This research continues many of the directions our team worked on prior to Anthropic, including: GPT-3, Circuit-Based Interpretability, Multimodal Neurons, Scaling Laws, AI & Compute, Concrete Problems in AI Safety, and Learning from Human Preferences. Identify patterns and trends across cases, and close the loop between enforcement outcomes and upstream improvements by feeding findings back to policy, detection and research teams to improve policies, classifiers and model behavior.