Is AlphaGo already four stones stronger?

[wineandgolover]

I'm told that this article from a Chinese newspaper claims that the current AlphaGo can give version 18 (the one that played Lee Sedol) four stones. And it's still getting stronger.

Holy crap.

I'd appreciate it if any Mandarin-fluent folks want to read it and confirm or tell me I've been misinformed.

[Majordomo]

From what I recall reading on reddit self-play (aka AlphaGo playing itself) can give exaggerated results in terms of improvement - but even so this sounds massive so if this is true how large is the gap between your average pro and the top pros for context? From what I gather there is very little in the way of handicap games between pros so is there even any data on this or just educated guesssing?

[macelee]

Your understanding of the text is correct. However as suggested this sort of comparison is rather meaningless. TWe can not draw any conclusion from this (assuming it is a fact), such as this new version can give top pro players 4 stones.

[Krama]

I hate deepmind. They don't say anything while they know the hype is great. It's like a scummy marketing campaign.

Would it be so hard for Demis or someone from the team to simply write one message on twitter or something simply explaining if they are still working on it and how much it has improved.

It would take him 2 minutes and it would confirm/denounce all the rumors on the interwebz.

[yoyoma]

Four stone improvement again?! Wow that is incredible! Sure self-play can be problematic but I think it's safe to say that it's quite a bit stronger now than before.

Krama, I share your frustration about the delays in getting good information. But Aja Huang will be giving a talk at a conference next Wednesday -- The 9th International Conference on Computers and Games (CG2016). And the organizers said they will be putting a video of his talk online so hopefully we will get some more information then.

http://computer-go.org/pipermail/comput ... 09432.html

In my talk I will focus on the source of AlphaGo's power: policy and value networks. I will also mention the status and conclusion of our investigation on AlphaGo's problem in the 4th game against Lee Sedol.

[wineandgolover]

Your understanding of the text is correct. However as suggested this sort of comparison is rather meaningless. TWe can not draw any conclusion from this (assuming it is a fact), such as this new version can give top pro players 4 stones.

In past presentations DeepMind did provide a disclaimer about drawing conclusions from computer v computer play, so I'm sure they'd agree. So, maybe we can't draw a firm conclusion, but it is really interesting.

One possibility is that version 19 takes advantage of the same weakness/bug that Lee Sedol leveraged in game 4, and does it reliably / repeatedly. Then it wouldn't mean too much beyond the fact that it isn't as vulnerable to branching, hard-to-read fights, with the occasional 1/10000 move thrown in.

But another possibility is that version 19 is just significantly stronger. Maybe not four stones versus a human pro better, but two wouldn't be that shocking, and who knows about three?

Perhaps AlphaGo will soon be revolutionizing human go? It is quite exciting to consider. This could be the best time in history to be a go fan.

[Jeromin]

I hate deepmind. They don't say anything while they know the hype is great. It's like a scummy marketing campaign.

Would it be so hard for Demis or someone from the team to simply write one message on twitter or something simply explaining if they are still working on it and how much it has improved.

It would take him 2 minutes and it would confirm/denounce all the rumors on the interwebz.

I agree. In the end, it seems like it will be up to the Zen team and similar projects to slowly replicate the accomplishments Google Deepmind is being so reluctant to share. They're not a non profit foundation sure, but it's not the Google style.
A 10/11 dan silicon opponent running on reasonable hardware would have been immensely interesting and useful for pros and the Go community in general.
No software, no AlphaGo vs Alphago games, no rematch with top pros, no news... It's like they're saying look how great we are, now get off our property!

[pookpooi]

Didn't Lee Sedol (before his match with AlphaGo) say to the AlphaGo v13 that he can give 3 stones to it and still won, so AlphaGo v18 is 4 stones stronger than v13 make sense to me. I don't know about this v19 though, but this is what Aja Huang said last week.

[Charles Matthews]

In the end, it seems like it will be up to the Zen team and similar projects to slowly replicate the accomplishments Google Deepmind is being so reluctant to share. They're not a non profit foundation sure, but it's not the Google style.

A 10/11 dan silicon opponent running on reasonable hardware would have been immensely interesting and useful for pros and the Go community in general.

Yes but ... you mean, several months have gone by, and we still don't have the thing at the top of everyone's wishlist, gratis. Why don't you ask for something really unreasonable? You might get that too.

No software, no AlphaGo vs Alphago games, no rematch with top pros, no news... It's like they're saying look how great we are, now get off our property!

From the little I know, there isn't necessarily the management time. It is a perfectly reasonable plan to allow the bot to play itself until a plateau is found, surely.

[Babelardus]

I hate deepmind. They don't say anything while they know the hype is great. It's like a scummy marketing campaign.

Would it be so hard for Demis or someone from the team to simply write one message on twitter or something simply explaining if they are still working on it and how much it has improved.

It would take him 2 minutes and it would confirm/denounce all the rumors on the interwebz.

I agree. In the end, it seems like it will be up to the Zen team and similar projects to slowly replicate the accomplishments Google Deepmind is being so reluctant to share. They're not a non profit foundation sure, but it's not the Google style.
A 10/11 dan silicon opponent running on reasonable hardware would have been immensely interesting and useful for pros and the Go community in general.
No software, no AlphaGo vs Alphago games, no rematch with top pros, no news... It's like they're saying look how great we are, now get off our property!

It's no different than it was with Deep Blue back in 1997.

That was a supercomputer with special, dedicated chips to play chess. While AlphaGo doesn't use dedicated Go hardware, it does use hardware developed by Google (the Tensor processor), which is available no-where else. Even if they would release the software, it would probably take a a lot of effort to rewrite the algorithms to run on common hardware, and it can't even be guaranteed that the algorithms would work well.

The only thing Google could do is disclose if they used any programming techniques that haven't been used before (like Monte Carlo, 2004-2006), and how they did it. Then these techniques could be incorporated in current engines that run on normal hardware.

[Kirby]

AlphaGo running on commodity hardware would probably still be pretty strong.

[yoyoma]

Searched twitter for alphago to see people tweeting about Aja Huang's speech. The only information I found was this Japanese tweeter:
https://twitter.com/1never

自分の発表終わって今からAlphaGoの第2著者の方の招待講演
現在のAlphaGoはNatureバージョンよりも相当強い
AlphaGo vsイ・セドルの第4戦での問題はすでに解決済み
次のAlphaGoの対戦相手に関してはノーコメント，ただし現在何かしらの協議中
AlphaGoはTensorFlowを使ってるわけではないらしい

Paraphrasing from google translate I get:
Current AlphaGo is much stronger than the Nature version.
The problem in game 4 vs Lee Sedol has already been resolved.
No comment on AlphaGo's next opponent, but we are in discussions about it.
Seems like AlphaGo doesn't use TensorFlow (an open source library by Google)

A tweet from another Japanese source said he will blog soon, but that there really wasn't much new information.

[Babelardus]

Seems like AlphaGo doesn't use TensorFlow (an open source library by Google)

According to this article, AlphaGo does use TensorFlow.

[yoyoma]

Rumor seems to be refuted:
http://computer-go.org/pipermail/comput ... 09464.html
http://www.dgob.de/yabbse/index.php?act ... 5047;image

A talk by Google, you can see a picture of a slide

* Current version stronger than Lee Sedol version but definitely not four stones!

[pookpooi]

"still considering what to do next"
How about triple challenge to ke jie, park jungwhan and iyama yuta (each are no.1 of strong go base country.)