GPT-4 explains GPT-2 neurons and releases a neuron explanations dataset

AI Impact Summary

AI research and engineering teams can leverage GPT-4-generated explanations to map neuron behavior in GPT-2, enabling faster hypothesis testing and targeted debugging. The dataset provides per-neuron explanations and scores, which can seed interpretability workflows, model auditing, and safety analyses. Because the explanations are imperfect, teams should treat them as a starting point for human review and validate critical claims against ground-truth checks.