Can it generate a chess program? If it's feed all the opens source engines. Then some Stockfish clone might pop out.
GPT-3 is interesting as it contains the whole internet. The GPT-3 model contains what? Input data / hyperparameters / weights / biases that's a few petabytes. I save my whole model into json and just load when I want to waste electricity. I haven't actually directly tried GPT-3. GPT-2 when I played with it was horrible.
JohnWoe wrote: ↑Sun Sep 05, 2021 11:48 pm
Can it generate a chess program? If it's feed all the opens source engines. Then some Stockfish clone might pop out.
On a smaller scale, I have recently suggested that instead of directly evaluating a position, the NN should be writing an evaluation function that is well suited to the position (or at least selecting which EF components to use) - link.
Human chess is partly about tactics and strategy, but mostly about memory
JohnWoe wrote: ↑Sun Sep 05, 2021 11:48 pm
Can it generate a chess program? If it's feed all the opens source engines. Then some Stockfish clone might pop out.
..
Dunno when it starts to write chess engines on its own:
Condor Galaxy 1 (CG-1) supercomputer up and running, "beyond GPU", 4 exaFLOP in AI (FP16), supporting up to 600 billion parameters, and extensible up to 100 trillion, CG-2 and CG-3 to CG-9 to be completed in 2024:
