A Friday night cruise with Elon Musk has shown the world just how far Tesla’s Full Self-Driving has advanced and gave viewers a preview of real-world problems. To take a line from Rob Zombie, the latest iteration of FSD may be "More Human Than Human," which could be an issue.
"Nothing but net!" proclaimed Musk as he live-streamed to X.com, a video with more than 42 million views. The Tesla CEO referred to neural nets, the technology underpinning Version 12 of FSD. Tesla is removing 300,000-plus lines of code running the current FSD and is now relying on the car's neural networks to drive the car in V12.
Driving through Palo Alto, California, Musk, with Ashok Elluswamy, Tesla’s Director of Autopilot Software, sitting shotgun, picked random places for the Model S to drive to, and it did so smoothly and almost perfectly.
Technology Behind FSD
The test drive started with a construction zone, which was easily handled, but interestingly, at the end of the cones, the car wasn’t jerking over to find a lane immediately. Instead, like a confident human driver, it easily made its way to the lane with no oncoming traffic. We also watched as it effortlessly navigated speed bumps, roundabouts, pedestrians, bicyclists, and, of course - traffic.
For those who have had successful drives with FSD, you may be thinking, what's the big deal? Musk underscored that this version of FSD does not rely on hard-coded programming or a constant internet connection to function. All necessary inference occurs locally, a significant departure from previous versions.
“This is all nets, baby, nothing but net,” Musk exclaimed, emphasizing the vehicle’s dependency on machine learning rather than programmed instructions. The software operates at full frame rate, taking eight cameras at 36 frames per second to decide what to do. No line of code tells the car what to do at an intersection, on the highway or around other vehicles.
The software was impressive to watch, with its countless improvements over the current V11 version of FSD. Musk's vehicle kept a reasonable speed throughout the drive, even though Autopilot was set to a maximum speed of 85 mph.
When Musk reached his destination, the vehicle also automatically pulled over to the side of the road to park itself instead of just stopping in the middle of the road like it does today.
The Real-World, Real Person Challenge
Instead, Tesla has shown the car countless videos of real-world driving so that it will learn what to do in those situations. But that has caused a problem. As the car noticeably did not come to full stops, Musk explained the challenge posed by the fact that less than 0.5% of drivers come to a complete stop at stop signs.
This makes it incredibly difficult to program the vehicle with such sparse data on people obeying the law. This lack of data highlights Tesla's obstacles in training its AI systems to operate in real-world scenarios where human drivers frequently do not adhere to the rules.
Musk's FSD V12 Stream in 10 Minutes
Watch the highlights of Musk's 45-minute video stream below.
Regulatory Hurdles
This appeared in a few other instances where the car behaved like a confident driver, not endangering anyone but technically breaking the rules of the road. FSD suddenly went from the nervous teenager learning to drive to a good driver. But we know how the NHTSA feels about Tesla not following the rules.
In February, the watchdog issued a recall stating that Tesla was acting too human — I mean not following the rules. It cited FSD Beta software:
Traveling or turning through certain intersections during a stale yellow traffic light;
The perceived duration of the vehicle’s static position at certain intersections with a stop sign, particularly when the intersection is clear of any other road users;
Adjusting vehicle speed while traveling through certain variable speed zones, based on detected speed limit signage and/or the vehicle’s speed offset setting that is adjusted by the driver.
Yes, I’m sure we all stop at every yellow light, come to a complete stop at every stop sign, and never go over the speed limit.
Kudos to Tesla and Musk for this real-world demonstration, and if there was any speculation that it was rigged, Musk did have to intervene at an advanced left-hand signal. This small glitch, Musk reassured, could be addressed with further training and data collection. There will be a lot of training announced Tesla’s commitment to invest $4 billion over the next two years in training computing, signaling the company's dedication to perfecting this groundbreaking technology.
Subscribe
Subscribe to our newsletter to stay up to date on the latest Tesla news, upcoming features and software updates.
Tesla signaled its intention to launch its first Robotaxi network in Austin, Texas, with company-owned and operated vehicles, back in January. This network will be Tesla’s first foray into truly autonomous vehicles - ones that aren’t being directly supervised by a driver. Later in February, Tesla’s executive team confirmed that the plans were on track for the launch of the Austin network both on X and during several interviews that they participated in.
At the end of February, we also found out that Tesla has applied for a Supervised Robotaxi license in California, where the network will also launch, but with safety drivers in place.
This is an ambitious plan, but FSD has really come a long way in the past year. FSD V12 was a massive step forward from V11, and V13 has made the experience smoother and safer than ever before. FSD V14 is expected to be another big step up with auto-regressive transformers and audio input.
Still on Track
Now, Elon has tripled down on the fact that Tesla will be launching their first autonomous robotaxi network in Austin - just two months away at this point. It seems that Tesla is fully set to launch their first fully unsupervised self-driving vehicles that will carry paying passengers in June.
Additional Cities
The best part is that Elon also confirmed that they’re targeting robotaxi networks launching in many cities within the United States by the end of this year. However, remember that this will be a Tesla-run network at first - Tesla owners won’t be able to add their vehicles to the Robotaxi fleet right away.
Adding Customer Vehicles
It will take some time before Tesla meets their strict internal safety requirements before it lets customers add their own vehicles to the network. Tesla’s executive team mentioned that they intend to let owners add their own vehicles to the fleet sometime in 2026. This happens to be the same time frame that Tesla plans to launch the Robotaxi across the United States, Mexico, and Canada — something that could only be done with customer-owned cars.
That final step will be bold—and it may come with complications, especially given that autonomous vehicle approval spans municipal, state or provincial, and even federal levels. There are plenty of regulatory hurdles ahead, but this is undoubtedly shaping up to be one of the most exciting times for Tesla.
Launch Event
With the Robotaxi network launching and Unsupervised FSD just around the corner, there’s a lot to get excited about. Tesla is expected to host a launch event at Gigafactory Texas in Austin to mark the debut of its first Robotaxi network. The company previously hinted that referral code users could receive invites—offering a rare chance to score an early ride in a Robotaxi outside the Hollywood studio lot.
Tesla’s engineering mule vehicles—used to test autonomy for future platforms—have resurfaced after an extended absence. The last time we saw them was back in July, when Tesla was gearing up for its initial We, Robot event. Since then, sightings have been scarce.
These vehicles typically signal that Tesla is testing new camera placements or validating FSD ground-truth data. This time, however, they appear to be outfitted exclusively with the updated camera hardware seen on the Cybercab.
New Sightings
The mules have now been spotted in Boston, Massachusetts, and Concord, New Hampshire—their first known appearances this far east. This suggests Tesla is actively collecting data to evaluate and optimize FSD performance in new regions. Thanks to Reddit user Ready_Medium_6693 for catching the one in Concord.
Elon Musk recently mentioned that Tesla plans to expand Robotaxi networks beyond the initial Texas and California launches. So while these sightings aren’t entirely unexpected, the speed of their arrival is. It suggests Tesla is confident in its ability to train FSD for local traffic rules and driving behaviors to the point it no longer needs a driver.
Bumper Cameras
The vehicle that’s been spotted in Boston is the usual Tesla engineering Model 3… except with one unique twist. It includes a front bumper camera. Shout out to @Dylan02939106 for catching the bumper camera in these photos.
The Refreshed Model 3 Mule with a Bumper Camera!
@Dylan02939106 on X
We previously released an opinion piece regarding whether the front bumper camera would be required for Unsupervised FSD. In our eyes, Tesla will require a bumper camera for Unsupervised FSD - if only to improve direct visibility in the bumper area of the vehicle. This is key for low-speed maneuvering in crowded areas like parking lots and garages, as well as for key features like Actually Smart Summon.
It isn’t surprising that Tesla is continuing to evaluate vehicles with a bumper camera - the Cybertruck, Cybercab, and Refreshed Model Y all have them now - and the rest of the lineup, including the flagship Model S and X, as well as the refreshed Model 3 - were supposed to eventually receive them as well.
With the Model S and Model X set for refreshes this year—and the front bumper camera on this Model 3 looking well-integrated rather than a temporary setup—it seems likely that front bumper cameras will soon become standard across Tesla’s entire lineup.
The Robotaxi Mule in Boston
@Dylan02939106 on X
In fact, Tesla’s newest iteration of the FSD computer has a slot for the Bumper Camera on all AI4/HW4 vehicles, so a future retrofit could even be possible.
These are some pretty exciting times - we may see Robotaxi networks actually deploy in time for Tesla’s lofty goal of “sometime in 2026” for cities throughout North America. Once many American cities begin to accept their deployment, it will be easier to seek homologation in Mexico and Canada, enabling deployment throughout the continent.
