A Song of “Full Self-Driving”

How Hard Is FSD & Where Are the Moats?

• One framing: if FSD is “hard,” Waymo’s years of work and sensor stack give it a big lead; if it’s “easy,” Tesla’s approach is easily copied and offers little moat.
• Pro‑Tesla view: the moat is a “data flywheel” – millions of cars collecting real‑world video, plus custom HW/SW and manufacturing scale.
• Skeptical view: other automakers also have cameras and connectivity; labeling and training are nontrivial; Tesla is late and has no clear sustainable edge.

Data vs. Sensors vs. Compute

• Some argue “the bitter lesson”: scalable compute and massive data dominate clever algorithms; Tesla’s global fleet and companies like Google (via YouTube/Street View) have key “world model” training data.
• Others reply that data volume won’t linearly solve FSD; current techniques may hit a wall regardless of data, and richer multi‑sensor inputs (lidar, radar, HD maps) may matter more than raw video.
• There’s pushback that unlabeled, low‑quality dashcam video is not analogous to the high‑quality text corpora used for LLMs.

Cameras‑Only vs. Lidar/Radar

• One camp claims Tesla’s camera‑only stack is fundamentally limited, citing: repeated hardware refreshes, removed sensors, well‑publicized failures (e.g., “Looney Tunes wall”, crashes into trucks/firetrucks).
• Another camp says this overstates things: FSD has significantly improved in the last few years; newer HW (HW4, upcoming HW5) has better vision and inference; specific failure cases have been mitigated.
• Multiple users point out competing EVs with radar arrays and higher‑resolution cameras that deliver robust ADAS (blind spot, cross‑traffic, intersection warnings) without claiming FSD.
• One detailed comment argues Tesla camera resolution equates to sub‑legal “visual acuity” at many distances, even on newer hardware.

Safety, Reliability & Autonomy Levels

• First‑hand FSD users report it can handle complex highway and urban maneuvers and is very useful—but still regularly: misreads lights, ignores signs/zones, chooses bad lanes, blocks merges, or pulls into traffic.
• Many treat it like advanced cruise control and explicitly say they would not trust it unsupervised.
• Recent crashes prompt debates over whether FSD “randomly” drives off road vs. driver error/override; no consensus.
• Waymo is generally acknowledged as SAE Level 4 (within geofences, with occasional tele‑operator assistance), while Tesla FSD remains Level 2 driver assist.

Musk, Hype, and Trustworthiness

• Several comments highlight lawsuits where Musk’s lawyers argued his FSD promises were mere “puffery” no reasonable investor should rely on; this is cited as a reason not to trust his timelines or claims.
• Others think criticism of Musk and the article’s political framing overshadow real technical progress visible in extensive user videos.
• There’s debate over how much credit Musk deserves versus execution by engineering teams, and whether his insistence on cameras‑only has harmed Tesla’s trajectory.

Competition, Scaling & China

• Some argue Waymo’s tech lead is offset by weak manufacturing and unclear path to millions of vehicles; they may end up a software/licensing provider.
• Chinese automakers are mentioned as already fielding lidar‑equipped FSD‑like systems at low cost and operating at substantial scale, though others think their tech still lags Waymo and relies heavily on human oversight.
• Ride‑hailing drivers and low‑cost human labor are seen as another competitive pressure on robotaxi economics.

Miscellaneous Technical & Factual Points

• Multiple corrections note that lane‑departure, adaptive cruise, and emergency braking often use cameras/radar today; lidar is not yet ubiquitous in mainstream ADAS, contrary to a line in the article.
• Some discussion touches on Google Street View: use of lidar/photogrammetry, prior Wi‑Fi data collection lawsuits, and the likelihood that Google captures extensive environmental metadata.