both CS and Zen have beaten pros with 4.

What's even more interesting to me is when the pros start taking stones. It will be amazing to see games with a computer better than pros even when the pros take a couple stones. I'd love to know how good a program can get - if there could be a program that beats pros even when pros take 9 stones.

[Eizerho]

Even perfect play couldn't beat a top pro giving 9 stones, unless the pro is forced to lose at gunpoint, the pro wants to joke around, it's 1 second blitz or something else ridiculous. Of course, I can't prove any of this, just like I can't prove 1 is the best move, even though I'd bet my life that it is.

I'm guessing that perfect play might be like 3 stones stronger than a top pro, but I'm just another noob.

[Eizerho]

Depends which server you are talking about.

That makes me think it would be fun to see one of the top bots on Tygem, which has much weaker dan ranks than KGS. If a bot can get 6d on KGS maybe it can already get 9d on Tygem!

According to http://www.computer-go.info/h-c/, CrazyStone lost to a amateur who is 7d on tygem. So I seriously doubt it can reach 9d there.

I also don't buy the claim that the strongest bots are 6d, 6d at blitz, maybe. Crazystone achieved 6d KGS with blitz time settings. I don't think bots are nearly as strong in slower time settings. For example, Zen19S is only 3d on KGS.

[Pippen]

I also don't buy the claim that the strongest bots are 6d, 6d at blitz, maybe. Crazystone achieved 6d KGS with blitz time settings. I don't think bots are nearly as strong in slower time settings. For example, Zen19S is only 3d on KGS.

Agreed. Just last year CS lost to FJD, a german 6d, in a serious tournament over 4 games. I think there is no program who could beat a regular 6d in a best of 5 match with serious thinking time, I think Zen or CS could beat 4d's though in such a match-up...no way they are still kyu's or shodan's. But either way: In 2100 processors will be so damn powerful that it won't matter much. At the end brute force will win it for Go programs too, it will be just more indirect then with Chess.