“Mechanistic interpretability research, led by figures like Chris Olah, is unlikely to sufficiently reverse engineer models like GPT-7 in time to solve alignment if AGI arrives soon.”