V15 "Beyond Humanity"? Don't rush to believe it—regulatory investigations and security data tell a different story.

V15 “Surpasses Human Driving”? Don’t Rush—Look Beyond the Narrative

Musk has more to say: FSD Version 15 will achieve “unmonitored driving while safety surpasses humans.” It sounds like yet another victory declaration for an end-to-end approach. But the plain truth is that the divide between Tesla’s and Waymo’s two different playbooks is now laid out clearly:

• End-to-end (Tesla): One neural network rules them all; in theory it can be used anywhere, but it’s prone to failing when it runs into weird, unusual boundary cases.
• High-definition maps + geofencing (Waymo): It’s rock-solid within the mapped territory, but if you want to expand? It’s slow—and expensive.

Musk is making plenty of noise, but reality on this end isn’t cooperating:

• The NHTSA is still investigating Tesla accidents.
• v14.2 actually took a step back in situations of “should we let it happen or not—whether it should go or not.”
• Leaked benchmark tests show that Tesla can’t outperform some rivals in chaotic urban simulations.
• The v14.3 preview puts the spotlight on improvements to reinforcement learning—to be honest, iterations of this training strategy may be more useful than simply scaling up the model.

My take is this:

• Whether it wins or not isn’t about how flashy the demo video is—it’s about whether it can handle things reliably at 3 a.m., with construction up ahead and nobody watching.
• Version numbers and tweets can’t be directly swapped for stock prices and regulatory licenses. Data assets and verification mechanisms are the real wall.

How different people should look at this

• Competitors (Waymo, Cruise): Cruise already had an accident and is now shrinking; Tesla chose to keep charging forward—whether it succeeds depends on whether V15 can produce solid, concrete proof.
• AI talent: Based on fleet data trends, I’ll give a “rough 70% probability” judgment—if V15 verification succeeds, more talent will move toward Tesla/xAI; if safety isn’t up to standard, the open-source ecosystem will look more appealing instead.
• Fleet operators: Most likely they’ll still choose a hybrid plan of “semi-automation + human backup,” waiting until unmonitored driving has truly proven safety in large-scale scenarios before deciding.

Regulation and Liability: Getting the engineering done doesn’t mean you can hit the road

Technical breakthroughs are only half the story. Safety validation for autonomous driving, who’s responsible when something goes wrong, and how policy will coordinate—those are the biggest variables.

• Views in the market are badly split:
  • Optimists are focused on v14.3’s “inference capability” improvements and believe the turning point is near.
  • Safety researchers say deviations and edge conditions haven’t really been solved at all, and systemic risks have been underestimated.
• From a valuation perspective:
  • After Musk posted the tweet that “V15 is coming,” the stock price didn’t move much, while FSD paid attachment rates actually fell.
  • If V15 safety can be verified by a third party, it may affect the regulator’s attitude. The European Union AI Act revision may, in principle, prefer the “closed-loop training—verification—deployment” path, but in the short term you can’t count on it.

How to understand this wave of narrative

• Musk is very good at setting the tempo, but the claim that “under unmonitored driving, it’s safer than humans” only counts if backed by verifiable data—tweets and version numbers don’t count.
• Smart money doesn’t bet on timing; it bets on data and verification capability:
  • Massive real-world vehicle data is a long-term moat.
  • Reinforcement learning and closed-loop verification determine how far an end-to-end route can go.
• What different people should do:
  • For people building tech / doing research: Right now, the most important thing is to truly integrate reinforcement learning into the training and evaluation closed loop.
  • For enterprises: Until unmonitored driving safety has been validated at large scale, it’s still best to be honest and go with “partial automation + human backup.”
  • For investors: Only now are you rushing in to buy the Robotaxi concept—you’re already late.

In short:

• Whether V15 is worth anything depends on data and verification, not on version numbers.
• How regulation is defined and how responsibility is assigned determines how quickly it can be implemented. In the short term, this matters far more than model parameters.

Importance: High
Category: Technical insights｜AI safety｜Market impact

Conclusion: If you’re chasing Robotaxi concept stocks, getting in now is already late. What really gives you an advantage is capital that makes long-term allocations around data assets and verification capability, as well as frontline builders who put real effort into reinforcement learning and closed-loop evaluation. There’s not much to gain in the short term—patience and verification capability are the deciding factors.