leela is official(?) better than sf9

[glennsamuel32]

thats pretty much an imputation from my personal point of view and my experience tells me not to do so if there is no evidence for such claims..to tell people that google manipulated the entries to get a result that was intended to get because otherwise it would not have been possible or likely is not the right way to talkchess imho...

Maybe you or the Google team or the Leela team can answer these questions I have...

1. When DeepMind challenged the Go players, it was in front of TV cameras.
What happened with Stockfish ??
A private competition with no pgn, no depth / score / time analysis ??
Just a half-ass pdf as a research paper with only the moves of the games A0 won ??

2. They decided not to challenge the current SF dev version which was clearly 50+ elo stronger ??

3. They allocated 1 gb for 64 threads -- because they ran out of memory ??
How much memory was A0 using for it's analysis tree ??

"leela is official(?) better than sf9"

I've already mentally dumped this thread into the BS section

[Spliffjiffer]

im not a personality that is involved in google or Lcx...
1. the competition between A0 and SF8 might be a hoax that never took place...imo it took place which is a pure question of believe or not believe..

2. imo SF8 was "new enough" to show the intimidation of this kind of "new" approach...isnt it???

3. the allocation of (only) 1GB ram for a 64-threaded a-b-engine is imo questionable as well...this hash-size is definately too small for 1 min thinking-time by using 64 threads (imo)...OTOH: i tested a-b-engines very very much even considering to change "hash-size" and "normally" the impact of finding "the better move" is marginal (considering as well that most lost positions didnt came from mishandeling endgames where hash is getting more important AFAI do consider it nowadays)...thats my experience with hash-sizes when they go beyond 1GB (still 256 mb make in most cases excellent results !!) but oc 64 threads is beyond my experiences !!..my personal conclusion: 1GB was inadequate but not really result-relevant

[Tobber]

1. When DeepMind challenged the Go players, it was in front of TV cameras.
What happened with Stockfish ??
A private competition with no pgn, no depth / score / time analysis ??
Just a half-ass pdf as a research paper with only the moves of the games A0 won ??

2. They decided not to challenge the current SF dev version which was clearly 50+ elo stronger ??

1. Not interesting enough, to beat a human world champion which no engine had done before that makes headlines.
Make a chess engine which is stronger than another chess engine, big deal, so what? DeepMind/Google is not interested in chess engines. The total market for chess engines is probably smaller than what Google spends on toilet paper in their office.

2. They used what was considered the stable version.

/John

[jp]

A private competition with no pgn, no depth / score / time analysis ??
Just a half-ass pdf as a research paper with only the moves of the games A0 won ??

DeepMind/Google is not interested in chess engines. The total market for chess engines is probably smaller than what Google spends on toilet paper in their office.

John, it would have been & continues to be very easy (almost zero effort needed) for DeepMind to make the program or at least the weight files public. Even easier (really zero effort) for the game scores and game logs. Please stop making excuses for them.

[Tobber]

A private competition with no pgn, no depth / score / time analysis ??
Just a half-ass pdf as a research paper with only the moves of the games A0 won ??

DeepMind/Google is not interested in chess engines. The total market for chess engines is probably smaller than what Google spends on toilet paper in their office.

John, it would have been & continues to be very easy (zero effort needed) for DeepMind to make the program or at least the weight files public. Please stop making excuses for them.

I'm not making excuses but you are blinded by your interest in chess engines. It's a very small and uninteresting part of the IT market. My guess is that they made it basically for internal use, maybe to show some new people how stuff works. Just a spin off from AlphaGo.
They are a commercial company, they publish what they want but after AlphaGo it's very likely they have the capability to really do something like A0.
/John

[jp]

I'm not making excuses but you are blinded by your interest in chess engines. It's a very small and uninteresting part of the IT market. My guess is that they made it basically for internal use, maybe to show some new people how stuff works. Just a spin off from AlphaGo.
They are a commercial company, they publish what they want but after AlphaGo it's very likely they have the capability to really do something like A0.
/John

No, I'm not blinded. Either you didn't realize different people are replying to you or you think everyone who replies to you is the same. Not that the other repliers are "blinded by interest in chess engines" either. They are just pointing out obvious problems. You can only claim people are "blinded by interest in chess engines" if they are demanding that some commercial company continues to work on a chess engine forever. No one ever demanded that.

Your guess is not a good one. If they did what you guess, they wouldn't write anything. Why bother writing at all? It takes a lot more effort to write what they did than to release a data file. Releasing a data file takes ZERO effort.

[Tobber]

I'm not making excuses but you are blinded by your interest in chess engines. It's a very small and uninteresting part of the IT market. My guess is that they made it basically for internal use, maybe to show some new people how stuff works. Just a spin off from AlphaGo.
They are a commercial company, they publish what they want but after AlphaGo it's very likely they have the capability to really do something like A0.
/John

No, I'm not blinded. Either you didn't realize different people are replying to you or you think everyone who replies to you is the same. Not that the other repliers are "blinded by interest in chess engines" either. They are just pointing out obvious problems. You can only claim people are "blinded by interest in chess engines" if they are demanding that some commercial company continues to work on a chess engine forever. No one ever demanded that.

Your guess is not a good one. If they did what you guess, they wouldn't write anything. Why bother writing at all? It takes a lot more effort to write what they did than to release a data file. Releasing a data file takes ZERO effort.

Can you give a good reason why they should release it? The paper written was a minimal report but claiming, as some people do, that they faked it is just silly. Lc0 has showed that the concept is working and DeepMind has access to some heavy hardware. Even for internal use they must show somebody what they have done and releasing an early version of the paper takes no effort. As I just wrote, defeating a human world champion is good publicity, defeating just another chess engine is not.
/John

[AW~]

Your guess is not a good one. If they did what you guess, they wouldn't write anything. Why bother writing at all? It takes a lot more effort to write what they did than to release a data file. Releasing a data file takes ZERO effort.

What do you think happens if more info is released? What would change?

[Laskos]

Your guess is not a good one. If they did what you guess, they wouldn't write anything. Why bother writing at all? It takes a lot more effort to write what they did than to release a data file. Releasing a data file takes ZERO effort.

What do you think happens if more info is released? What would change?

Maybe our expectations from Lc0. It's possible that Lc0 with 11xxx nets were already close to A0 on the same hardware. It's possible that A0 was tactically weak (extremely weak in tactical test-suites) and weak in endgames too, like Lc0. It's possible that using a tricky book for SF8 with exits to very tactical games, A0 would have not won easily against SF8 in DeepMind conditions. Having more info about actual games, about some tuning and training parameters used in that paper would certainly have helped. Also, the selection of openings and the selection of shown full games are skewed. It happens often in science literature, but it's not well received when the flaws and the bias are shown when reproducing the results.

[AW~]

Maybe our expectations from Lc0. It's possible that Lc0 with 11xxx nets were already close to A0 on the same hardware. It's possible that A0 was tactically weak (extremely weak in tactical test-suites) and weak in endgames too, like Lc0. It's possible that using a tricky book for SF8 with exits to very tactical games, A0 would have not won easily against SF8 in DeepMind conditions. Having more info about actual games, about some tuning and training parameters used in that paper would certainly have helped. Also, the selection of openings and the selection of shown full games are skewed. It happens often in science literature, but it's not well received when the flaws and the bias are shown when reproducing the results.

I dont expect anything from Leela so I guess that leaves me out. Im hopeful DeusX improves because at least its within range of some opening evals.
I dont see the value of another 1000 10 depth engine vs 10 depth engine games. IMO, both engines were handicapped.... something the SF supporters spend no time considering since that gives the advantage to A0. Only by having the A0 engine in hand do I think we can get to the truth.