Widely pursued, far from settled
Interpretability research is being actively pursued across the AI industry and academia. Major labs including Anthropic, OpenAI, and Google DeepMind maintain public interpretability research efforts, and independent academic groups contribute a substantial share of the field's published work, often developing techniques and critiques that shape how industry researchers approach the problem. That breadth of participation is a sign of how seriously the underlying question is taken, not a sign that the question is close to fully answered.
It's important to be honest about where the field actually stands. Techniques like circuit analysis, sparse autoencoders, and probing have produced real, useful, and specific findings about how particular small pieces of particular models work. Nobody has anything close to a complete, reliable account of everything happening inside a large modern model. The tools available today are genuinely useful and genuinely partial at the same time, and both halves of that sentence matter.