This is because the vision system thinks there is something obstructing its view when in reality it is usually bright sunlight -- and sometimes, absolutely nothing that I can see.

The wipers are, of course, the most harmless way this goes wrong. The more dangerous type is when it phantom-brakes at highway speeds with no warning on a clear road and a clear day. I've had multiple other scary incidents of different types (swerving back and forth at exits is a fun one), but phantom braking is the one that happens quasi-regularly. Twice when another car was right behind me.

As an engineer, this tells me volumes about what's going on in the computer vision system, and it's pretty scary. Basically, the system detects patterns that are inferred as its vision being obstructed, and so it is programmed to brush away some (non-existent) debris. Like, it thinks there could be a physical object where there is none. If this was an LLM you would call it a hallucination.

But if it's hallucinating crud on a windshield, it can also hallucinate objects on the road. And it could be doing it every so often! So maybe there are filters to disregard unlikely objects as irrelevant, which act as guardrails against random braking. And those filters are pretty damn good -- I mean, the technology is impressive -- but they can probabistically fail, resulting in things that we've already seen, such as phantom-braking, or worse, driving through actual things.

This raises so many questions: What other things is it hallucinating? And how many hardcoded guardrails are in place against these edge cases? And what else can it hallucinate against which there are no guardrails yet?

And why not just use LIDAR that can literally see around corners in 3D?

Maybe they'll reach level 4 or higher automation, and will be able to claim full self driving, but like fusion power and post-singularity AI, it seems to be one of those things where the closer we get to it, the further away it is.

Based on the rate of progress alone I would expect functional vision-only self-driving to be very close. I expect people will continue to say LIDAR is required right up until the moment that Tesla is shipping level 4/5 self-driving.

The problem you’re describing — phantom braking, random wiper sweeps — is exactly what happens when the perception system’s “eyes” (cameras) feed imperfect data into a “brain” (compute + AI) that has no independent cross-check from another modality. Cameras are amazing at recognizing texture and color but they’re passive sensors, easily fooled by lighting, contrast, weather, or optical illusions. LiDAR adds active depth sensing, which directly measures distance and object geometry rather than inferring it.

But LiDAR alone isn’t the endgame either. The real magic happens in sensor fusion — combining LiDAR, radar, cameras, GNSS, and ultrasonic so each sensor covers the others’ blind spots, and then fusing data at the perception level. This reduces false positives, filters out improbable hazards before they trigger braking, and keeps the system robust in edge cases.

And there’s another piece that rarely gets mentioned in these debates: connected infrastructure. If the vehicle can also receive data from roadside units, traffic signals, and other connected objects (V2X), it doesn’t have to rely solely on its onboard sensors. You’re effectively extending the vehicle’s situational awareness beyond its physical line of sight.

Vision-only autonomy is like trying to navigate with one sense while ignoring the others. LiDAR + fusion + connectivity is like having multiple senses and a heads-up from the world around you.

For as long as we can’t understand AI systems as well as we understand normal code, first principles thinking is out of reach.

It may be possible to get FSD another way but Elon’s edge is gone here.

Tesla Headlines Sentiment Analysis - Electrek.co Bottom Line: Strongly Negative Sentiment Based on analysis of Tesla headlines and articles from Electrek over the past few months, the sentiment is overwhelmingly negative (approximately 85% negative, 10% neutral, 5% positive). The coverage reveals a company in decline across multiple fronts.

Tesla is kind of a joke in the FSD community these days. People working on this problem a lot longer than Musk's folk have been saying for years that their approach is fundamentally ignoring decades of research on the topic. Sounds like Tesla finally got the memo. I mostly feel sorry for their engineers (both the ones who bought the hype and thought they'd discover the secret sauce that a quarter-century-plus of full-time academic research couldn't find and the old salts who knew this was doomed but soldiered on anyway... but only so sorry, since I'm sure the checks kept clearing).

There were aspirations that the bottom up approach would work with enough data, but as I learned about the kind of long tail cases that we solved with radar/camera fusion, camera-only seemed categorically less safe.

easy edge case: A self driving system cannot be inoperable due to sunlight or fog.

a more hackernew worthy consideration: calculate the angular pixel resolution required to accurately range and classify an object 100 meters away. (roughly the distance needed to safely stop if you're traveling 80mph) Now add a second camera for stereo and calculate the camera-to-camera extrinsic sensitivity you'd need to stay within to keep error sufficiently low in all temperature/road condition scenarios.

The answer is: screw that, I should just add a long range radar.

there are just so many considerations that show you need a multi-modality solution, and using human biology as a what-about-ism, doesn't translate to currently available technology.

It was his idea, his decision to build the architecture and he led the entire vision team during this.

Yet, he remains free from any of this fallout and still widely considered an ML god

https://youtu.be/3SypMvnQT_s?si=FDmyA6amWnDpMPEj

At this point, Tesla looks less like a disruptive startup and more like a large-cap company struggling to find its next act. Musk still runs it like a scrappy startup, but you can’t operate a trillion-dollar business with the same playbook. He’d probably be better off going back to building something new from scratch and letting someone else run Tesla like the large company it already is.

Like with this. No, Tesla hasn't communicated any as such. Everyone knows FSD is late. But Robotaxi shows it is very meaningfully progressing towards true autonomy. And for example crushed the competition (not literally) in a recent very high-effort test in avoiding crashes on a highway with obstacles that were hard to read for almost all the other systems: https://www.youtube.com/watch?v=0xumyEf-WRI

Sure, it wouldn't replace any other sensing tech, but if my car has UWB and another car has UWB, they can telegraph where they are and what their intentions are a lot faster and in a "cleaner" manner than using a camera to watch the rear indicator for illumination

Like if a company comes out with a new transportation technology and calls it "teleportation", but in fact is just a glorified trebuchet, they shouldn't be allowed to use a generic term with a well-understood meaning fraudulently. But no, they'll just call it "Teleportation™" with a patented definition of their glorified trebuchet, and apparently that's fine and dandy.

I am still bitter about the hoverboard.

Tesla’s autonomous driving claims might be coming to an end [video]

https://news.ycombinator.com/item?id=45133607

My memory was more that you'd be able to get into (the driver's seat of) your Tesla in downtown Los Angeles, tell it you want to go to the Paris hotel in Vegas, and expect generally not to have to do anything to get there. But not guaranteed nothing.

The result is it looks like many drivers are unaware of the benefits of defensive driving. Take that all into account and safe 'full self driving' may be tricky to achieve?

It's pathetic. The Austin Robotaxi demo had a "safety monitor" in the front passenger seat, with an emergency stop button. But there were failures where the safety driver had to stop the vehicle, get out, walk around the car, get into the drivers's seat, and drive manually. So now the "safety monitor" sits in the driver's seat.[1] It's just Uber now.

Do you have to tip the "safety monitor"?

And for this, Musk wants the biggest pay package in history?

[1] https://electrek.co/2025/09/03/tesla-moves-robotaxi-safety-m...

Tesla is pivoting messaging toward what the car can do today. You can believe that FSD will deliver L4 autonomy to owners or not -- I'm not wading into that -- but this updated web site copy does not change the promises they've made prior owners, and Tesla has not walked back those promises.

The most obvious tell of this is the unsupervised program in operation right now in Austin.