Long story short: model is 3000 times smaller than ChatGPT 4, learned from Stockfish games, doesn't search, doesn't play itself. Basically a LiChess Tools user :D . The purpose is not to beat SF, but to learn principles and maybe generate them as algorithms.
www.youtube.com/watch?v=uz83G-2ny8QNot a large language machine. It uses AO NN, and SF as oracle. But it does compete against one that is using language model. That is my understanding, from browsing the deepmind paper.
My take home is that it was saying that whatever chess SF16 is the top engine about, the same NN that was able to do RL chess, is also able to compress or embed pretty much the same as SF16, at least enough to be GM level as assumed engine ELO always do.
There even had a control group play humans to tie their engine vs engine world back to human ELOs.
I did not watch the video at all. I work from the paper. Is what I wrote needing some fixing?
Also, it did not learn from SF games (that is SF that "learns" from Leela's "data"). It learn directly from SF score as it oracle trainer. So A"0" becoming a transformer of SF16... yes, that is a training technique that was popular first in LLM stuff. making a "smaller" NN from a bigger NN. but what it says is that the current size of A"0" NN, and its basic input vector formalism (was basically FEN with some condiment of PGN history, but not enough to stop saying it was a FEN), is enough to capture the chess that SF16 is hard-wired to play (or that its dev. captured, or that its SF-Dev class of engine captured over the engine pool ages).
qui dit mieux? (also the narrator in the video, is trying to sound like a robot).
Edit: really, SF games. Well I guess it is fair that they would use its world of chess, if trying to embed it. But, here is a thought, they might also use positions that were never fed to SF, and espouse its phenotype there too.. it could then probably improve its games with or against SF16 itself...
The youtube uloader links that are its own, do no give access to papers.. only the archive link about the paper I looked through. (It as a summary table of its engine ELO and important characteristics such as input form (PGN versus FEN). Since LLM are about sequence of strings, they are good at "historical" short and long cross-correlation grabbing. But chess is inherently position determinisic, it is more of a spatial problem, yet, if all the PGN also include the information of the position at the start.. (that is where I don't know how the chat-gpt3 in there is faring, not very good, but what did it really chew on?).
So, when is it making its way to LT? :). You might want to use lc0 instead for a real second opinion to complete SF point of view given by lichess.
You can spot our own Matthew Sadler in the video there : https: [slash slash] www [dot] youtube [dot] com/@SiliconRoadChess/videos (I hate the links to embedding mechanism in Lichess)
The problem with large language models is that they are ... well... large :) Can't bundle them in the extension. It would be interesting to have people get the extension and have to wait for 2 GB to download.
I have been thinking of some server options for LT, like having one server online and providing some functionality there, but that involves a lot of time and effort and some resources that I don't want to invest right now.
Anyway, any discussion in this direction is premature at this point.
