FSD - HIGH-CONVICTION BET AND THE BULLY EFFECT ?

 

GROK:

Tesla FSD is already operating unsupervised in limited robotaxi form and is on a credible (if still gradual and geography-constrained) path to broader unsupervised capability, though camera-only physics impose real constraints that prevent “every eventuality everywhere” reliability in the near term. 

The extract’s core technical critique of pure vision systems remains largely valid, but real-world progress since those arguments has narrowed some gaps while confirming others.

Current status (as of mid-August 2026)

Tesla already runs fully unsupervised Robotaxi rides (no safety driver or in-vehicle monitor) in geofenced areas of six U.S. cities: the entire Austin metro, Dallas, Houston (Texas), plus Miami, Orlando, and Tampa (Florida). 

The fleet has logged more than 380,000 unsupervised miles with what Tesla describes as “zero notable incidents” (any reported events involved other parties striking stationary vehicles). 

Miles are growing at double-digit rates week-over-week. Early builds of FSD v15 (roughly 40% of planned major improvements) are already running on this fleet.

Consumer FSD remains supervised (v14.x series is the current production software; it is highly capable in clear conditions but still requires driver attention). 

Unsupervised capability for personal HW4/AI4 vehicles is targeted for a phased, geography-by-geography start in Q4 2026 at the earliest, after further validation. HW3 vehicles (millions of cars) lack the memory bandwidth and cannot achieve unsupervised FSD without computer + camera upgrades. 

Scaling remains deliberate and limited—active unsupervised vehicles are still only around 90–100, with broader rollout tied to full v15 and regulatory comfort.

Musk’s timelines have slipped repeatedly for a decade, so treat “widespread by end of 2026” claims with caution; the practical window for meaningful consumer unsupervised and larger robotaxi scale is late 2026 into 2027.

Camera-only limitations vs. progress

The extract ( https://teslasharetrend.blogspot.com/2026/01/perfect-fsd-impossible.html) correctly identifies fundamental constraints of passive cameras: 

they can be degraded or blinded by heavy rain/fog/snow/spray, sun glare, dust, low light, dirty/obstructed lenses, or extreme dynamic range. 

Depth is inferred rather than directly measured, high-speed or heavily occluded scenarios shrink safety margins, and pure vision has no independent active-ranging redundancy. 

These are physics issues, not just “not enough data yet.” 

NHTSA continues probing visibility-related crashes, Tesla has pursued camera-cleaning patents and software mitigations (e.g., forcing more conservative modes in rain), and Waymo executives continue arguing that vision alone hits a safety ceiling short of robust Level 4/5 performance across all conditions.

At the same time, Tesla’s end-to-end neural nets, massive real-world data (supervised fleet well past 13 billion miles), occupancy networks, and iterative releases have produced large practical gains. 

Supervised FSD now handles complex urban driving, many weather conditions, and long-tail behaviors far better than earlier versions. 

Unsupervised robotaxi operation in relatively structured U.S. cities demonstrates that the system can already function without a human in the loop under controlled operational design domains. 

Multi-sensor fusion (cameras + lidar + radar) provides clearer redundancy and is currently superior for the hardest edge cases and adverse weather; 

Tesla’s approach trades some robustness for far lower cost and easier scaling across millions of vehicles. 

Neither path is “solved for every eventuality”—the long tail is real—but vision-only is no longer obviously incapable of limited unsupervised use.

True all-weather, all-road unsupervised Level 4/5 everywhere remains difficult precisely because of the sensor physics the extract highlights. 

Expect continued incremental expansion of operational design domains rather than a sudden “works everywhere” switch.

Chaotic street scenarios (large Indian cities and similar)

Dense, high-entropy environments like Mumbai, Delhi, or Bangalore are substantially harder. 

Unpredictability is the baseline rather than the edge case: 

weak or absent lane markings, multi-directional mixed traffic (cars, two-wheelers, rickshaws, pedestrians, animals, carts), sudden unsignaled maneuvers, variable road widths, monsoon-hidden hazards, and different social driving norms. U.S. suburban/highway or even moderate urban success does not transfer directly.

Tesla has collected data in Indian cities (hiring operators and testing in Mumbai/Delhi areas around 2025 onward). 

Visualization clips of chaotic Indian traffic circulating in 2026 illustrate the challenge—the system’s perception can look overwhelmed by the density and irregularity. Local data is valuable for training the long tail, and some Indian startups are also using chaotic streets as a teaching environment. 

However, full unsupervised operation on open public roads in such cities is a more distant prospect. 

Progress is more realistic first on controlled corridors, highways, geo-fenced zones, or with heavy local adaptation and regulatory approval. Supervised FSD could still deliver useful assistance earlier, but intervention rates and safety validation bars will be higher.

Legal and liability context

The extract’s second section is broadly correct. 

When FSD (especially marketed as “Full Self-Driving”) contributes to a crash, Tesla faces product-liability claims, failure-to-warn arguments, and regulatory scrutiny regardless of the driver’s formal duty to supervise. 

Precedents, jury outcomes, class actions, and NHTSA probes already exist. 

Once true unsupervised operation expands, manufacturer liability generally increases further. This does not prevent deployment, but it raises the evidentiary bar for safety claims and encourages conservative rollouts.

Bottom line: 

Unsupervised FSD is no longer theoretical—Tesla is already doing limited robotaxi versions of it and continues closing the gap with software iteration and data. 

Camera-only systems can reach useful unsupervised capability in favorable domains and will keep improving, but the physical limits of vision mean multi-sensor approaches retain advantages for maximum robustness, especially in weather and extreme chaos. 

Large Indian-style cities represent one of the hardest remaining tests and will lag structured environments by a meaningful margin. 

The trajectory is positive and evidence-based rather than pure hype, yet “unsupervised everywhere, all conditions, no caveats” remains a multi-year challenge rather than an imminent certainty.

Musk’s rigid commitment to vision-only is a high-conviction philosophical and economic bet, not pure “arrogance,” but the lack of sensor redundancy remains a legitimate vulnerability that critics (including competitors) continue to highlight. 

Adding affordable lidar is more feasible in 2026 than it was years ago, yet it collides with Tesla’s core thesis, data strategy, and cost model. The board’s relative inaction is best understood through governance and incentive realities rather than simple neglect.

Tesla’s stated rationale for pure vision

Musk and Tesla repeatedly frame the approach around first principles:

• Humans drive successfully using eyes (cameras) + a biological neural net. A sufficiently advanced digital neural net plus multiple cameras should match or exceed that, without needing “crutches.”
• Multi-sensor fusion introduces sensor contention: when cameras, lidar, and radar disagree, which wins? Tesla argues this ambiguity can create hesitation or unsafe decisions and has cited it as a reason they disabled radar earlier.
• Cost and scalability for a mass-market + robotaxi model. Cameras are already standard on every Tesla at negligible incremental cost. Even with lidar prices having fallen sharply (automotive-grade units now often in the $200–$1,000 range at volume, down dramatically from tens of thousands), multi-unit coverage still adds meaningful BOM, power draw, packaging, aesthetics, and manufacturing complexity across millions of vehicles. Pure vision keeps the hardware cheap enough for the economic case of unsupervised robotaxis on consumer cars.
• Data purity and the “moat.” Tesla’s enormous fleet generates pure camera video at massive scale. End-to-end neural nets are trained on that modality. Introducing lidar would require new fusion architectures, additional labeling/training pipelines, and risk diluting the vision-only advantage they have already invested years into.

They have demonstrated unsupervised robotaxi operation (no safety driver/monitor) in limited geofenced U.S. cities using this stack, with hundreds of thousands of miles claimed and software continuing to iterate (early v15 on the fleet). That progress is real, even if the scale remains modest compared with multi-sensor rivals.

Where the critique lands

Lidar costs have collapsed enough that “affordable incorporation” is no longer science fiction—solid-state and MEMS units are entering mass production at price points that premium and even some volume vehicles can absorb for Level 3/4 features. Active sensing provides direct 3D ranging that works better in darkness, heavy precipitation, glare, dust, or when cameras are partially occluded. Waymo and others treat cameras, lidar, and radar as complementary rather than competing, arguing that pure vision hits a reliability ceiling short of the “nines” needed for widespread unsupervised operation without geographic or weather limits. Independent analyses of early Tesla robotaxi crash rates have sometimes painted a less favorable picture than Tesla’s internal framing, and NHTSA scrutiny has focused on visibility-related failures.

Adding lidar would not erase Tesla’s software/data strengths; it could de-risk the hardest edge cases while still leveraging the neural-net “moat.” The resistance is less about current unit cost than about rewriting the architecture, admitting the pure-vision path needed supplementation, and protecting the narrative that has justified years of hardware simplification (removal of radar and ultrasonics).

Why the board largely sits on its hands

Tesla’s board is not a classic independent check on the CEO. 

Musk is the largest shareholder, chairman, and the central figure whose vision drives both product direction and the stock’s high valuation multiple (heavily tied to autonomy/robotaxi success under his technical approach). 

Directors have historically aligned with or deferred to that vision; challenging the foundational sensor strategy would amount to a public repudiation of the founder’s multi-year bet and could destabilize execution and investor confidence. 

Shareholder pressure in 2026 has focused more on missed timelines, HW3 upgrade liabilities, and scaling speed than on forcing a lidar pivot. Governance critics have long noted limited independence on key issues, but there is no public evidence of a board-level revolt specifically over sensors.

In short, the vision-only stance is coherent within Tesla’s first-principles framing and economic model, and it has produced tangible unsupervised capability in constrained domains. 

It is also rigid, and the missing redundancy is a real technical risk that cheaper lidar makes harder to dismiss on pure cost grounds. 

A pragmatic hybrid (vision-primary + selective lidar) could strengthen the product without abandoning the AI advantages—but that would require Musk to treat lidar as complementary rather than a doomed crutch, which has not been his posture. 

The board’s silence reflects the power structure more than indifference to outcomes.