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. Analyze the economics of model training and research compute, including ROI across different dimensions like training run sizes, chip types, data, etc, and distill into actionable framework for capacity planning and deployment.