Fluent dreaming for language models (AI interpretability method) — AI Alignment Forum