Tesla updates FSD stack
Tesla AI released FSD Supervised v14.3, claiming roughly 20% faster reactions via an MLIR‑rewritten compiler, stronger reinforcement‑learning training, and improved handling for low‑visibility and small‑animal avoidance. The release also teases upcoming pothole avoidance and driver‑monitoring improvements, showing Tesla continues to iterate its autonomy stack with software releases rather than large hardware shifts. Faster inference and improved edge‑level behavior are the kind of changes that directly affect system‑level latency and safety trade‑offs in deployed autonomy. (x.com)
A self-driving update can look like nothing happened on the outside while a lot changed underneath, and Tesla’s new Full Self-Driving version 14.3 is exactly that kind of release: the steering wheel stays the same, but the software that turns camera video into driving decisions was rebuilt. (tesla.com) Tesla’s system is still an advanced driver-assistance feature, not an autonomous car, and Tesla says the vehicle drives “with your active supervision” using eight exterior cameras and on-board artificial intelligence. (tesla.com 1) (tesla.com 2) The basic job of this software is simple to describe and hard to do: take a flood of camera images, figure out where lanes, signs, people, and cars are, and choose a steering or braking move in time. Tesla says its computer processes 2 billion data points in real time, which is why tiny delays inside the software stack matter. (tesla.com) One piece of that stack is the compiler, which is the layer that turns a trained artificial-intelligence model into instructions the car’s chip can run, like turning sheet music into motions a pianist can actually play. Tesla says version 14.3 rewrote that compiler and runtime “from the ground up” with Multi-Level Intermediate Representation, or MLIR, and claims that cut reaction time by 20%. (electrek.co) (notateslaapp.com) That claim matters because a faster model is not just a benchmark win on a laptop; in a moving car, shorter delay means the same camera frame can turn into a brake, steer, or yield decision sooner. Tesla also says the rewrite improves “model iteration speed,” which means engineers can push new versions through the pipeline faster after training. (electrek.co) (notateslaapp.com) Tesla says it also upgraded the reinforcement learning stage of training, which is the part where a model improves by getting scored on its behavior, more like a flight simulator with points than a hand-written rulebook. In the release notes, Tesla ties that change to broader gains across many driving scenarios rather than to one narrow trick. (notateslaapp.com) Another change is the vision encoder, which is the front end that turns raw camera pixels into a machine-readable picture of the road, like compressing a messy scene into the details that actually matter. Tesla says the new encoder improves rare and low-visibility scenarios, strengthens three-dimensional geometry understanding, and expands traffic-sign understanding. (electrek.co) (notateslaapp.com) The visible driving tweaks sound small until you picture where they happen: Tesla says version 14.3 reduces unnecessary lane biasing, cuts minor tailgating behavior, improves parking spot selection and maneuvering, and handles emergency vehicles, school buses, complex traffic lights, and small animals better. Those are edge cases in the engineering sense, but they are the exact moments drivers notice first. (electrek.co) (notateslaapp.com) Tesla is also using the update to preview what comes next, including pothole avoidance, more “reasoning” improvements, and driver-monitoring sensitivity changes. That fits Tesla’s larger strategy in 2026: keep the hardware mostly fixed, keep collecting fleet data from more than six million vehicles, and move the capability forward with over-the-air software releases. (notateslaapp.com) (tesla.com)
