Tesla (Autopilot / Dojo)

Contribution to AI

Betting perception entirely on cameras was a minority position when Tesla took it, and the consequences — good and bad — became the field's clearest running experiment in learned driving. From late 2014 the sensor suite shipped as standard equipment on production cars, which turned a customer fleet into a data-collection instrument: candidate networks could be run silently against what human drivers actually did, and rare events could be mined by shipping a detector and waiting for the returns. No research group with a few dozen instrumented vehicles could sample the world at that rate, and the argument that data volume substitutes for sensor richness has been tested at a scale nobody else could reach. The in-house inference computer of 2019 set an industrial precedent, a car maker designing its own neural network silicon instead of buying Nvidia's, and its throughput was what made processing eight cameras at full frame rate practical; radar came out of new Model 3 and Model Y builds two years later. The multi-task HydraNet design, one shared visual backbone feeding dozens of task-specific heads, became a common answer to running many predictions inside a single latency budget. In 2024 the city-driving stack became one network trained on video, retiring roughly 300,000 lines of hand-written C++. The December 2023 recall of about two million vehicles established the other half of the legacy: a learned driving policy is a component that regulators can declare defective and require patched over the air.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).