A Bicycle Needs Just Two Neurons to Stay Upright — and That Should Humble Every AI Lab

A decades-old mathematical result shows that the control problem behind riding a bike is solvable with a two-neuron circuit. The implications for how we think about biological efficiency versus machine learning complexity are quietly radical.

Written by Iris Vale · Bureau Tech · May 30, 2026

The Surprising Claim

It takes two neurons to ride a bicycle. That sentence sounds like a joke, but it is the headline finding of a paper circulated on Fermat's Library — a platform that publishes annotated versions of notable scientific papers — and it has drawn renewed attention via Hacker News.

To be precise: the paper argues that the *control problem* underlying bicycle balance is solvable by a circuit of two neurons. This is a mathematical lower bound, not a claim about neuroanatomy. Your brain does not literally delegate cycling to two cells. But the result is still striking, because it tells us something about the *minimum complexity* required to solve a problem that most humans spend weeks learning.

What 'Two Neurons' Actually Means

In control theory — the branch of mathematics concerned with how systems are steered toward desired states — a 'neuron' in this context refers to a simple computational unit that takes inputs, applies a threshold or weighting function, and produces an output. This is the McCulloch-Pitts model of a neuron, not a biological cell with dendrites and synaptic chemistry.

The bicycle-balancing problem is formally an *inverted pendulum* problem: keep a top-heavy object upright on a moving base by applying corrective forces. It is notoriously tricky to solve in robotics. The paper's contribution is showing that a two-unit feedback controller is sufficient to close the loop.

Why This Is Not Trivial

Minimal-circuit proofs matter for several reasons.

First, they set a floor. If two neurons suffice, then any system using more is either redundant, more robust, or solving a harder version of the problem. Distinguishing between those three possibilities is non-trivial and scientifically productive.

Second, they constrain theories. Neuroscientists building models of motor control have to explain why biological systems use far more neural resources than the minimum. Redundancy? Noise tolerance? Generalization to varied terrain? Each answer implies a different architecture.

Third, and most relevant to applied technology: minimal models are attack surfaces. A system governed by a simple control law is, in principle, easier to perturb. If a robot or autonomous vehicle uses a near-minimal controller for balance, an adversary who understands that controller can craft inputs — unusual road surfaces, sensor spoofing, unexpected payloads — that push the system outside its stability envelope. This is not hypothetical; it is a live concern in autonomous vehicle security research.

What We Do Not Know

The paper, as surfaced, does not appear to make claims about *which* two neurons in the human nervous system correspond to this model, or whether any biological organism has been shown to use a two-neuron circuit for balance. That would be a much stronger and more controversial claim.

It is also not clear from the available summary whether the two-neuron result holds under realistic noise conditions — sensor noise, actuator lag, variable terrain — or only in an idealized mathematical setting. Robustness to noise is where minimal controllers typically fail, and where biological systems typically excel.

Bureau has not independently reviewed the full paper. The claims here are drawn from the Fermat's Library annotation and the Hacker News discussion thread.

The Broader Pattern

Results like this one recur in the history of computational neuroscience. In the 1980s, researchers showed that a small number of neurons could encode surprisingly complex sensory maps. In the 2010s, single-layer networks were shown to solve problems previously thought to require depth. Each time, the lesson is the same: human intuitions about complexity are poorly calibrated.

For technologists building systems that interact with the physical world — robotics, autonomous vehicles, exoskeletons, prosthetics — the practical takeaway is to test minimal models before assuming complexity is necessary. For security researchers, the takeaway is the inverse: do not assume that a simple-looking controller is easy to understand or predict in adversarial conditions.

What to Watch

The Fermat's Library annotation format invites community commentary, and the Hacker News thread is likely to surface critiques of the mathematical assumptions. Watch for challenges to the noise-tolerance of the two-neuron model and for any neuroscientists weighing in on whether the model maps onto known biological circuits. Those responses will determine whether this result is a curiosity or a genuine benchmark.

Key takeaways

The core claim: bicycle balance — a task humans find difficult to learn and impossible to fully articulate — can be captured by a two-neuron control model.
The result is mathematical, not a biological assertion that your brain literally uses only two neurons to ride; it is a lower-bound proof about the minimum representational complexity required.
This kind of minimal-circuit result has historically been used to stress-test assumptions in control theory, robotics, and neural network design.
The gap between 'minimum neurons needed' and 'neurons actually used' is itself a research question — one with direct relevance to understanding redundancy and robustness in biological systems.
For security and trust-and-safety researchers, minimal-model results matter: systems that can be controlled by surprisingly simple rules are also potentially easier to manipulate or spoof.

FAQ

Does this mean the human brain uses only two neurons to ride a bike?

No. The paper makes a mathematical claim about the minimum complexity of a control model, not a biological claim about which specific neurons are active during cycling. The human motor system involves many brain regions and millions of neurons; the two-neuron result sets a theoretical floor, not a description of anatomy.

What is an inverted pendulum, and why does it matter here?

An inverted pendulum is a classic control-theory problem: keep an inherently unstable, top-heavy object balanced by applying corrective forces at its base. A bicycle with a rider is a real-world inverted pendulum. The problem is used as a benchmark in robotics and control engineering because it is simple to state but non-trivial to solve stably.

What is Fermat's Library?

Fermat's Library is a platform that publishes scientific papers with inline annotations, allowing readers to add explanatory notes in the margins. It is designed to make dense academic papers more accessible. It is not a peer-reviewed journal; papers it hosts may have been published elsewhere or may be preprints.

Are there security implications to minimal control models?

Potentially, yes. A system governed by a simple, well-characterized control law may be more vulnerable to adversarial inputs — sensor spoofing, unusual physical perturbations, or edge-case scenarios — than a more complex or redundant system. This is an active area of research in autonomous vehicle and robotics security.

What would make this result more or less significant?

The result becomes more significant if the two-neuron model is shown to be robust under realistic noise and perturbation conditions, or if a biological circuit matching the model is identified. It becomes less significant if the model only works in idealized, noise-free mathematical settings that do not reflect real-world riding conditions.

Citations

It Takes Two Neurons to Ride a Bicycle — Fermat's LibraryThe control problem of bicycle balancing can be modeled with a two-neuron circuit; paper annotated and published via Fermat's Library.
Hacker News discussion thread (Bureau research source)The paper surfaced renewed community attention via Hacker News, where technical commentary on the mathematical assumptions is ongoing.
Inverted Pendulum — ScholarpediaThe inverted pendulum is a standard benchmark problem in control theory and robotics, directly relevant to bicycle-balance modeling.