What Reinforcement Learning Teaches Us About Social Change

Richard Sutton, the father of reinforcement learning, called LLMs a dead end. His critique is that they don’t learn from experience, predict tokens not outcomes, mimic but can’t act and have no goals.

It’s a technical debate that lands squarely in the social sector because social change requires us to learn in real time by trying, failing, repairing, adapting and persisting. Exactly what Sutton calls the “era of experience.”

Interview: [link]

This is why I’ve argued that wide abstraction is so dangerous in social change. Pattern-matching machines can’t sit in the mess, repair relationships, or build trust and don’t “learn” the way people and communities do.

My recent essay in India Development Review (IDR) Review unpacks this risk, and asks how we keep tools in service to connection.

Originally written for LinkedIn on 29 September 2025. View original →