GM Andrew Tang vs Leela Chess Zero

@DVRazor

Not sure about whether Leela beat Fruit 2.1, but remember that algorithmically Leela is not so different from AlphaZero (both use reinforcement learning), which obtained an ELO > Stockfish with under 4 hours of training, and dominated Stockfish in its match.

So I agree this is all very remarkable, but it should not be surprising that an open source implementation of an algorithm similar to AlphaZero's can beat a 2695.

Sure, but Google used so much calculating power, Leela's devs say they need years to catch up on that to get Leela to the same level.

There are some differences between Leela and Alphazero that should be considered. First, Google used specialized hardware for both trainning and interface (not CPUs or GPUs, as it is with Leela). I'm impressed that she could beat an IM running on domestic hardware.

Second, A0 software was actually the work of years, because it all started in 2015 with the first AlphaGo, which wasn't stronger than the best Go players at the time. They made much more progress in terms of software than people realise.

So, at this point Leela probably hasn't done half the ammount of steps that A0 did in 4 hours, and each of her individual steps made less progress aswell.

@asymptoticfreedom I suggest you to read the paper: arxiv.org/pdf/1712.01815.pdf in detail and not advertisement news articles written either to promote the paper or to make some sort of sensation.

4 hours is not the 4 hours a usual person could imagine. It is not even 4 hours at all.

What it actually meant was: "using 5,000 first-generation TPUs to generate self-play games and 64 second-generation TPUs to train the neural networks". Then type 'tpu vs gpu comparison' and check how does it compare (estimates tell that TPUs are 10-20x faster). By the way the training time was 9h and not 4h. 4h is when they claimed that it was finally better than stockfish (not mentioning by how much better).

Btw if you will play the new version of the new version of stockfish vs the old version of stockfish (Stockfish 9 64-bit 4CPU vs Stockfish 8 64-bit 4CPU: http://www.computerchess.org.uk/ccrl/4040/rating_list_all.html) then judging by this caclulator for (3444 vs 3387) you will get approximately the following score from 100 matches:

Win 20 games, draw 75 games, loose 4 games. So now the unproven (do not forget that there is no way to verify their results. They could just ran 1mln stockfish vs stockfish games and selected a few with spectacular wins) 28, 72, 0 do not look so sensational.

The main point is: when someone sells you something - try it on your own or at least read the actual specification instead of relying on flashy brochures designed to sell you this cool new thing.

@nikolajtesla
Thank you for your rational and informative comment. As you see most people doesn't care about the details only the buzzwords.

This match can be extremely easy for the Gm depending on 2 elements how fast the computer is and how the GM wants to play quality or quantity games for example for one strategy could be to play random and lose quickly the first 3 phases of the match and be behind 0-14 then for the remainder of the time which will probably be 1 h 40 min depending how quickly the 3 phases go and then in the 4th match challenge the computer to ultra bullet for the remaining amount of the time and if the computer is not super fast crush it like 25-0 so the final score could be 25-14 in favor of the Gm if he chooses that strategy :P

Go 唐 Go

WHat time control do you think tang will chose?

ultra as it gives him the advantage plus he can get more games in

I won this program in superblitz mode :)
http://play.lczero.org/

Move History
1. e4 c5
2. Nc3 Nc6
3. g3 Nf6
4. Bg2 e6
5. Nge2 d5
6. exd5 exd5
7. d4 cxd4
8. Nxd4 Be7
9. Bf4 O-O
10. Qd2 Bc5
11. Nxc6 bxc6
12. O-O Ba6
13. Rfe1 Ng4
14. Nd1 Qb6
15. c3 Rfe8
16. Bf3 Rxe1+
17. Qxe1 Nf6
18. Qd2 Ne4
19. Bxe4 dxe4
20. b4 Be7
21. Be3 Qb5
22. Nb2 Qe2
23. Qxe2 Bxe2
24. Re1 Bf3
25. Nc4 Bf6
26. Bd4 Rd8
27. Bxf6 gxf6
28. Na5 Rd3
29. Nxc6 Rxc3
30. Nxa7 Rc2
31. a4 e3
32. fxe3 h5
33. Nb5 Rg2+
34. Kf1 Rxh2
35. Nd4 Be4
36. a5 Bd3+
37. Kg1 Rb2
38. Rd1 Be4
39. b5 Rg2+
40. Kf1 Rxg3
41. a6 Rg2
42. b6 Rb2
43. b7 Bxb7
44. axb7 Rxb7
45. Kf2 Rb2+
46. Kf3 h4
47. Rh1 h3
48. Rxh3 Kg7
49. Rg3+ Kf8
50. Nf5 Rb5
51. Kf4 Rb4+
52. e4 Ke8
53. Rg8+ Kd7
54. Rf8 Ke6
55. Re8+ Kd7
56. Re7+ Kc8
57. Rxf7 Kd8
58. Rxf6 Kd7
59. Rd6+ Kc7
60. Rd4 Rxd4
61. Nxd4 Kd6
62. Kf5 Kc5
63. e5 Kxd4
64. e6 Kd5
65. e7 Kd6
66. e8=Q Kc7
67. Qe6 Kd8
68. Qf7 Kc8
69. Ke6 Kb8
70. Kd6 Ka8
71. Kc6 Kb8
72. Qb7#