Tesla updates supervised FSD

Tesla just changed the part of Full Self-Driving that users never see: the software translator that turns camera input into driving decisions. In version 14.3, Tesla says a rewritten artificial intelligence compiler cuts reaction time by about 20% on Hardware 4 cars getting software build 2026.2.9.6. (electrek.co) A compiler is the layer that takes a model trained on giant computers and makes it run on the chip inside the car. Tesla says it rebuilt that layer and the runtime on Multi-Level Intermediate Representation, a toolchain format used to optimize machine-learning workloads for specific hardware. (electrek.co) Tesla’s system is still not a robot driver you can ignore. Tesla’s own support pages say Full Self-Driving is a supervised driver-assistance feature, requires active driver attention, and does not make the vehicle autonomous. (tesla.com) The basic setup is camera-first driving assistance. Tesla says new vehicles use eight external cameras, process about 2 billion data points in real time, and rely on a vision-based model instead of radar-and-lidar sensor fusion. (tesla.com) That camera-first design is why bad weather and bad lighting matter so much here. In March 2026, the National Highway Traffic Safety Administration upgraded its Tesla Full Self-Driving visibility probe to an Engineering Analysis covering about 3.2 million vehicles after crashes in glare, fog, and dust conditions. (electrek.co) Version 14.3 is aimed straight at those edge cases. The release notes say Tesla upgraded the reinforcement-learning stage used to train the driving network and improved behavior in low-visibility driving, emergency-vehicle response, and pothole avoidance. (notateslaapp.com) Reinforcement learning is the training method where the model gets scored for choices, like a driving coach giving points for smooth merges and taking points away for bad calls. Tesla says that training change improved a wide variety of driving scenarios before the model was packed into this release. (notateslaapp.com) Tesla also says it tightened the driver-monitoring system in the same update. The release notes describe better eye-gaze tracking, better handling for eyewear, and higher accuracy in variable lighting, which means the car is getting stricter about whether the human is still actually supervising. (evshift.com) The rollout itself is still limited and hardware-specific. Multiple reports say version 14.3 first landed on Hardware 4 vehicles, with Tesla starting a gradual owner rollout after employee beta testing in early April 2026. (electrek.co) So this update is not Tesla flipping a switch to unsupervised driving. It is Tesla making a supervised camera-based system react faster, see a little better when conditions get ugly, and hand less of the easy work to the human without letting the human off the hook. (tesla.com)