TalkRL: The Reinforcement Learning Podcast David Abel on the Science of Agency @ RLDM 2025
Sep 8, 2025
David Abel, a Senior Research Scientist at DeepMind specializing in agency, dives into the fascinating intersection of reinforcement learning and philosophy. He discusses how clear definitions can drive advancements in the field and the relationship between plasticity and empowerment in intelligent agents. The conversation also covers the significance of mutual and directed information in communication theory and challenges traditional views on continual reinforcement learning, advocating for a more dynamic understanding of agency in AI.
AI Snips
Chapters
Books
Transcript
Episode notes
Philosophy Guides Precise Definitions
- David Abel treats 'science of agency' as blending philosophy with formal definitions to make progress.
- He argues iterative, precise definitions can open new research trailheads and insights.
Plasticity And Empowerment Are Duals
- Plasticity measures how much observations can influence an agent's future actions over time.
- Empowerment measures how much the agent's actions can influence future observations, making them dual concepts.
Swap Perspectives To Reveal Symmetry
- Viewing agent and environment symmetrically reveals environment 'plasticity' and 'empowerment.'
- Swapping perspectives shows plasticity of one equals empowerment of the other.
