https://www.youtube.com/watch?v=gQ2qQFwmMEg

Yesterday Korean news channel YTN reported that a Tygem account suspected to be AlphaGo has been gaining attention. There are several reasons that this specific account may be the mysterious robot - it's british, its name("deepmind") is identical to AlphaGo's developers, and moreover its playing style has certain quirks that's usually attributed to machines. And it suddenly became much stronger since last december - right after Google announced the match with Yi Sedol(the account itself was created in 2014). At one point it scored 4 consecutive wins against some of Tygem's strongest players - who may very well be top pros.

Kim Chan Wu 6P, who is also an A.I. programmer, analyzes "deepmind"'s games and concludes it has an edge over human players in its calm judgement ability. For example, at 1:29 of the report, he says that "deepmind"'s sacrifice of the center to gain profit at the bottom shows that it picked the move that has the highest probability of winning the game, which is typical of machines, and Yi Sedol might have a hard time if he gets too worked up about local profits. OTOH, though, "deepmind" seems to struggle against 'unusual' moves, especially in the opening, as in the example at 1:50 shows. In that game, the human player(black) suspected his opponent was a robot and decided to play a very unorthodox opening and won. The report concludes that how well AlphaGo can handle Yi's creative probes may be the determining factor of the match.

(At the moment, "deepmind" is a solid 9d with a 11-9 record on recent matches. His winning rate might be a bit skewed if he kept playing very strong players on purpose, and if that's the case I think his record is very impressive - certainly enough to give Yi a run for his money. It is also very possible "deepmind" is AlphaGo running on weak hardware, so it might appear weaker than it really is.)

Not true, that game they showed was deepmind vs. sky15. Sky15 is a Tygem 9d player who frequently plays this opening (as both Black AND White), and did not prepare that opening against deepmind specifically because he or she thought deepmind was AI. Here is the game record in case anyone's interested:

Sky15 is particularly strong at playing Ko.

Aja Huang said it was he playing (he's 5D EGF), it might be Fan Hui too since he had join Deepmind team since his lost. It might be AlphaGo's calibration process, or a trap lay by deepmind's team (they're aware that everyone is watching out for any robot account in Tygem), I don't know.
Or maybe this, deepmind has more than one Go program, in the latest lecture Demis Hassabis said he has all kind of experiment of so many method.

Thanks, good find!

I also found the first game shown in the report; it's zyoda vs deepmind, 2016-02-21.

I don't know how to upload game records here, so sorry about that

[Please move this topic to Computer Go section]

_________________
Tsumego/Tesuji apps for iPad, iPhone & Android devices:http://www.lifein19x19.com/forum/viewtopic.php?f=18&t=7511

Keeping in mind all the caveats about "deepmind", Haylee has a much better record as a 9d if I remember well. So there might still be hope for Lee Sedol

AlphaGo has multiple strengths, depending on how many CPUs/GPUs it's using, as you can see from the Nature article.

So if this "deepmind" account is actually AlphaGo we still don't know how strong the AlphaGo playing Lee Sedol will be, as I suspect they can't afford(?) to keep 1202 CPUs etc used in the Fan Hui match dedicated to this account. It's 'possible' they're only using 1 CPU, but the decrease in strength implied by doing this would be very noticeable.

Anyway, if it is actually AlphaGo, and I strongly doubt they would set up an account called "deepmind" at this stage, then I think it's inevitable that it will be stronger than this account during the Lee Sedol match.

On a subsidiary point, if Haylee is 9d on this server, then I'd expect a full AlphaGo to be stronger than that.

Oh yes they could. Look, that sort of thing depends on the budget assigned to the project. It is also a reason why we don't really have a fair idea of what the stronger MCTS programs could do were they not the projects of individuals and small academic groups which can't afford to dedicate that much hardware << yes not linear, diminishing returns severe, but we're looking at an order of magnitude or more crunch power >>

Look, high as the cost of using that hardware might seem to you, what do you imagine the cost of the development team? How many systems analysts and programmers for how many hours? And if you think the cost of some (comparatively) few hours running a bot is high, how about the 7x24 running for "training"?

It's also worth mentioning that the win loss percentage reveals very little about the strength of the player at the 9d level. Strong pros who play on Tygem will only play against other 9d players who they feel are near their level, and constantly refuse challenges from "weak" 9d players if you ever watch one of them stream. As a result, some of the best 9d players on Tygem have lacking records. 다시(P) / Qichu(P) for instance, is a top Tygem 9d player who has a record of 406-391, and used to have a below 50% winrate, but tons of people will watch his games because he only plays against the best players.

Mogo used to use 3200 cores supercomputer
And about Zen http://computer-go.org/pipermail/comput ... 04894.html

Where could I see those live streams? Is it on twitch or some asian streaming site? Also I don't suppose they speak in English, do they?

I see your point, but since this topic is closely related to the many Yi vs AlphaGo threads here (of which the number will probably explode next week - after all, this match is the biggest event in professional go right now - one can argue it's also a big event in A.I. research, but the weight it has in the Go community vs A.I. community is incomparable IMO; to the former, it's an event that will mark the greatest turning point in the 2500+ year history of the game while to the latter it's just a flashy advertisement of one of the more important breakthroughs in A.I. research over the years.), I think it will be a lot more practical and convenient to the users to keep related threads in one place. After all, isn't that the whole point of categorizing threads? OC, that's just my opinion so I'm fine with moving the thread if others think that's better

I seriously doubt Aja Huang was serious when he said all the games of "deepmind" were played by himself. AFAIK he's a 6d(european?) which I don't think is strong enough to pull out deepmind's exploits over the last few months which includes a few dozen wins against confirmed pros and winning streaks over Tygem's best players.

If one understands what a hassle it is to create a 9d account there (you have to fight your way up all the way from 3d), there's a good chance this account is a relatively rare and important asset to the team, especially if you consider their limited time to prepare the match(unless OC AlphaGo is so strong that it can create 9d accounts for breakfast, but even then doing that would be quite undesirable since they'll risk dragging huge attention. ).

Thus I don't think it's very reasonable to completely lay aside such an asset just for the sake of secrecy, when there's still tons of ways to bluff your strength while using it as a test bed; like constantly making certain types of mistakes or several human players sharing the account with AlphaGo.

Also, the "why would deepmind create an account as obvious as 'deepmind'?" argument isn't very convincing because this account was created long before the team became famous. It's probably not the team's only test bed account, but IMO there's enough indicators pointing out that this account is at least occasionally used as a test bed as well. Especially when the account is 'on a good day'.

Aja is 5d EGF and not a lot stronger than me: we actually played shortly after the Fan Hui match (but he kept schtum!) and both had chances to win. So I don't think he would be 9d Tygem, certainly not beating pros.

Demis Hassabis confirmed that userid deepmind is id used by AlphaGo at tygem to practice.
172 games played by AlphaGo but there are lot more games in TYGEM played by deepmind. Early games maybe played by Aja Hwang.

http://www.tygem.com/news/news/viewpage ... &findword=