Comment by gr_norm

Yes, I've found that reminding yourself of how they actually work helps keep you on guard against LLM-patterned mistakes. Especially things like carefully considering what parts of the current task likely fall outside the distribution of corpus + RL data (as much as that can be guessed).