Dashcam 0.17-second clip measures only part of FSD stack

A recent dashcam clip claims Tesla FSD v14 responded to a cut-in vehicle in 0.17 seconds. That figure captures a narrow slice of the pipeline, and social sharing has treated it as proof of full-stack performance. Electrek reported that FSD v14.3 delivered roughly 20% faster reaction time through an MLIR compiler rewrite, with improved handling of small animals and vulnerable road users. Those are documented changes. They do not answer what a single clip actually measured.

What the clip times

The video starts its stopwatch when the object appears in the vision output and stops when the vehicle begins to slow. That interval covers the planning decision and the onset of braking. Perception latency before the object registers in vision, and actuator delay between the brake command and deceleration, sit outside the count. In our reading, a 0.17-second window samples only part of the full perception-to-actuation loop. Tesla has published no median or tail latencies for the complete path, so an isolated timestamp cannot establish fleet-level safety.

What v14.3 actually documents

Tesla's official release notes for v14.3 (2026.2.9.6) credit the MLIR rewrite with about 20% faster reaction time. The same notes describe reinforcement learning that rewards proactive safety around small animals, plus stronger response to emergency vehicles, school buses, and right-of-way violators. Tesla Oracle frames those gains as proactive protection for animals and vulnerable road users. None of these sources spell out internal planner thresholds or cost-weight adjustments. The public record points to faster reactions and training priorities, not a published breakdown of how the planner weights each hazard class.

Why the stack matters

Think Autonomous traces Tesla's historical FSD architecture as a modular stack where perception (HydraNet vision and occupancy networks) feeds a separate planning module. That separation matters when interpreting any clip that clocks only the moment vision labels an object through the first brake response. Policy and regulation need distributions across thousands of miles, not single events amplified on social platforms.

AutonomyEV's opinion

Require release of binned latency data before reaction-time videos shape rules or fleet narratives. Single-frame timestamps tell little about median performance or tail risk across the fleet. Documented v14.3 gains are real, yet they still leave the full loop unmeasured in public data.