How close is AI interpretability to being solved?

A researcher who worked on mechanistic interpretability at OpenAI says that, with modern agentic tools and a closer look, the vast majority of neurons in models up to trillions of parameters now look legible. He thinks interpretability could be solved soon. Another commenter says the field has a long way to go and suggests a possible “interpretability phase transition,” citing a paper on changing feature representations in a toy model.