AlphaZero - Stockfish game 3

[TommyTC]

SF analysis in this position is completely misguided.

I propose the following human play after 49. ... Ke8. Instead of 50. Qf4 you should look at:
50. Bd5! Qg8
51. Qa4! +-
51. ... Rd8 (if 51. ... d6 to protect a7 then all the Re6 defense ideas fail for black and white wins easily and if one sneakily tries to distract whites queen from a7 with 51. ... Qh8 then after Qf4! Qg8 Qc7 Qh8 Qd6! - watch the pathetic black Queen - black also wins easily as he won 2 important tempi. SF is utterly confused here probably due to null move pruning)
52. Qxa7 Qh8 (what else as Ke8 looses another pawn)
53. Qa5 Ke8
54. Qc3 Qf8
55. a4 +- and the position plays itself

I agree with your assessment and that 50. Bd5! is an improvement.

SF is clearly confused in this position. She is an exchange + 2 pawns ahead, yet is smart enough to know that White has at least compensation so displays 0.00.

I also agree completely that SF is pruning a bit too much at the lower search depths, perhaps as you stated because of "null move pruning", although even LMR (?) might also be to blame since White has 40 possible moves in the starting position. I found the following variation interesting starting with the position after 49...Kf8:

4rk1q/p2prp1p/5RpP/2p5/7Q/1B4P1/P4PK1/8 w - - 0 50

50. Bd5 Rc8 51. Qf4 Qg8 52. Qa4 d6 53. Qa6 Rd8

From depth 40 to about depth 50 (sometimes later), SF says 0.00, then at depth 51 the eval starts creeping up: 0.14, 0.56, 1.13, 1.72, until depth 51 is completed. The best move is 54. Qa5. I cleared hash tables and tried it again. This time similar results but it showed 54. g4 as the best move. I then played 54. Qa5 and it immediately evaluated it at 1.72.

There are still moves to examine before saying this position is completely winning for White, but it is clear that White has an advantage after 49...Kf8. One possible try is: 49...Kf8 50. Bd5 Rc8 51. Qf4 Qg8 52. Qa4, and I've started looking at 52...Qh8 and 52...Ke8, since 52...d6 seems to be losing.

[IQ]

4rk1q/p2prp1p/5RpP/2p5/7Q/1B4P1/P4PK1/8 w - - 0 50

50. Bd5 Rc8 51. Qf4 Qg8 52. Qa4 d6 53. Qa6 Rd8

From depth 40 to about depth 50 (sometimes later), SF says 0.00, then at depth 51 the eval starts creeping up: 0.14, 0.56, 1.13, 1.72, until depth 51 is completed. The best move is 54. Qa5. I cleared hash tables and tried it again. This time similar results but it showed 54. g4 as the best move. I then played 54. Qa5 and it immediately evaluated it at 1.72.

There are still moves to examine before saying this position is completely winning for White, but it is clear that White has an advantage after 49...Kf8. One possible try is: 49...Kf8 50. Bd5 Rc8 51. Qf4 Qg8 52. Qa4, and I've started looking at 52...Qh8 and 52...Ke8, since 52...d6 seems to be losing.

If I remember my analysis correctly in the 50. ... Rc8 line after 51. Qf4 Qg8 52. Qa4 is not accurate .... i think here black draws. My idea here was 52. Bb7! Rce8 (only move and therefore white delayed the whole Ke8 freeing manover) and now 53. Qd6 c4 (everything else is worse) and 54. a4! + /- now that black can't free himself we just advance the a-pawn (remember we will sooner or later take a7 and do not worry about the c4 pawn, we take it under preferrable circumstances) - i analyzed all resulting endings to a white win, and would be hugely surprised if there is a miracle defense.

[IQ]

SF is clearly confused in this position. She is an exchange + 2 pawns ahead, yet is smart enough to know that White has at least compensation so displays 0.00.

I am afraid that SF is not "smart" enough to see the compensation but rather "blind" enough to not see how white makes progress and thus confuses himself with a number of threefold repititions.

[clumma]

Threads. They used a 32 core machine but ran it with 64 threads, the little rascals.

Source for this claim?