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. Security and hardening: Own the BMC security posture: secure and measured boot, root of trust, attestation (SPDM), authenticated update (PLDM FW Update), rollback protection, attack-surface reduction.