Each of those segments runs under its own regulatory and deployment constraints, so delivering Claude to them means standing up compliant infrastructure with cloud partners, landing new models in restricted environments, bringing our application suite to parity with commercial, and clearing frameworks like FedRAMP, repeated for every segment. 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.