Comment by zamalek

I wonder how a transformer (even an existing LLM architecture) would do if it was trained purely on chess moves - no language at all. The limited vocabulary would also be fantastic for training time, as the network would be inherently smaller.