Stephen Downes

This is something we tried to argue in the 2000s with connectivism, though the response was often, "there's no similarity between computer and human learning," or even, "computers can't learn, period." Here's where we are in 2025: "Reinforcement learning is at its core the art of learning through experience, trial and error, feedback, and gradual refinement. Whether it's an AI mastering a complex task or a child learning to ride a bike, the principles remain the same." This article explains the concept. And I want to underline here, this is a learning theory, and it's the sort of learning theory that makes the old stand-bys of behaviourism, cognitivism and constructivism seem quaint and uninformed. What's not in the article are the mathematical foundations, physical implementations, applications and deployment. But they're all there too.