LEARNING FROM AN INHUMAN TEACHER

How wise do our mentors need to be?

This article really made me think. The piece is worth reading but one of the (many) good points came from this diagram:

Embedded in this is how a mentor can know more and use that to guide the mentee. Its a game of metering -- you need to know what information to give and when. But it can be hard to know what order to implement teachings in. What is someone close to grasping? What does someone need more layers of learning in order to understand? What is a reasonable mistake to overlook now knowing that it can be fixed in the future?

Alright here comes AI

Inevitably, this has been disrupted with the explosion of AI into everything. Games such as Pokemon, Stratego and Camel Up have had huge bot efforts put into them and now they outperform humans. These bots are at the frontiers of their respective games.

The catch is -- these may not be good mentors. Knowing how to win doesn't mean you can mentor someone.

Stop talking abstractly.

Okay fine. Lets talk about chess.

Chess has experienced the consequences of black box computer domination before most other domains. As long as I've been alive humans have definitively lost the battle against computers.

This expertise means any computer can find the best move to play. But the best humans and the best bots don't play the same. To learn from a bot you need an intermediary that filters your mistakes through a "is this humanly attainable" filter. This would put you on the orange path in the diagram.

Can you do something about it?

This thought process inspired me to make the Marginal Maia Mentor

The basis of this is the Maia model which is designed to play like a human and the Stockfish model which is the most powerful bot ever. Using both of these together I show you your moves if someone marginally stronger would have made a better move.

Specifically the tool filters your moves to make flashcards if a marginally better player (100 Elo stronger) would make and if that move is substantially better than the one you made (10% winning chance increase, like Lichess).

By doing this you will be shown improvements if someone within reach would have done better. You won't be shown the cataclysmic mistake you may have made if someone slightly better than you would have also made that mistake. That's fine. We'll learn it when we're ready.