New paper outlines teaching superhuman concepts from AlphaZero to GMs

https://arxiv.org/abs/2310.16410

Fascinating ideas in this paper. I wonder if this concept could be used to extract known concepts and teach them separately to bots so that you didn’t have the common bot complaint of not playing “human” mistakes. If the bot knew the concepts a 800elo player typically knew, would it make “better” mistakes?

The paper discusses using AlphaZero to teach new concepts via puzzles, and interacting with puzzles, but I am just starting out in my chess journey. I want to learn, and am intimidated by playing humans. If I could play a bot, that played human concepts, including mistakes, I would benefit from that experience.  

Further, the paper discusses a possible exploration into making the experience of learning from AlphaZero interactive! “Nonetheless, it would be interesting to augment this phase with an interactive component: e.g., for each puzzle, humans can actively engage with AZ by playing moves and asking AZ what its response is. This interactive element would allow humans to investigate counterfactual scenarios, allowing for a deeper understanding why AZ did not select their solutions or approaches.”

Just really wild stuff.

This is most interesting!