Just a few days ago, I was pointed to a new scientific paper by Google DeepMind
https://arxiv.org/pdf/2402.04494.pdf
by a Reddit thread – thanks very much!(https://www.reddit.com/r/chess/comments/1alx0t0/google_deepmind_presents_grandmasterlevel_chess/ It’s amazing as it’s from Google DeepMind, it’s about chess, but it’s not about AlphaZero! The Google DeepMind researchers were trying to build a “searchless” chess engine: an engine that doesn’t calculate at all but finds great moves by “understanding” the position through evaluation only. They cite a quote they attributed to Jose Raul Capablanca (although I’ve also heard it attributed to Emanuel Lasker and Alexander Alekhine so it’s obviously a very tempting quote!) which says “I see only one move ahead. But it’s always the correct one!” which sums up the idea very well!
How did they attempt to realise this? They pulled together ten million Lichess games and got Stockfish 16 to analyse all the positions from those games. That gave “Google Searchless” (as we’ll call it from now) a body of knowledge with which to evaluate any position.
The idea of a strong searchless engine is not new. I’ve written a lot about Leela with a search restricted to one node (1 half-move calculation maximum) which achieves exactly the same effect. I’ve used this “hobbling Leela”(!) a lot in my training, If you’ve read “The Silicon Road to Chess Improvement” or regularly watch videos on my Silicon Road YouTube channel, you’ll know that I love training against Leela like this at because
- Leela plays very fast (instantly)
- Leela plays good-looking, solid moves but..,
- Because Leela’s is not calculating it can also make blunders which is quite hopeful for a human player!
Whenever I’m trying to understand a specific position or start with a fresh opening, I like to try things out against Leela at 1 node first. Before the real hardcore engine stuff starts hitting me, it’s useful to try out my human ideas against something a bit less strong! Something that will play a lot of reasonable moves, but maybe fall for some of the little traps that I’ve developed! I recommend it to everyone! Not many people listen 😉 but there was a very strong player recently who said he trained a lot like this and found it very useful, so fingers crossed!
For my book “The Silicon Road to Chess Improvement” I ran a 101-game and published quite a few games from that match in there. I made 78% against Leela but the games that Leela was winning were incredible! It was navigating Open Sicilians full of tactics without calculation! I was thinking “but come on, there must be a blunder soon!” I estimated Leela’s strength around 2450-2550 albeit a rather unreliable one!
Why do I find this valuable and worthwhile? Well, it’s very interesting to try to learn from engines; obviously, though we can’t calculate like engines, there’s no reason why we cannot evaluate like them (that’s just a question of knowledge and insight and applying it). If an engine playing without calculation – or just a minimal amount – can reach a level of 2900-3000, that gives you hope as a human player that with a better quality of evaluation, you might be able to reach that level too!
So how strong is Google Searchless? The paper states that Google Searchless had played against both bots and humans at Lichess. Against bots it got to an ELO of 2299 and against humans at blitz, it got a Lichess rating of 2895 over 174 games…
As always with these papers, there’s lots of scientific detail – a lot of which goes right above my head! – but there’s not quite enough on the chess for me to really get a grip of what that means! The rating you get when playing humans depends on a few factors. For example, were the games with increments or not? What sort of players were they? Were they strong players that you drew with, or a series of weaker players and you beat a lot of them? Looking at the 7 games presented at the end of the paper, my feeling was that the human opposition was not that strong… or maybe having a bad day 😉 Lichess (blitz) ratings are also not the same as over-the-board ELO ratings. I was once told that someone of my over-the-board strength (between 2600-2700) should have a Lichess blitz ELO of 3000 (never managed by the way!) so maybe 2895 equates to around 2500/2500 over-the-board?
In this video – https://youtu.be/zMHf3lmP0WE – I take a look at a couple of “Google Searchless” games. In general, I think the style of play – active, aggressive – is quite similar to that of Leela at 1-node and I would guess that the strength is in the ballpark of the Leela I faced in that 101-game match. Of course there have been a lot of developments on Leela since then, and based on the games I’m playing now with Leela, I have the feeling that a fresh match would be much closer. Leela’s biggest fan – grizzled veteran TCEC chatter mrbdzz – suggested I play a fresh match and stream the whole match online! 😊 We’ll see 😉!
In any case, very happy and excited to see Google DeepMind pick this up and I hope that there are more developments to come! Of course, the best thing EVER would be a match between Leela 1-node and Google Searchless… well an engine fan can hope! 😊
Hey!! Thanks Matthew for this saturday morning coffee chess brake. I always thought it was the words of Smyslov (Only 1 move ahead), while Capablanca’s hand put the pieces on good squares.
And yes, mrbdzz is right, you should do a match against Leela and stream it live. But before that, you should ask for specifics openings to your subsribers.
Have a great weekend Matthew.
Ah indeed, that might be a fun idea to ask for some opening suggestions! (No Bongcloud or Grob though! ;)) Best Wishes, Matthew
Hi Matthew! Thanks for the post. I haven’t dug into the paper yet, but just from reading your post it seems to me that their feat is not particularly impressive – after all, as you yourself say, 1 node Leela achieves comparable results.
OTOH, much depends on the details. Leela is based on AlphaZero’s neural net, which encodes previous 50 positions in the game (they did this to avoid having to explicitly code the 50 move rule or the threefold repetition rule). What’s to stop a neural net from implicitly encoding search?
Ythe shallow benchmarking is typical for Google’s chess papers and predictably disappointing… Hopefully we will get an open-source implementation soon and we can learn more from it 🙂
Hey, thanks for the comment. Indeed, the Leela devs posted a tweet claiming that they had something stronger already using the same transformer techniques. I’m not technically competent enough to judge that, but my impression from looking at the games is that Leela would hold its own well with Google Searchless. I am very keen on searchless engines (I think they give excellent insight into how human players might get much stronger by improving their evaluation) so hoping for some more chess details to emerge and with a bit of luck this is the start of something, rather than just a single paper! Best Wishes, Matthew
I like playing against the grob but i’ll never play it unless a mouse slip 1.g4 instead of 1.g3 which happened once but yark !!!
No i would be more into a Caro-Kann Kortchnoi / Bronstein-Larsen, othherwise a Rossolimo.
🙂 🙂