What happened to the LC0 project?

[Andreas Geiger]

Dear folks,

where is the enthusiasm gone this project had in the beginning? LC0 reached some time the top top level, and the supporters dissapeared in big numbers. The respectfull ChessDBCN supported the project for some time with respectfull CPU-time, and seems to be gone now. Seems to me LC0 would become to strong and would be reaching top top level if we just would let it go. Big thx to ChessDBCN for what was done so far! Your oppinion is appreciatet. 

[Dave Whipp]

To my mind, it achieved its original scientific objective (reproducing the A0 result); made a couple of innovations to the basic architecture (e.g. MLH); and then tried to keep up with Stockfish through optimization and tuning. But in the fight of "best self-learning engine" Vs "best engine, unlimited", the latter wins. The fields of RL and deep learning have continued to innovate; but LC0 has stuck with the basically the same architecture. Stockfish subsumed the essentials; LC0 didn't innovate (of, if they did, then they didn't communicate); so people drifted away.

I think LC0 achieved success; spawned multiple projects that took advantage of that success; but hasn't found a second act.

On Sat, May 22, 2021 at 5:47 AM Andreas Geiger <Isidor.K...@gmx.de> wrote:

Dear folks,

where is the enthusiasm gone this project had in the beginning? LC0 reached some time the top top level, and the supporters dissapeared in big numbers. The respectfull ChessDBCN supported the project for some time with respectfull CPU-time, and seems to be gone now. Seems to me LC0 would become to strong and would be reaching top top level if we just would let it go. Big thx to ChessDBCN for what was done so far! Your oppinion is appreciatet. 

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/c1d961c0-4368-443b-9b14-fbd28157a177n%40googlegroups.com.

[Ivan Ivec]

Leela has a very good evaluation, that's for sure.
But you need to make it faster.

In my opinion, Leela (or any other engine) needs two networks: one for evaluation, and one for search decisions.

[Ivan Ivec]

After more thought it seems to me that current main run net is impressive, but too big for the current hardware.

I expect that Run 2 and/or Run 3 will be successful after more training games.