Within the coming decades, artificial general intelligence (AGI) may surpass human capabilities at a wide range of important tasks.

In this talk Richard outlines a case for expecting that AGIs could learn to pursue goals which are undesirable (i.e. misaligned) from a human perspective. He argues that if AGIs are trained in ways similar to today's most capable models, they could learn to act deceptively to receive higher reward, learn internally-represented goals which generalize beyond their training distributions, and pursue those goals using power-seeking strategies.

Richard is an AI safety researcher, currently working on the Governance team at OpenAI. He previously worked as a research engineer on the AGI safety team at DeepMind. Originally from Auckland, he studied at Oxford and Cambridge; while at the latter he designed and ran the AGI Safety Fundamentals course