AlphaZero is a hoax.... (READ)

@mdinnerspace I mean we understood that the optimal SF8 version would have done better. But where did you get this 2200-2300elo performance from ?

I will have to go back to locate the source. It came from discussions of the match by contributors who ran the games through various engines. Just look at the game linked here, 3 inaccuracies, 4 mistakes, 1 blunder. A typical game. The limitations placed on SF rendered it at less than it's optimal strength. It was playing questionable moves every game in critical positions. Most likely due in main to the 1 minute/move time constraint and the lack of an opening tablebase; a primary strength with out which SF finds itself in less than optimal positions.
AlphaZero had full access to all of its hundreds of millions self taught practice games and choose a very select opening repertoire that it favored. It used these openings to repeatedly take advantage of SF early mistakes.

I am not "defending" SF nor trying to negate the tremendous job the programmers did with DeepMind. In fact just the opposite. Standard programming used for SF, Komodo, Houdini imo just makes them "glorified calculators", "blunder checkers" and so on. They lack strategical awareness and can not be "trusted" in much of their evaluations.
The self taught logarithms that simulate "human intuition" used by the programmers of DeepMind are clearly superior and most likely soon out date the way current engines are programmed.
However, AlphaZero posses no evaluation function. As it stands, the only things learned in chess theory are done by reviewing the games. AlphaZero has nothing to say except making it's moves.

I fail to see how AlphaZero is any more useful than Stockfish, at least at present.

AlphaZero's evaluation is based upon MCTS+NN scores, so its evaluation function could be expressed in terms of those scores.

In the original AlphaGo zero paper they included a graph of the relative strength when they switched off the MCTS part and just played the highest valued move of the the neural nets first pass. This halved the Go ELO but amazingly still outplayed the human crafted opponents. I can't wait for the full paper on the chess version, what if the 'eval' network it produces can reveal new features to us when we peer inside ml4a.github.io/ml4a/looking_inside_neural_nets/

A proof :

Let's take the 6th game.

60. ... Rc5

That's awful! Loss of 6 points (according to Lichess) !

34. ... Knight d4

One point loss.

5rk1/6p1/pp4P1/P1r1RPK1/2PR4/8/8/8 w - - 1 61

is the link for 60. ... Rc5

1r1r3k/p2p2p1/1ppq1pBp/8/n1Pn1PQP/4R1P1/P5K1/2B1R3 w - - 3 35

is the link for 34. ... Knight d4

The difference between hardware 'horse power' between A0 and stockfish is gigantic, see post #14, but stockfish can only use maximal 64 cpu cores and can not be accelerated with gpus.
I can not see a hoax here, just the clever use of supercomputer calculating power, but it was an unfair match, because if A0 use only equivalent performance to 64 cpu cores and stockfish run with optimal settings, then stockfish will win, i guess.

@Ervin259 Stockfish changes its mind as it thinks deeper. The lichess analysis is generally not very deep (though more than good enough for most human purposes).

I am able to replicate all of the "bad" moves people claim SF made during this match by letting my own computer run long enough. My computer is fairly powerful - about half of the speed of what Google was using and much more than what you'll get from a lichess computer analysis.