David Silver (Deepmind) inaccuracies

[Cardoso]

I am happy I have watched no videos of theirs and read no papers of theirs. Very Happy.

Maybe that's why you are asking so many questions. Please read the paper. All good things take effort. Believe me, it's a very interesting paper.

[shrapnel]

I am happy I have watched no videos of theirs and read no papers of theirs.

I suspected as much.
For you, Ignorance, truly, is BLISS !

[M ANSARI]

I really don't understand why all the hate for A0. Clearly they chose SF because SF is considered the strongest chess engine. Nothing wrong with that assumption and while it might not have been TCEC champion, nothing I have seen changes the fact that we all know that SF is one of the strongest if not the strongest engine. There is probably less than 30 ELO difference in the strength of the top 3 engines so anyone of them would have been a good sparring partner. Remember that SF played on a 64 thread computer ... so very strong hardware. I do agree that not all the conditions were optimal for SF and that the time control of 1 minute per move or having no book probably tilted things a bit towards A0. But the point is that no other chess engine would have been able to perform similar to A0.

So let's acknowledge that something major has happened in the world of computer chess and of course we want further games and further proofs of how strong A0 is. The idea of using several GPU cards to run very fast real time MC games has always been an interesting idea. With billions of lines tested you are bound to hit a promising but obscure line which the brute force of CPU powered SF can quickly verify it is a good line. Hopefully this is something someone can make happen.

[Cardoso]

Or maybe a Sinclair ZX Spectrum
I had one of those, in my teenager days, and played chess on it. Of course I could win many games, but who couldn't on that machine?
Lyudmil has more than once denied the validity of both classic alphabeta basic concepts as well as the new AlphaZero approach. On a previous post he doesn't give much importance to branching factor and asked me if there are any differences between search and eval! I have advised him to study and learn how to program a chess engine, and any doubts would be gladly answered here. But he keeps bombarding us with disbelief, denial and sometimes with harsh criticism. And keeps wanting to discuss things he isn't (as far as I know) willing to dedicate some time to learn.

[Rebel]

#2. What does David Silver mean when he says: "this previous world champion"?

I think that he means that Stockfish 8 was the previous TCEC winner, which would be correct, although you can have doubts that TCEC is not a FIDE recognized championship. Houdini now being the TCEC champion. I don't think Mark Lefler would have liked much being Alpha Zero's guinea pig instead of SF, although he would have liked learning from the games.

To me it sounds as brutal commerce, we AZ are now WC

When you think Rebel was once close to being it...

What do you mean Rebel once was close. It would probably be helpful if we stuck to the facts.

One handsome guy! That is, back then.

[Milos]

The idea of using several GPU cards to run very fast real time MC games has always been an interesting idea. With billions of lines tested you are bound to hit a promising but obscure line which the brute force of CPU powered SF can quickly verify it is a good line. Hopefully this is something someone can make happen.

For the Nth time, A0 was running at exactly same machine as SF and that machine was used to actually run search and NN top level Tensorflow specialised data manipulation (what they didn't even bother to mention in the paper, btw. highly ethically questionable move). In addition to that, they used 4 TPUs that are highly specialised cutting edge research hardware for vector-matrix multiplication only. It's like you run SF plus 4 highest performing FPGA boards to run move-gen, eval and qsearch that would probably give SF tenfold nps boost.
So things are more than fishy with DeepMind. And considering the nature of all of that and multibillion dollar advertising company known for unethical behavior such as Google running the show, having doubts is only logical way. Ppl just believing things blindly are nothing but naive suckers.

[shrapnel]

So let's acknowledge that something major has happened in the world of computer chess and of course we want further games and further proofs of how strong A0 is.

Obviously.
But best of luck trying to convince folk like Milos and Tsvetkov, whose whole world seems to have turned upside down with the emergence of AO.
You could argue with them till you are blue in the face, but to no avail.
Perhaps they need help of a different sort.

[M ANSARI]

The idea of using several GPU cards to run very fast real time MC games has always been an interesting idea. With billions of lines tested you are bound to hit a promising but obscure line which the brute force of CPU powered SF can quickly verify it is a good line. Hopefully this is something someone can make happen.

For the Nth time, A0 was running at exactly same machine as SF and that machine was used to actually run search and NN top level Tensorflow specialised data manipulation (what they didn't even bother to mention in the paper, btw. highly ethically questionable move). In addition to that, they used 4 TPUs that are highly specialised cutting edge research hardware for vector-matrix multiplication only. It's like you run SF plus 4 highest performing FPGA boards to run move-gen, eval and qsearch that would probably give SF tenfold nps boost.
So things are more than fishy with DeepMind. And considering the nature of all of that and multibillion dollar advertising company known for unethical behavior such as Google running the show, having doubts is only logical way. Ppl just believing things blindly are nothing but naive suckers.

I think you misunderstood what I was trying to say. I meant that SF could maybe benefit by using multiple GPU cards to use the massive capability of modern GPU cards to do real time MC analysis on a position and maybe find a move that the CPU can then search and confirm as being good. I remember some time back a program called "Freezer" or "Fridge" or something like that ... it was really good at finding breakthroughs in closed positions and could avoid the "horizon effect". It used MC search and I was really impressed with it in endgame and closed positions. I do agree that DM was using a completely different type of hardware and thus cannot really be compared to the 64 thread hardware of SF. But some of the moves DM made would take a very long time for SF to see. Given enough time I think SF would see those good moves as all those "positional" moves by DM had a tactical confirmation of their viability. Maybe this was just a simple engine that was using tremendously powerful hardware to mimic AI, but that would mean that Google is part of massive fraud and lying to the public. I have a hard time believing that as it could easily backfire 1000x worse for something that is really insignificant to the general public, and has very little gain. I have to admit I am also a little bit skeptical about some things and would certainly want to see more games (especially the drawn games) ... but you have to remember that chess engines are programmed by inferior chess players (humans) and the idea that something better than a human can teach chess is not so outrageous.

[MikeGL]

#2. What does David Silver mean when he says: "this previous world champion"?

I think that he means that Stockfish 8 was the previous TCEC winner, which would be correct, although you can have doubts that TCEC is not a FIDE recognized championship. Houdini now being the TCEC champion. I don't think Mark Lefler would have liked much being Alpha Zero's guinea pig instead of SF, although he would have liked learning from the games.

To me it sounds as brutal commerce, we AZ are now WC

When you think Rebel was once close to being it...

What do you mean Rebel once was close. It would probably be helpful if we stuck to the facts.

One handsome guy! That is, back then.

+1
Agreed.
Inside and out

[Lyudmil Tsvetkov]

Why does not it do the very same with black, not a single king side fianchetto, Bg7?

Perhaps the fact chess is an asymmetric game might be part of the explanation.
Anyway NN have the specific disadvantage of not being able to explain their behaviour. So it's no use asking why A0 play this or that move because NN can't explain it's decisions. Not even the A0 team can explain A0's moves.

No part, Bg7 is also the best move for black, after taking control of center.