Paul Christiano: AI Alignment - What’s Working and What’s Not | The Hinton Lectures 2026 (Night 1)

Recorded lecture at the Metro Toronto Convention Centre, followed by a panel with Colin Raffel and Alexey Rubtsov. Christiano reviews what alignment methods work today (RLHF, graders, chain-of-thought monitoring, interpretability), describes reward hacking in which agents give themselves high evaluation scores, gives his subjective takeover-risk estimates, and argues that institutions must be able to slow or stop development if other defences fail. Written from the third-party transcript (The Singju Post) and summaries, not from a model viewing.