It’s not very good at explaining stuff though. One example: ask it to explain a subtle joke, and it will keep failing in funny ways.

This is not surprising though, as these kinds of models (LLMs) were specifically optimized for generation, not explanation.