“OpenAI's automated interpretability technique was run at scale on all 300,000 neurons of the GPT-2 model.”