Tesla ships FSD v14.3 update
Tesla AI pushed Full Self‑Driving (Supervised) v14.3 with a rewritten AI compiler that the company says speeds reaction times by about 20% and improves handling of rare scenarios like emergency vehicles and small animals. The update also teases upcoming features such as pothole avoidance, showing Tesla’s continued incremental approach to on‑road autonomy. Faster reaction and better corner‑case handling are the two immediate technical claims supporting safer supervised driving. (x.com)
A car running driver-assistance software has to turn camera frames into steering and braking commands in fractions of a second, and every extra delay eats into the distance available to react. Tesla says its new Full Self-Driving v14.3 update cuts that reaction time by about 20% with a rewritten artificial-intelligence compiler and runtime. (tesla.com) (electrek.co) A compiler is the layer that turns a trained neural network into instructions a car computer can actually run, like translating a recipe into steps your exact oven can follow. Tesla’s claim is that it rebuilt that translation layer around Multi-Level Intermediate Representation, a compiler framework created in the LLVM ecosystem. (electrek.co) (evxl.co) That matters because Full Self-Driving is still a supervised system, not a robotaxi mode you can ignore. Tesla’s own support pages say the feature requires active driver supervision and does not make the vehicle autonomous. (tesla.com 1) (tesla.com 2) The first public rollout appears to be software version 2026.2.9.6, and reports this week say Tesla started with a small early-access group rather than the whole fleet at once. That is the usual pattern for a safety-sensitive release, because Tesla can watch for regressions before pushing it wider. (teslanorth.com) (notateslaapp.com) Tesla’s release notes say v14.3 also upgrades the reinforcement-learning stage of training, which is the part where the model improves by getting scored on millions of driving decisions. In plain terms, Tesla is saying the system was retrained to handle more awkward edge cases, not just run the old model faster. (evshift.com) (evxl.co) The company’s concrete examples are unusually specific: emergency vehicles, school buses, yellow-light behavior, and small animals. Those are the kinds of rare situations that break driving software, because they show up less often than lane-keeping on a sunny highway. (digitaltrends.com) (teslaoracle.com) There is also a second story hiding under the release notes: Tesla is trying to make future updates ship faster. If the compiler and runtime are easier to retarget, Tesla can swap in bigger or newer neural networks without rewriting as much low-level code each time. (electrek.co) (evxl.co) That is why the teaser features matter even though they are not fully here yet. Reports tied to the rollout say Tesla is already pointing to pothole avoidance and broader behavior upgrades as follow-ons, which suggests v14.3 is as much infrastructure work as visible driving polish. (digitaltrends.com) (blockchain.news) The hard part is that faster reactions do not automatically mean safer driving in every case. A system can respond 20% quicker and still make the wrong choice, which is why the real test for v14.3 is whether those corner cases get handled more consistently over millions of supervised miles. (tesla.com 1) (tesla.com 2) So this update is less “Tesla solved self-driving” and more “Tesla rebuilt part of the engine room.” If Tesla’s compiler rewrite really lets the car think faster and lets engineers iterate faster, v14.3 could end up being remembered less for one feature than for the stack underneath all the next ones. (electrek.co) (evxl.co)
