Today’s “FSD” has its limitations and requires supervision, but your description of it is not anything like my experience even on a HW3 vehicle. In fact, in many years of using Autopilot and various “FSD Beta” and “FSD (Supervised)” versions for several tens of thousands of miles I’ve literally never seen it “slam on the brakes suddenly for shadows” or “veer into the wrong lane”. I’m not a cult member and my next car won’t be a Tesla because I cannot support Musk after the horrible things he has done these last 2-3 years, but “FSD” is phenomenal when used appropriately and with the right expectations about what it is and what it isn’t. And it has improved a ton over the years, too.
The end-to-end solution was a real game changer, and while the previous solution was still useful and impressive in its own right, moving to the new stack was a night and day difference. With V13 finally taking advantage of HW4, and all the work they’ve been doing since then (plus upcoming HW5 introduction), it’s totally within the realm of possibility that they achieve viable L4 autonomy beyond this kind of small scale demo (and I hope some form of L3 maybe on HW4 before long for customer vehicles).
They collect video, not images, along with other sensor and control data.
It’s not a sunk cost fallacy, it’s a technical strategy that is very logical and showing compelling results (though as of yet unproven for achieving robust L3 or L4 autonomy, demos notwithstanding).
Driving over a double yellow is expected and legal in normal driving, such as when making a left turn or going around an obstruction.
In this example it looks like it oscillated between two different routing choices (turning left and going straight), and by the time it decided the correct route was to go straight, it found itself misaligned with the lane it should have been in. Instead of moving all the way back to the right, it kind of “cheats” its way into the upcoming left turn lane. This isn’t something it should do in this situation, but it’s likely emulating human behavior seen in situations like this which appeared in its training data, where people cut across the center line(s) ahead of the turn lane forming when they can see that it is clear.
The thing a lot of people get wrong is that they think the most valuable data for Tesla to collect are the mistakes or interventions. Really what they need the most is a lot of examples of drivers doing good job of handling complex situations. One challenge, though, is separating out those good examples from the less good or bad ones, as human drivers are notoriously bad at, well, driving.
They used it on test mules to create labeled training data for their older monocular depth / Occupancy Network models (which they still use as part of the supervisory policy enforcement and active safety layer, alongside the end-to-end model).
They’ve never had LiDAR in their cars and it would not ever have been practical for them to have done so. Nobody has any mid or long range LiDAR in vehicles as the scale that Tesla sells.
In a perfect world, AV software wouldn’t be necessary. We don’t live in a perfect world. So we need defense-in-depth, covering prevention, mitigation, and remediation.
The end-to-end solution was a real game changer, and while the previous solution was still useful and impressive in its own right, moving to the new stack was a night and day difference. With V13 finally taking advantage of HW4, and all the work they’ve been doing since then (plus upcoming HW5 introduction), it’s totally within the realm of possibility that they achieve viable L4 autonomy beyond this kind of small scale demo (and I hope some form of L3 maybe on HW4 before long for customer vehicles).