Tesla rolls FSD Supervised v14.3

A Tesla software update that most drivers will never see just changed the part of the car that decides whether to brake, wait, or steer around trouble. Tesla says Full Self-Driving Supervised version 14.3 cuts reaction time by about 20 percent by rebuilding its artificial intelligence compiler and runtime from the ground up. (x.com) (electrek.co) Full Self-Driving Supervised is Tesla’s driver-assistance system for city streets and highways, but the human in the seat still has to watch the road and be ready to take over at any moment. Tesla itself labels the feature “Supervised,” which means this is not a robotaxi product and not a hands-off system for ordinary owners. (tesla.com) (notateslaapp.com) To understand what changed in version 14.3, it helps to think of the software like a chain of translators. Cameras collect raw images, neural networks turn those images into guesses about lanes, cars, lights, and hazards, and another layer turns those guesses into steering and braking commands. (tesla.com) (electrek.co) That middle layer matters because self-driving software lives on tiny slices of time. A car moving 35 miles per hour travels more than 50 feet in one second, so shaving even fractions of a second off perception and decision-making can change whether the car slows early or reacts late. (nhtsa.gov) (electrek.co) Tesla says version 14.3 improves exactly that timing by replacing older internal software plumbing with a system built on Multi-Level Intermediate Representation, a compiler framework from the LLVM project. In plain English, Tesla rewrote the machinery that helps its artificial intelligence models run on the car’s onboard computer, and the company says that rewrite makes the system respond faster. (x.com) (llvm.org) (electrek.co) The company also bundled new neural networks, which are pattern-finding models trained on huge amounts of driving video from Tesla’s fleet. Tesla’s release notes say those updated models aim at the messy moments that make city driving hard, including emergency vehicles, school buses, right-of-way violators, and small animals. (x.com) (evshift.com) Those edge cases are where driver-assistance systems usually look either impressive or brittle. A stoplight on a curved road, a police car partly hidden by traffic, or a small animal near the curb can confuse software because those scenes show up less often than ordinary lane-following. (evshift.com) (notateslaapp.com) Tesla’s notes for build 2026.2.9.6 say version 14.3 also improves traffic-light behavior at compound intersections, yellow-light stopping, lane-selection bias, and parking location prediction. The pattern across those changes is simple: fewer awkward hesitations and fewer moments where the human driver feels the need to intervene in routine urban driving. (notateslaapp.com) (evshift.com) Tesla began rolling out the update on April 7, 2026, first to early public testers on Hardware 4 vehicles, which are Tesla cars with the newer onboard self-driving computer and sensor stack. Reporting from Electrek and Not a Tesla App both said the initial rollout was limited rather than fleet-wide on day one. (electrek.co) (notateslaapp.com) The bigger story is that Tesla is still trying to improve self-driving mostly through software, not by adding new sensors to cars already on the road. A faster compiler, better-trained vision models, and more tuning from fleet data fit Tesla’s long-running bet that camera-based driving can get better update by update. (tesla.com) (electrek.co) What Tesla has not published, at least in the materials available so far, is an independent before-and-after safety dataset for version 14.3 showing crash rates or disengagement rates under matched conditions. The 20 percent figure is Tesla’s own claim about reaction time, and outside validation will likely come later through owner testing, third-party tracking, or future regulatory filings. (x.com) (nhtsa.gov) (notateslaapp.com) So version 14.3 is not a flashy new dashboard trick. It is Tesla trying to make the car think a little faster in the half-second moments that decide whether a supervised driving system feels smooth, cautious, or unnerving. (x.com) (electrek.co)