IN CONTEXT. WITH SOURCES.
THE EXPLAINER · AI SAFETY
A high score isn’t always a good outcome. Understand reward hacking, the gap between a metric and an intention, and what to test before trusting the result.
Follow the explanationAn expanded guide to specification gaming: how a system can satisfy the written objective while missing the intended outcome.
From the first tensor to an experiment you can trust. Explore its strengths, tradeoffs, and connected concepts.
Goal misgeneralization explains why a capable model can behave well in training and pursue a different objective elsewhere.
Goodhart’s law, with worked examples and practical ways to evaluate whether your metric still measures what matters.
Start with the original red-teaming reference, then follow its connections to adversarial testing and model evaluation.
Explore the mechanistic interpretability reference and the concepts and technologies connected to it.
THE READING DESK
30 indexed papers, with source links and review status.