Tesla has officially released Full Self-Driving software version 14.3, featuring enhancements in parking prediction, reaction times, and handling of rare driving scenarios. The update includes a rewritten AI compiler for 20% faster reactions and upgrades to reinforcement learning. Early user impressions note smoother highway driving but some persistent navigation issues.
Tesla rolled out FSD V14.3 on April 7, with users like Sawyer Merritt beginning downloads in their Model Y vehicles. Key updates include improved parking location pin prediction displayed on a map with a P icon, more decisive parking maneuvers, and mitigation of unnecessary lane biasing and minor tailgating. The company rewrote its AI compiler and runtime using MLIR, achieving 20% faster reaction times and quicker model iterations, alongside enhanced responses to emergency vehicles, school buses, and right-of-way violators. Handling of small animals improved through focused reinforcement learning on challenging examples, while traffic light management at complex intersections advanced via fleet-sourced data. The neural network vision encoder was upgraded for better low-visibility scenarios and traffic sign recognition, with better management of rare objects in vehicle paths and automatic recovery from temporary degradations. Sawyer Merritt shared first drive impressions on April 8, praising more comfortable highway follow distances, quicker takeoffs from stop signs, and eliminated lane change hesitation compared to V14.2.2.5. He noted the system correctly read detour signs en route to a Supercharger but initially struggled with navigation there, with parking spots selected closer to destinations but no major overall change. Blinker activation timing improved, though speed profiles remained similar. Merritt observed that while changes are not revolutionary yet, V14.3 sets the stage for expanded reasoning, pothole avoidance, and better driver monitoring in future updates.
