Get Blake Richards on Dwarkesh Podcast

Blake Richards is suddenly central to the hottest question in AI, how to make foundation models truly agentic, after co-authoring a December 23, 2025 paper that introduces “internal RL,” a method that intervenes in a model’s residual stream to create temporally abstract actions and crack sparse reward tasks. As 2026 begins with every lab racing to build safer, more capable agents, hearing directly from a Google Paradigms of Intelligence researcher who helped develop this approach could shape how practitioners train and evaluate agents today.

He also published June 2025 work on cascading eligibility traces, a neuroscience-grounded recipe for precise credit assignment over seconds to minutes, exactly the kind of mechanism people argue next-gen agents will need. Richards sits at the crossroads of Google’s cutting-edge research and academic neuroscience at McGill and Mila, and he is actively engaging the community as @tyrell_turing, so his playbook is immediately relevant to builders and policymakers right now.

As of January 20, 2026 he has not appeared on the Dwarkesh Podcast before, so this would be a timely first that lets Dwarkesh’s audience interrogate the internal RL paradigm at the moment it is breaking into the mainstream. If you want this conversation to happen when it matters most, please support this nomination today.