Technical AI Safety Fellowship

A reading group on interpreting, evaluating and aligning advanced AI systems, from RLHF and reward misspecification to mechanistic interpretability, control and chain-of-thought faithfulness.

What you needNo background required; backgrounds in CS, engineering, and math encouraged.

Express interest

An introductory program covering the most important questions at the heart of AI safety and alignment research.

Over eight weeks of 90-minute discussions, fellows explore the technical challenges in understanding, supervising and shaping the goals of increasingly powerful AI systems.

Curriculum

  1. Reinforcement learning from human feedback (RLHF)
  2. Reward misspecification and goal misgeneralization
  3. Mechanistic interpretability and model evaluations
  4. Unlearning, red-teaming and scalable oversight