Tesla FSD
Tesla FSD
One-line summary: Tesla's vision-only driver-assist and robotaxi stack — supervised Level 2 for consumers, "unsupervised" (remote-supervised) in a handful of geofenced cities.
What it is
Tesla's "Full Self-Driving (Supervised)" is the company's end-to-end neural-network driving system running on either HW3 or AI4 (formerly HW4) in-car compute. v13 introduced a transformer-based "world model" with a ~10-second temporal buffer and voxelized 3D scene representation. v14 (rolling out 2026) pushes further on AI4, with a "v14 Lite" planned for HW3 2026-04-20-autoresearch-tesla-fsd.
The same stack powers Tesla's Robotaxi service — unsupervised (no human in front seat, remote monitoring only) in Austin, Dallas, and Houston as of April 18, 2026 2026-04-20-autoresearch-tesla-fsd.
Why it matters to autonomous-driving
Tesla is the most visible counter-bet to the Waymo/sensor-fusion school: it stakes a full consumer-and-robotaxi business on vision-only perception at scale. Whether that bet works or fails is one of the defining experiments of the field.
Key facts
- Hardware generations: HW3 (legacy), AI4 (current), AI5 (delayed to early 2027, ~9× HW4 memory per Musk, January 2026) 2026-04-20-autoresearch-tesla-fsd.
- Sensor stack: Cameras only. No LiDAR, no radar. Unique among major OEMs 2026-04-20-autoresearch-tesla-fsd.
- SAE level: Level 2 (supervised). California's CPUC explicitly states Tesla "is not operating an autonomous vehicle service" 2026-04-20-autoresearch-tesla-fsd.
- Unsupervised fleet (April 2026): Austin ~4–8 Model Ys of ~42 total; Dallas and Houston ~25 sq mi geofences each, launched April 18, 2026 2026-04-20-autoresearch-tesla-fsd.
- Community-tracker disengagement rate: city miles to critical disengagement fell from ~4,109 (v14.1 peak, October 2025) to ~809 after v14.2 2026-04-20-autoresearch-tesla-fsd.
- HW3/AI4 gap (community data): AI4 ~450 miles between critical disengagements vs. HW3 ~120 miles 2026-04-20-autoresearch-tesla-fsd.
Strengths (from our perspective)
- Operates on arbitrary streets rather than an HD-mapped ODD — wider domain than mercedes-drive-pilot.
- Fleet-learning flywheel: large installed base of camera-equipped vehicles feeding training data.
- First-person operator assessment from Karpathy (post-tenure, October 2025). andrej-karpathy led Tesla AI / Autopilot 2017–2022. He still rates Tesla's strategy as the more scalable bet relative to lidar-heavy alternatives, with the caveat that he's not fully independent. andrej-karpathy in 2025-10-17-dwarkesh-patel-andrej-karpathy-summoning-ghosts: "I just think people again are sometimes a little bit too naive about some of the progress and I still think there's a huge amount of work and I think Tesla took in my mind a lot more scalable approach and I think the team is doing extremely well... But I do think Tesla is taking the more scalable strategy and it's going to look a lot more like that." Hold as one weighted operator view, not consensus.
Weaknesses (from our perspective)
- Three concurrent nhtsa investigations into FSD as of April 2026, including an Engineering Analysis (EA26002) covering ~3.2M vehicles 2026-04-20-autoresearch-tesla-fsd.
- Independent safety-report analysis (Phil Koopman) finds Tesla's public numbers methodologically unfit to support its "10× safer than humans" claims 2026-04-20-autoresearch-tesla-fsd — see fsd-safety-data-validity.
- Scale gap vs waymo: ~8 Austin unsupervised cars vs Waymo's ~500,000 paid rides/week 2026-04-20-autoresearch-tesla-fsd.
- Vision-only bet remains unproven for L4 — see vision-only-vs-sensor-fusion and is-vision-only-sufficient-for-l4-autonomy.
- Karpathy's nines-progress data point (October 2025). Even with the "more scalable" assessment above, Karpathy is explicit that the work isn't close to done. andrej-karpathy in 2025-10-17-dwarkesh-patel-andrej-karpathy-summoning-ghosts: "while I was at Tesla for was it five years or so, I think we went through maybe three nines or two nines. I don't know what it is, but like multiple nines of iteration, there's still more nines to go." 2017–2022 advanced FSD by 2–3 nines of reliability; subsequent nines still pending. See demo-to-product-gap-march-of-nines for the framework.
Open questions
- will-nhtsa-ea26002-trigger-fsd-recall
- is-vision-only-sufficient-for-l4-autonomy
- does-remote-supervised-robotaxi-qualify-as-driverless