← Retour au blog
tech 31 May 2026

It Takes Two Neurons to Ride a Bicycle: A Revolution in AI

In 2004, Matthew Cook demonstrated that a two-neuron network could control a bicycle. This discovery challenges our understanding of AI and machine learning.

Introduction

In 2004, a NIPS conference unveiled a surprising breakthrough in artificial intelligence: it takes only two neurons to ride a bicycle. This discovery by mathematician and computer scientist Matthew Cook not only challenged traditional AI approaches but also questioned our understanding of learning mechanisms.

The Bicycle Challenge

Riding a bicycle is an activity many take for granted, yet it involves a complex set of movements and reactions. For machines, learning to ride straight remains a massive challenge. Before this study, attempts to automate cycling required hundreds of hours of learning or complex algebraic analyses of the bicycle's motion equations.

Two Neurons: A Revolutionary Network

Cook's proposition was simple: a two-neuron network could guide a bicycle in a desired direction. Unlike traditional approaches, this model required neither extensive training hours nor complex algebraic calculations. In fact, it mimicked human behavior by focusing on long-term stability while accepting short-term instabilities.

Implications for Artificial Intelligence

This minimalist approach has profound implications for AI development. It shows that simple systems can accomplish complex tasks, challenging the belief that more complexity necessarily equates to better performance. This paves the way for more efficient and faster applications in various fields, from robotics to industrial automation.

Practical Applications

Systems inspired by this research could transform sectors such as autonomous mobility and robotics. For instance, designing drones or autonomous vehicles with simplified architectures could reduce development costs and improve energy efficiency.

Conclusion

Matthew Cook's study on controlling a bicycle with a simple two-neuron network has upended our understanding of machine learning. It prompts a reconsideration of the complexity of current AI models in favor of more elegant and efficient solutions.

Let's discuss your project in 15 minutes.

AI neural networks machine learning automation Matthew Cook
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call