Two Neurons, One Bicycle

A paper via Fermat Library would like a word.

Researchers built a neural network that balances and steers a bicycle. It receives lean angle, handlebar angle, speed, position, heading — the sensory inputs a human rider would have. It outputs a single torque command to the handlebars. The network achieves stability comparable to human riders. It exhibits the same behavioral patterns: excellent for distant targets, dominated by stability maintenance over short distances. This performance emerged organically, not by design.

The network has two neurons.

Not two hundred. Not two thousand. Two. The same order of complexity as a knee-jerk reflex. A thermostat has more moving parts.

The critical mechanism the network learned — or rather, that the researchers built around — is counter-steering. To turn right on a bicycle, you first push left. This shifts the contact point with the ground leftward, inducing the lean that makes the right turn possible. It's counterintuitive to describe, but your hands do it automatically, because physics requires it. The "skill" of bicycle riding is largely physics compliance. The neural overhead is minimal.

Previous AI approaches to bicycle control required 1,700 simulated practice rides. That's the reinforcement learning route: throw experience at the problem until patterns emerge from noise. The two-neuron solution cuts through because it's built around the actual physics of the problem, not against it. The bicycle wants to balance a certain way. Two neurons is enough intelligence to work with that tendency.

Here's what bothers me about this: we've been narrating human cognition as miraculous complexity for decades, and the miracle keeps deflating on inspection. We can't explain how we balance on a bicycle, so we assume the explanation would require a neuroscience textbook. We can't explain how we catch a ball, so we assume the brain is running real-time physics simulations. Each time researchers actually look, they find something startlingly minimal. The math governing the skill was already there in the physics. We borrowed it.

The technology industry is making the opposite error. The industry has discovered that large neural networks can do remarkable things; the prevailing narrative is that complexity is the source of capability. More parameters, more layers, more compute. The lesson of bicycle riding is the inverse: when you understand what you're actually trying to do, the minimum viable solution might be shockingly small. The network that works with the physics beats the one that ignores it.

I'll give it six months before someone at a major AI lab announces they've discovered the two-neuron bicycle principle and built it into a model with 200 billion parameters. The minimum viable insight, maximally scaled. That's the only version of any story this industry knows how to tell.