Reinforcement learning via backpropagation was a fundamental shift from supervised learning because it eliminated the need for paired labeled data, enabled exploration and mistakes (the only path to genuine intelligence), and mirrored how humans and animals learn through trial-and-error.