Tesla updates FSD stack
Tesla AI pushed FSD Supervised v14.3 with faster reaction times, improved reinforcement‑learning training and better handling of low‑visibility and rare scenarios—updates that touch parking, emergency response and edge cases. Elon Musk also confirmed FSD v15 will use a much larger model, suggesting the company sees gains from scaling model size after making progress with smaller models. Those releases indicate Tesla is iterating both software and model‑scale in parallel to improve on‑road behaviour. ( )
Tesla’s driver-assistance software just got a speed upgrade in the part most owners never see: the code that turns camera input into steering and braking decisions. Tesla says Full Self-Driving Supervised v14.3 rewrote its artificial-intelligence compiler and runtime with Multi-Level Intermediate Representation, a software layer from the LLVM toolchain, and that change cut reaction time by 20% on Hardware 4 cars. (notateslaapp.com, electrek.co) A compiler is the translator between a neural-network model and the computer chip running it. If that translator gets faster, the same cameras and the same chip can respond sooner, which is why Tesla tied this update to braking, lane choice, and other split-second moves instead of to a new sensor. (electrek.co, digitaltrends.com) Tesla also changed how the system learns from trial and error, which is called reinforcement learning. The v14.3 release notes say Tesla upgraded that training stage across a wide range of driving situations and then used harder examples from the fleet to improve traffic-light behavior, small-animal handling, and other safety decisions. (notateslaapp.com, evshift.com) The other big change is in the vision encoder, which is the part that compresses raw video into a scene the car can understand. Tesla says that piece now does better in low-visibility conditions, recognizes more traffic signs, and builds stronger three-dimensional geometry for rare scenarios where objects are hanging, leaning, or extending into the car’s path. (notateslaapp.com, teslanorth.com) Those changes show up in small driving behaviors that owners actually notice. Tesla says v14.3 reduces unnecessary lane biasing and minor tailgating, makes parking-spot selection more decisive, improves parking-pin prediction on the map, and responds better to emergency vehicles, school buses, and right-of-way violators. (notateslaapp.com, evshift.com) Tesla also says the software is better at staying calm when part of the system temporarily degrades. The release notes say it can maintain control and recover automatically without driver intervention more often, which is Tesla’s way of saying fewer unnecessary disengagements when conditions get messy. (notateslaapp.com) This is still a supervised system, and Tesla says that plainly in the same notes. The company says the feature can start from a parked position, make turns, change lanes, and park at the destination, but it also says the car is not autonomous and the driver must remain attentive. (notateslaapp.com) The next step is not just polishing the current software but making the model itself much bigger. Elon Musk has said Tesla’s next Full Self-Driving version, v15, will use roughly 10 times more parameters, which is the count of adjustable values inside the model, after Tesla spent the past year squeezing more out of smaller models and better video compression. (notebookcheck.net, teslarati.com) That pairing is the real story in this release. Tesla is tuning the driving stack in two directions at once: faster software plumbing in v14.3 so the current model reacts sooner, and a much larger v15 model so future versions can reason through more edge cases without changing the camera-only hardware already on the car. (notateslaapp.com, electrek.co, notebookcheck.net)
