Tesla’s ‘Supervised Full Self-Driving’ (FSD) in buyer automobiles hasn’t improved all 12 months, primarily based on the perfect accessible knowledge beforehand praised by CEO Elon Musk.
Now Musk factors to having to attend till later this 12 months, however await what?
Tesla’s final main FSD replace, v13, was launched in December 2024.
Musk had beforehand claimed that v13 would allow “a 5 to 6x enhance in miles between disengagements in comparison with v12.5.”
The automaker by no means launched any disengagement knowledge to show any enchancment. Due to this fact, we have now needed to depend on crowdsourced knowledge. There’s a explicit dataset that Musk himself beforehand shared positively, suggesting that the restricted dataset is considerably reflective of what Tesla is seeing in its personal knowledge.
Primarily based on this dataset, v13 on Tesla’s HW4 automobiles was an actual enchancment, however it solely introduced a 2x enchancment in miles between important disengagement – nowhere close to the declare “5 to 6x” enhance.
As we beforehand reported, HW3 automobiles are nonetheless caught on v12, and Musk has admitted that the {hardware} won’t ever help the promised unsupervised self-driving functionality, with no plans to rectify the scenario in sight.
Now, six months after Tesla launched v13, this system has stagnated because the automaker shifted all its efforts to a “robotaxi” pilot program in Austin, Texas.
Tesla has launched a brand new model, v13.2.9 (left), however it has been performing worse than the earlier replace (v13.2.8 – proper) after over 5,000 miles of information:
The most recent knowledge on Tesla FSD v13.2.9 factors to 371 miles between important disengagements.
As we beforehand reported, the robotaxi pilot program in Austin is a transferring of the goalpost for Tesla, which has been promising that each one its buyer automobiles constructed since 2016 would turn into able to unsupervised self-driving with future software program updates.
It operates solely in a geo-fenced space of Austin, the place Tesla is particularly coaching its neural nets to be optimized for the realm. Moreover, it’s utilizing “loads of teleoperation” to help the fleet, one thing that may’t scale to buyer automobiles.
The hope is that Tesla’s optimization and give attention to this pilot challenge in Austin will in the end end in Tesla enhancing FSD in buyer automobiles.
Musk has now commented on this effort:
It’s a brand new model of software program, however will merge to the primary department quickly. We now have a extra superior mannequin in alpha stage that has ~4X the params, however nonetheless requires plenty of sharpening. That’s most likely prepared for deploy in just a few months.
Rapidly after claiming a 4x enhance in parameters, Musk stated that this could be coming “later this 12 months”:
~4.5X enhance in params must be prepared for huge launch later this 12 months. Tremendous frugal use of reminiscence bandwidth, caching precisely what is required & squeezing microseconds out of the whole lot are wanted to keep up the body fee. And the entire system must be retrained.
It’s value noting that Musk’s timelines for FSD releases have traditionally been extraordinarily late.
The higher query is what this long-awaited replace will convey to Tesla homeowners?
Electrek’s Take
The promised and paid-for unsupervised self-driving? No. The “unsupervised” self-driving that Tesla is launching as a part of the pilot program in Austin shouldn’t be transferable to the shopper fleet. It’s geofenced in a small space round Austin, Texas, and it depends on teleoperation, which doesn’t scale to tens of millions of automobiles like Tesla promised.
It’s additionally vital to notice that it’s not the primary time that Musk has promised a major enhance in parameters. The CEO stated that FSD v12.5 on HW4 was a “5x enhance in parameters” and that was fairly disappointing.
FSD v12.5 on HW4 (left) solely introduced a 22% enhance in miles between important disengagement in comparison with v12.3 (proper):
The truth is, the miles between important disengagements plummeted with different v12.5 level updates, and it in the end ended at 184 miles between important disengagements, considerably beneath v12.3:
Due to this fact, it’s exhausting to get too excited a couple of new “~4.5x enhance in parameters” when that’s what occurred the final time Musk known as for it.
Moreover, at the moment, Musk said that HW4 may help an “8x enhance in parameters,” and it was round this time that he started to precise much less confidence in his feedback about HW3.
It took one other 6 months earlier than he lastly admitted that HW3 wouldn’t help unsupervised self-driving, and Tesla principally stopped making any vital updates on the {hardware} since.
Tesla can also be shortly approaching the bounds of HW4 with current updates.
I believe it’s turning into clear that the robotaxi launch in Austin is simply one other distraction from the truth that Tesla can’t ship on its promise of constructing tens of millions of automobiles delivered since 2016 able to “unsupervised self-driving.”
I’m certain that the trouble goes to end in enhancements in FSD in buyer automobiles later this 12 months, however it received’t be to the extent wanted to realize unsupervised self-driving with out teleoperation, which once more shouldn’t be scalable.
If Tesla can get nearer to 1,000 miles between important disengagements, it could be good, however 99% of the worth of FSD lies in degree 4-5 unsupervised self-driving, and we received’t be even near that. And that’s what individuals paid for.
