> It’s ridiculously hard to find the optimal combination of prompts and examples and fine-tuning, etc. It’s a very large space, there are no easy abstractions to allow you to search through the space, LLMs are unpredictable and fragile, and these experiments are slow and expensive.

Regardless of the actual experiment outcome, I think this is a super valuable insight. "Should we provide legal moves?" section is an excellent case study of this- extremely prudent idea actually degrades model performance, and quite badly. It's like that crocodile game where you're pushing teeth until it clamps onto your hand.