'Mittens ChessBot' (Chess.com) Stuns Chessworld

[towforce]

BIG CHESS NEWS: ALPHA ZERO (16) REAPPEARS TO PLAY 'MITTENS".....WOW!!

https://www.youtube.com/watch?v=_0o7pkxvFHk

[Event "10 mins Chess.Com"]
[Site "Chess.com"]
[Date "2023.01.07"]
[Round "?"]
[White "Mittens Bot (Chess.com)"]
[Black "Alpha Zero 16"]
[Result "0-1"]
[ECO "A47"]
[WhiteElo "3500"]
[BlackElo "4000"]
[Annotator "Ross-SF Ipman Se-3"]
[PlyCount "82"]
[TimeControl "10 mins"][pgn]1. d4 b6 ?! {
Alpha Zero 16 Throws Books Out!!} 2. Nf3 Bb7 3. Bf4 Nf6 4. e3 Nh5 !? 5. Bg5 (5.
Nbd2 ! Nxf4 6. exf4 e6 7. Bd3 f5 8. O-O Be7 9. Qe2 O-O 10. c3 Nc6 11. Ne5 ! {
London System....} Nxe5 {[]}) 5... h6 6. Bh4 g5 7. Bg3 d6 8. Bd3 (
8. h4 !? Nxg3 9. fxg3 g4 10. Ng1 ?! h5 11. Ne2) 8... e6 9. O-O (9. c4 !? Nd7
10. Nc3) 9... g4 10. Nfd2 (10. Bh4 ?! Bxf3 11. Bxd8 Bxd1 12. Bh4 Be7 13. Bxe7
Bxc2 14. Bxc2 Kxe7 15. Nc3) 10... Nxg3 11. fxg3 h5 12. Qe1 Bh6 13. Qf2 Qe7 14.
e4 (14. Nc3 Nc6 15. Rae1 Nb4 16. a3 Nxd3 17. cxd3) 14... Nc6 15. Nb3 (15. c3
O-O-O) 15... O-O-O 16. Nc3 (16. N1d2 Rdf8 17. c3 Kb8) 16... Rdf8 17. Rae1 Kb8
18. a3 (18. a4 Qg5) 18... Qg5 19. Nd1 (19. d5 Ne7 20. a4) 19... Qg7 20. c3 Bg5
21. Qe2 h4 !{Crushing...} 22. gxh4 Bxh4 23. g3 f5 24. gxh4 Rxh4 25. Qg2 Rfh8 26. Rf2 Qh6
27. Kf1 (27. Nd2 Rh3 28. Bf1) 27... f4 28. a4 e5 29. Bb5 (29. d5 Ne7 30. a5 Rh3
) 29... a6 30. Bc4 Ne7 31. dxe5 dxe5 32. a5 b5 33. Bf7 (33. Bd3 Rh3 34. Nc5 g3)
33... Nf5 34. Nd2 g3 35. Nf3 gxf2 36. Nxf2 Ne3+ 37. Rxe3 fxe3 38. Nxh4 Qxh4 39.
Qg3 exf2 40. Qxh4 Rxh4 41. h3 Rf4 0-1
[/pgn] & Black won....Alpha Zero!!

AlphaZero has never been released to the public. I wonder who the keeper is who can wheel it out against great new Chess.com bots?

[towforce]

chess.com have told Hikaru that they've been monkeying with Mittens' algorithm over the last few days, and this is why there have been strange and inexplicably bad moves:



Hikaru is not impressed with this behaviour!

Edit: maybe my theory is correct, and the Mittens engine runs on a server (so that an NN can be used)? If so, the publicity Mittens has generated will have resulted in high usage, and those server costs might have been higher than they were expecting (or their servers ran out of capacity).

[Werewolf]

BIG CHESS NEWS: ALPHA ZERO (16) REAPPEARS TO PLAY 'MITTENS".....WOW!!

https://www.youtube.com/watch?v=_0o7pkxvFHk

[Event "10 mins Chess.Com"]
[Site "Chess.com"]
[Date "2023.01.07"]
[Round "?"]
[White "Mittens Bot (Chess.com)"]
[Black "Alpha Zero 16"]
[Result "0-1"]
[ECO "A47"]
[WhiteElo "3500"]
[BlackElo "4000"]
[Annotator "Ross-SF Ipman Se-3"]
[PlyCount "82"]
[TimeControl "10 mins"][pgn]1. d4 b6 ?! {
Alpha Zero 16 Throws Books Out!!} 2. Nf3 Bb7 3. Bf4 Nf6 4. e3 Nh5 !? 5. Bg5 (5.
Nbd2 ! Nxf4 6. exf4 e6 7. Bd3 f5 8. O-O Be7 9. Qe2 O-O 10. c3 Nc6 11. Ne5 ! {
London System....} Nxe5 {[]}) 5... h6 6. Bh4 g5 7. Bg3 d6 8. Bd3 (
8. h4 !? Nxg3 9. fxg3 g4 10. Ng1 ?! h5 11. Ne2) 8... e6 9. O-O (9. c4 !? Nd7
10. Nc3) 9... g4 10. Nfd2 (10. Bh4 ?! Bxf3 11. Bxd8 Bxd1 12. Bh4 Be7 13. Bxe7
Bxc2 14. Bxc2 Kxe7 15. Nc3) 10... Nxg3 11. fxg3 h5 12. Qe1 Bh6 13. Qf2 Qe7 14.
e4 (14. Nc3 Nc6 15. Rae1 Nb4 16. a3 Nxd3 17. cxd3) 14... Nc6 15. Nb3 (15. c3
O-O-O) 15... O-O-O 16. Nc3 (16. N1d2 Rdf8 17. c3 Kb8) 16... Rdf8 17. Rae1 Kb8
18. a3 (18. a4 Qg5) 18... Qg5 19. Nd1 (19. d5 Ne7 20. a4) 19... Qg7 20. c3 Bg5
21. Qe2 h4 !{Crushing...} 22. gxh4 Bxh4 23. g3 f5 24. gxh4 Rxh4 25. Qg2 Rfh8 26. Rf2 Qh6
27. Kf1 (27. Nd2 Rh3 28. Bf1) 27... f4 28. a4 e5 29. Bb5 (29. d5 Ne7 30. a5 Rh3
) 29... a6 30. Bc4 Ne7 31. dxe5 dxe5 32. a5 b5 33. Bf7 (33. Bd3 Rh3 34. Nc5 g3)
33... Nf5 34. Nd2 g3 35. Nf3 gxf2 36. Nxf2 Ne3+ 37. Rxe3 fxe3 38. Nxh4 Qxh4 39.
Qg3 exf2 40. Qxh4 Rxh4 41. h3 Rf4 0-1
[/pgn] & Black won....Alpha Zero!!

AlphaZero has never been released to the public. I wonder who the keeper is who can wheel it out against great new Chess.com bots?

Very interesting. I'd assumed this was a dead project now

[AndrewGrant]

BIG CHESS NEWS: ALPHA ZERO (16) REAPPEARS TO PLAY 'MITTENS".....WOW!!

https://www.youtube.com/watch?v=_0o7pkxvFHk

[Event "10 mins Chess.Com"]
[Site "Chess.com"]
[Date "2023.01.07"]
[Round "?"]
[White "Mittens Bot (Chess.com)"]
[Black "Alpha Zero 16"]
[Result "0-1"]
[ECO "A47"]
[WhiteElo "3500"]
[BlackElo "4000"]
[Annotator "Ross-SF Ipman Se-3"]
[PlyCount "82"]
[TimeControl "10 mins"][pgn]1. d4 b6 ?! {
Alpha Zero 16 Throws Books Out!!} 2. Nf3 Bb7 3. Bf4 Nf6 4. e3 Nh5 !? 5. Bg5 (5.
Nbd2 ! Nxf4 6. exf4 e6 7. Bd3 f5 8. O-O Be7 9. Qe2 O-O 10. c3 Nc6 11. Ne5 ! {
London System....} Nxe5 {[]}) 5... h6 6. Bh4 g5 7. Bg3 d6 8. Bd3 (
8. h4 !? Nxg3 9. fxg3 g4 10. Ng1 ?! h5 11. Ne2) 8... e6 9. O-O (9. c4 !? Nd7
10. Nc3) 9... g4 10. Nfd2 (10. Bh4 ?! Bxf3 11. Bxd8 Bxd1 12. Bh4 Be7 13. Bxe7
Bxc2 14. Bxc2 Kxe7 15. Nc3) 10... Nxg3 11. fxg3 h5 12. Qe1 Bh6 13. Qf2 Qe7 14.
e4 (14. Nc3 Nc6 15. Rae1 Nb4 16. a3 Nxd3 17. cxd3) 14... Nc6 15. Nb3 (15. c3
O-O-O) 15... O-O-O 16. Nc3 (16. N1d2 Rdf8 17. c3 Kb8) 16... Rdf8 17. Rae1 Kb8
18. a3 (18. a4 Qg5) 18... Qg5 19. Nd1 (19. d5 Ne7 20. a4) 19... Qg7 20. c3 Bg5
21. Qe2 h4 !{Crushing...} 22. gxh4 Bxh4 23. g3 f5 24. gxh4 Rxh4 25. Qg2 Rfh8 26. Rf2 Qh6
27. Kf1 (27. Nd2 Rh3 28. Bf1) 27... f4 28. a4 e5 29. Bb5 (29. d5 Ne7 30. a5 Rh3
) 29... a6 30. Bc4 Ne7 31. dxe5 dxe5 32. a5 b5 33. Bf7 (33. Bd3 Rh3 34. Nc5 g3)
33... Nf5 34. Nd2 g3 35. Nf3 gxf2 36. Nxf2 Ne3+ 37. Rxe3 fxe3 38. Nxh4 Qxh4 39.
Qg3 exf2 40. Qxh4 Rxh4 41. h3 Rf4 0-1
[/pgn] & Black won....Alpha Zero!!

AlphaZero has never been released to the public. I wonder who the keeper is who can wheel it out against great new Chess.com bots?

Very interesting. I'd assumed this was a dead project now

Almost certainly a youtuber running a random engine, calling it AlphaZero, and farming content. Which seems to work for him.

[Jesse Gersenson]

Edit: maybe my theory is correct, and the Mittens engine runs on a server (so that an NN can be used)? If so, the publicity Mittens has generated will have resulted in high usage, and those server costs might have been higher than they were expecting (or their servers ran out of capacity).

If it were running on a server, and too many people were using the engine move choices would not be affected but the engine would lose games on time.

The cost to run the engine, if it is on a server, is near-zero.

All of their other "personality bots" are WASM.

[towforce]

Edit: maybe my theory is correct, and the Mittens engine runs on a server (so that an NN can be used)? If so, the publicity Mittens has generated will have resulted in high usage, and those server costs might have been higher than they were expecting (or their servers ran out of capacity).

If it were running on a server, and too many people were using the engine move choices would not be affected but the engine would lose games on time.

Not necessarily: Hikaru said that chess.com had turned down Mittens' level: this implies to me a lower amount of CPU time per move, which further implies a reduced depth of search.

The cost to run the engine, if it is on a server, is near-zero.

If every Mittens user needs exclusive use of a powerful GPU to run the NN quickly, the cost could be high.

All of their other "personality bots" are WASM.

Ah - thanks for that. My knowledge was outdated!

I might be wrong, but given that WASM code runs in the browser, that many chess players won't have a high-specification GPU, and that downloading a large NN into the browser is likely to be problematical for some chess.com users, it seems unlikely to me that Mittens is a WASM bot: far more likely that, in order to get the strength and consistency, they made Mittens' move selection something that's done on a server. This would neatly explain:

1. The sudden jump in playing strength

2. Why chess.com would choose to (temporarily according to Hikaru) throttle her playing strength

Of course, this is speculation on my part: I would welcome any actual information anybody is willing and able to share.

[towforce]

OK - so maybe Mittens is a WASM program like the other bots (I cannot see any webservice calls as I play Mittens - but I'm not 100% sure that you still do see them under WASM).

I am making the following new guesses:

* Mittens is a WASM program

* to get the extra strength, a 20-40 mb NN is downloaded into the WASM in the browser

* no attempt is made to use the user's graphics card to speed up NN processing

However, if this guess is correct, then I'm mystified as to why chess.com would turn down Mittens' strength?

[jdart]

I watched a video of GM Nakamura beating this bot with pawn odds in his favor:


He had several difficult decisions to make during the game and probably made some inaccuracies, but in the end he won.
What most impressed me is that a lot of time he was calculating, he wasn't looking at the board.

[carldaman]

Noticed the same thing myself about Hikaru. It's possible that if he's a good visualizer, then looking at the current board position could get in the way of his calculations, so he needs to look away.

[Aldus]

I watched a video of GM Nakamura beating this bot with pawn odds in his favor:


He had several difficult decisions to make during the game and probably made some inaccuracies, but in the end he won.
What most impressed me is that a lot of time he was calculating, he wasn't looking at the board.

He said somewhere (I don't remember the clip) he vividly sees the chessboard with his eyes closed. Many GMs have a photographic memory.

Here: