Driver remains legally + operationally responsible. Eye-tracking enforces attention. "Full Self-Driving" branding is misleading; "Supervised" is doing real work.
End-to-end neural net
v12+ replaced modular pipeline with one big network. Smoother edge cases, less interpretable failures.
Where FSD shines vs still fails
✓ Shines
Highway driving — smooth lane keeping
Stop-and-go traffic (the worst commute beat)
Familiar routes w/ clear markings
Parking-lot navigation (v13 jump)
Highway-to-exit transitions
✗ Still fails (vigilance needed)
Construction zones w/ cone detours
Stopped emergency vehicles (NHTSA concern)
Heavy rain / snow / fog
Low-sun glare
Unusual traffic control (cops, flaggers)
Phantom braking (rarer than 2022, not gone)
The Grok Integration
Grok lives in the car as an LLM assistant — Q&A, trip planning, vehicle control via natural language ("warm the cabin to 72°F"). Not part of the FSD driving stack. Two separate systems.
Vision-Only Bet
No LiDAR. No radar (mostly). Tesla's argument: humans drive with vision. Counter: LiDAR provides geometry no camera matches in bad weather. Bet still in question.
Tesla vs Waymo vs Mercedes
What Grok actually does in-car
Waymo (Robotaxi) Level 4 in defined service areas. No driver.
Tesla FSD Level 2. Anywhere validated. Driver required.
Mercedes Drive Pilot Level 3 in highway corridors only.
Trip planning "Italian restaurant open at 9pm w/ parking"
Vehicle control via NL "Schedule service Tuesday morning"
Hands-free Q&A while driving General knowledge + context