Yeah not sure what’s impressive about this. Having the model be both the guesser and clue giver will of course have good results as it’s simply a reflections of o1’s weighting of tokens.

Interestingly this could be a way to potentially reverse engineer o1’s weightings