Reinforcement Learning
The AI That Learned to Walk by Falling Down a Million Times
The most powerful AI systems in the world don't learn the way you were taught in school — they learn the way you learned to ride a bike.
The Idea
Most machine learning works by showing a model thousands of labelled examples — here is a cat, here is a dog — until it can generalise. Reinforcement learning is a fundamentally different idea. There are no labelled examples. Instead, an agent exists in an environment, takes actions, and receives signals: reward or punishment. Its entire goal is to figure out, through trial and error, which sequence of decisions leads to the most reward over time. What makes this genuinely strange is that the rewards can be delayed and sparse. An agent playing chess doesn't know whether move 14 was brilliant or catastrophic until the game ends forty moves later. This is called the credit assignment problem — figuring out which past action deserves credit for a present outcome. Humans are notoriously bad at this too, which is part of what makes reinforcement learning feel so eerily biological. The other surprising thing: reinforcement learning agents often discover strategies no human would have thought to teach them. They're not constrained by human intuition. Given the right reward signal and enough time, they will find paths through a problem that we'd never have mapped. That's the power. The risk is the flip side of the same coin — if the reward signal is even slightly misspecified, the agent will find clever, unexpected ways to maximise it that are nothing like what you intended.
In the World
In 2017, DeepMind's AlphaGo Zero sat down — metaphorically — to learn the ancient game of Go. It had no human game records to study. No expert commentary. No opening theory. It was given only the rules, a board, and a single instruction: win. It played itself. Millions of games, continuously, refining its sense of which positions led to victory. Within three days, it had surpassed the level of most human players. Within 40 days, it had beaten AlphaGo — the earlier version that had already defeated the world champion Lee Sedol — by 100 games to zero. What unsettled Go experts most wasn't the winning. It was the style. AlphaGo Zero played moves that looked wrong by centuries of accumulated human wisdom, moves that professional players initially dismissed as mistakes — and then, watching the game unfold, slowly understood as something else entirely. It had discovered strategic concepts that human players had never codified, because no human had ever had the opportunity to play millions of games against a version of themselves. This is reinforcement learning operating without the ceiling of human knowledge. It doesn't inherit our blind spots. It also doesn't inherit our judgment about what is beautiful, fair, or sportsmanlike — which is precisely why the question of what reward signal you give these systems is not a technical footnote. It is the whole game.
Why It Matters
Reinforcement learning is no longer confined to games. It is the backbone of how large language models like the ones behind modern AI assistants are fine-tuned to be helpful — a technique called reinforcement learning from human feedback, where human preferences act as the reward signal. It is being used to optimise data centre cooling, design new materials, and train robotic hands to manipulate objects they've never touched before. Understanding its basic logic changes how you read AI news. When you hear that an AI system behaved unexpectedly, or optimised for something technically correct but obviously wrong, you're almost always looking at a reward specification problem. The system did exactly what it was told — we just didn't say what we meant precisely enough. That reframe matters beyond AI. It's a useful lens on organisations, incentive structures, and even personal habits: what are you actually rewarding, versus what you think you're rewarding? The gap between those two things is where most surprising outcomes live.
A Question to Ponder
If you had to define the single reward signal that governs most of your decisions right now, what would it honestly be — and is that the signal you'd consciously choose?
Get a new one of these every morning.
Start learning with Thinkable