BOTW: 2018.11.30

Stockfish 10

Stockfish is now 10 years old! The open-source engine has been the world's strongest chess entity* for more than 5 years and continues to win every computer competition it enters. With contributors from all around the world, and a well-developed testing framework, Stockfish improves continuously and development versions appear almost daily. Once a year they commit to a new version name, and on November 29 they released Stockfish 10, so now is a good time to update.

The Stockfish team release compiles for different OS's -- PC, Linux, Mac, Android -- but also release different versions for PC which perform differently on various hardware. Intel i7 users should get the ones with the BMI2 instructions, while AMD users should stick with the POPCNT versions; YMMV, but you can always test each one on your own by using the -bench parameter.

*not counting the (mostly?) unavailable AlphaZero, see below.

AlphaZero Returns... sort of

AlphaZero is (probably) the strongest chess entity of all time. But it is also a science and PR project which is not available to the public. Last year the AlphaZero developers released a paper about it, along with 10 amazing games it played against Stockfish 8... and then it disappeared.

But during the Carlsen - Caruana World Championship match Demis Hassabis, head of Google DeepMind and the AlphaZero project, dropped in to watch and was interviewed at the start of game 8. You can watch the interview, but the tempting bits were, "we are going to release more games", and, "I think people are going to find some of these newer games, with an even stronger version of AlphaZero, quite fascinating, I hope." The latter suggests that DeepMind have done further training on AZ!?

We now know that the games they intend to release will be part of a book "Game Changer" by GM Matthew Sadler and WIM Natasha Regan, who previously collaborated on "Chess for Life". DeepMind has been letting Sadler use AZ to analyze chess, and Sadler has posted videos summarizing what he has discovered about the Carlsen-Caruana match games. It's not quite as interesting as you might hope, but it's hard to do better for a review of the match than a very strong GM using the world's strongest and most exclusive chess engine. You can find Sadler's videos here:

If you want more computer insight into the Carlsen - Caruana match, especially the amazing endgame in game 6 -- where Carlsen created a piece-down fortress which could have been broken, but not by any human -- see Ken Reagan's blog post.

Chess.Com Speed Chess Championship 2018: Semis and Final

The Semifinals and Final of the 2018 Chess.Com Speed Chess Championship take place November 30-December 2, 2018. The semis feature:

  1. GMs Wesley So vs Jan-Krzysztof Duda (Nov 30; 1pm EST)
  2. GMs Hikaru Nakamura vs Levon Aronian (Dec. 1; 1pm EST)

the winners meet in the Final Sunday, December 2, at 3pm EST.

Duda is a rating underdog vs So, but has already eliminated favourites Karjakin and Grischuk. Naka is the clear favourite to win it all. 

The format for the matches is 3 hours of online play, broken into three formats:

    • 90 minutes of 5+2 blitz, 
    • 60 minutes of 3+2 blitz,
    • 30 minutes of 1+1 bullet

Live Games (with IM Daniel Rensch and GM Robert Hess)

"Computer games are boring" ??

... is what people who don't look at computer games (used to?) say.

Here's a bonus game: Stockfish 10 vs Leela in a fascinating King's Indian where Leela does what other engines do not: play ...g4-g3 and sac LOTS of material for an uncalculatable attack. Both sides find amazing resources.

Nerd Info: 

  • White: Stockfish 10
  • Black: Leela Chess Zero (0.19.0, network 31682)
  •  TC: 2 + 2
  • OB: Nunn Test Suite
  • Hardware: AMD Ryzen Threadripper 1950X 16-Core Processor 3394 MHz.
    • SF was using 16 cores; Leela was using twin 1080Ti's. 
    • Speed:
      • W=34.4 plies; 24,296kN/s 
      • B=13.6 plies; 167kN/s 

Note: this last number was reported by the Fritz GUI, but it is clearly incorrect: on this hardware Lc0 runs closer to 8kN/s and maxes out in simple positions around 45kN/s. This lower node count is typical for neural net programs, which have much larger (and so slower) evaluation functions than the A-B engines we're all used to.

I have added a few notes, and a couple of similar games, but most of the notes are directly from the engines. The numbers following the moves indicate: the evaluation of the engine which played that move/ply depth; the number of seconds spent on that move; the best move according to the other engine (if different from the move played). This gets particulary interesting around move 30, when the engine evaluations deviate wildly.


() - ()
 Round:  Result: