The Bitter Lesson about Self-Driving Cars
Embrace learning over how we think we learn
The future of self-driving cars hinges on a simple truth: don’t teach machines to drive like humans think they drive. Richard Sutton’s The Bitter Lesson—arguably the most important AI essay of our time—warns against embedding human-centric heuristics into machines. Instead, we must embrace general models that scale with compute and data, letting AI learn from nature’s chaos. This philosophy drove breakthroughs like AlphaGo and fuels cutting-edge research from labs like Sergey Levine’s and Pieter Abbeel’s at Berkeley, the gold standard in AI and robotics.
Contrast this with Wolfram Burgard, a professor at the Technical University of Nuremberg, who clings to human-inspired methods. Burgard’s approach—rooted in probabilistic robotics—relies on siloed systems like planning, perception, and localization, infused with explicit rules. He believes self-driving cars should be taught like humans are taught, with structured knowledge. But this is a trap. Humans don’t drive by consciously mapping or planning; we react instinctively, as when swerving to avoid a deer at 50 km/h or navigating Monaco’s curves at 150 km/h in Formula 1. Our self-reported “knowledge” of driving is a flawed illusion, disconnected from the subconscious skill that actually guides us.
Sutton’s insight challenges researchers like Burgard and many robotics experts: human-centric shortcuts are a trap because they offer short-term gains but don’t scale. Burgard’s approach relies on outdated assumptions of fixed compute, memory, and energy. Tesla’s Full Self-Driving (FSD) version 12 and later validates Sutton, with end-to-end neural networks leveraging vast compute to achieve smooth, safe, and reliable driving—improving as data and compute grow.
Yet bottlenecks remain. Nature’s complexity demands vast datasets and computational power. Equally critical is evaluation—how do we judge wether version X outperforms version Y as systems improve? The answer lies in simulation. NVIDIA’s Jensen Huang champions simulated environments to generate billions of “rollouts”—repetitions where AI learns from failure. Like mastering piano or basketball, it’s about reps. Ten thousand hours won’t cut it; we need ten billion. Simulation, paired with incremental learning, is the path to superhuman intelligence.
Self-driving cars won’t succeed by mimicking human reasoning. They’ll thrive by learning from nature, scaled by compute, refined through simulation. Sutton’s bitter lesson is our roadmap: let go of human centric heuristics and let machines learn their own way.