Tesla FSD Supervised v14.3

A Tesla on Full Self-Driving Supervised is still a driver-assistance car, not a robotaxi: Tesla says the system can drive “almost anywhere” only with your active supervision, and the owner manuals say you must stay attentive and be ready to take over at all times. (tesla.com 1) (tesla.com 2) The way it works is simpler than the name sounds. Tesla says the car mainly uses eight external cameras as its eyes, then runs artificial-intelligence models on the in-car computer to turn those video feeds into steering, braking, lane changes, and turns. (tesla.com 1) (tesla.com 2) Those artificial-intelligence models are called neural networks, which are software systems trained on huge piles of examples the way a person gets better after seeing the same kind of situation thousands of times. Tesla says it trains Full Self-Driving Supervised on billions of miles of anonymous real-world driving data. (tesla.com) The bottleneck is not just seeing the road. A car also has to turn those camera images into a decision fast enough that a child stepping off a curb or a car braking ahead does not arrive one beat too late. (tesla.com 1) (tesla.com 2) That is why Tesla’s new v14.3 update is mostly an under-the-hood release. Tesla says it rewrote the artificial-intelligence compiler and runtime from the ground up using Multi-Level Intermediate Representation, a software framework for turning trained models into code that runs more efficiently on the car’s hardware. (electrek.co) (notateslaapp.com) Tesla says that rewrite cuts reaction time by about 20%, which is the kind of gain carmakers chase because the same model can feel more natural if it starts braking, yielding, or steering a fraction of a second sooner. Tesla also says the rewrite speeds up model iteration, which means its engineers can test and ship new driving models faster. (electrek.co) (notateslaapp.com) The other big change is training. Tesla says it upgraded the reinforcement-learning stage, which is a method where a model improves by being rewarded for better choices, and it also upgraded the vision encoder, the part that turns raw camera frames into a machine-readable map of the scene. (electrek.co) (notateslaapp.com) Tesla says those model changes are aimed at rare and low-visibility scenarios, stronger three-dimensional geometry understanding, and broader traffic-sign understanding. In plain English, that means better guesses when rain, glare, darkness, odd road layouts, or unusual objects make the scene harder to read. (electrek.co) (notateslaapp.com) The release notes also get very specific about behavior. Tesla says v14.3 improves emergency-vehicle handling, small-animal handling, traffic-light and traffic-sign understanding, parking-spot selection, maneuvering, and recovery from awkward situations, while also reducing unnecessary lane biasing and minor tailgating. (electrek.co) (notateslaapp.com) The rollout is limited, at least at first. Reports on April 7 and April 8 said v14.3 was going to Hardware 4 vehicles in the United States, including Model S, Model 3, Model X, Model Y, and Cybertruck, with early-access testers seeing it before any wider release. (electrek.co) (notateslaapp.com) (teslarati.com) So the story in v14.3 is not that Teslas suddenly became self-driving on April 7, 2026. The story is that Tesla is trying to make a supervised camera-based driving system react faster, understand messier edge cases better, and behave less awkwardly in the last 20 feet of a trip, while still telling drivers to keep their hands ready and their eyes on the road. (tesla.com) (tesla.com) (electrek.co)