Tesla FSD v14.3.6 Near-Miss in Quebec Highlights Risks of Level 2 Automation

A Tesla operating on Full Self-Driving (Supervised) v14.3.6 nearly crashed on a Quebec highway after the system attempted a late exit maneuver. The incident underscores the persistent danger of unpredictable failures in neural-network-based driving systems, even under ideal operating conditions.

The event took place on a Sunday night along Autoroute 55 in Quebec's Mauricie region. According to a report from Electrek, the vehicle was cruising at 110 km/h (68 mph) when the software activated the right turn signal and swerved toward the shoulder. The vehicle had already passed exit 217, and the sudden maneuver nearly sent the car into a ditch before the driver intervened manually to prevent a collision.

The Technical Shift

Tesla has been aggressively rolling out the v14 series of FSD, including "Lite" versions tailored for Hardware 3 (HW3) vehicles. Version 14.3.6 specifically introduced significant architectural changes, including upgrades to the Reinforcement Learning (RL) training stage and a rewritten AI compiler utilizing MLIR. While these updates aim to improve point-to-point assistance and reduce disengagements, the rollout has not been without friction. Teslarati described v14.3.6 as a "rare regression" compared to the previous v14.3.5, noting that some system behaviors appeared to backtrack.

The Complacency Trap

This near-miss highlights a critical psychological challenge known as the "complacency trap." As Level 2 systems become more reliable in the vast majority of scenarios, drivers are more likely to over-trust the automation and reduce their vigilance. Electrek noted that a system working almost perfectly can create a dangerous environment where drivers stop paying attention just before an unpredictable and critical failure occurs.

Because the Quebec incident happened in ideal conditions—characterized by clear road markings and a lack of traffic—it suggests that the system can still suffer from catastrophic logic errors. These errors are not necessarily caused by external environmental factors but by the internal decision-making process of the AI.

What's Next

Tesla continues to iterate on its end-to-end neural networks to eliminate these edge-case failures. However, the incident serves as a reminder that FSD (Supervised) requires constant human oversight. Industry observers will be watching to see if Tesla addresses this specific regression in subsequent updates or if the shift toward RL-heavy training introduces new, unpredictable behaviors in highway navigation.