RT @OpenAI: We applied GPT-4 to interpretability — automatically proposing explanations for GPT-2's 300k neurons — and found neurons respon…