In the last week, two former Tesla self-driving/Autopilot program leaders have commented on the state of autonomous driving, telling a very different story than their former boss, Elon Musk.
Elon Musk has been notoriously wrong about predicting when Tesla would solve self-driving.
The CEO first announced that “all Tesla vehicles produced since 2016 have all the hardware necessary to achieve full self-driving,” and then claimed, every year from 2019 to 2025, that Tesla would deliver the capability through software updates by the end of each year.
He reiterated the prediction recently, saying Tesla would remove the safety monitor from its robotaxi service in Austin and enable “unsupervised self-driving” in consumer vehicles by the end of 2025.
There’s been a running gag at Tesla about engineers finding out that the company is supposed to deliver something as Musk announces it publicly – leading to a large discrepancy between what Tesla is working on and what Musk claims it will deliver.
Self-driving has been a good example.
While Musk has repeatedly claimed for the last 6 years that Tesla is on the verge of solving autonomy, the people actually working on the technology are not really in agreement. Some of them who left are starting to speak out.
Andrej Karpathy, Tesla’s former head of artificial intelligence, who led the Autopilot and self-driving programs at Tesla from 2017 to 2022, recently warned that autonomous driving is far from a solved problem.
In a new podcast this week, the AI expect again warned that autonomy is not solved:
He said that he would “push pack” on the idea that what we are seeing in the world of autonomy today, with Tesla and Waymo, means it is a solved problem.
Karpathy said:
“[…] I think basically what takes the long amount of time and the way to think about it is that it’s a march of nines and every single nine is a constant amount of work, so every single nine is the same amount of work, so when you get a demo and something works 90% of the time, that’s just the first nine, and then you need the second nine, and third nine, fourth nine, fifth nine, and while I was at Tesla for five years or so, i think we went through maybe three nines or two nines. I don’t know, but like multiple nines of iteration, there’s still more nines to go, and so that’s why these things take so long […]”
Some data support what the engineer is claiming, as the latest FSD Beta software updates that the Tesla team delivered under his leadership did result in a significant reduction in driver intervention, but the progress has been much less evident since:
The first few ‘9s’ deliver a much greater impact, statistically, than the next ones, even though, as Karpathy pointed out, the next ones are just as important and they are just as tricky as the previous 9s.
While he highlights that there’s still a lot of work to be done, Karpathy did say that he belives Tesla’s approach to be more scalable.
He is not the only former Tesla Autopilot program leader to speak out recently.
Sterling Anderson is recognized as the first Autopilot program leader at Tesla in 2015-2016. He now leads global products at GM, which announced this week that it plans to launch level 3 autonomous driving in 2028.
During the event announcing the new autonomous driving timeline at GM, Anderson took a jab at his former employer:
Our customers have driven over 700 million hands-free miles with Super Cruise without a single accident attributed to the technology. I led Autopilot, and you can’t say that for Autopilot. I think this is the long-term play: we build trust with customers by delivering safe products.
He has a point considering Tesla is currently flooded with lawsuits related to accident that involved its Autopilot or ‘Full Self-Driving’ features.
The GM executive favor the more careful approach to autnomous driving.
Electrek’s Take
As I often point out, there’s what Elon says, and there’s what Tesla’s lawyers say.
Elon’s own lawyers say Tesla shareholders shouldn’t listen to him, calling his statements “mere corporate puffery.” That’s an actual quote.
I do believe that Tesla will achieve unsupervised self-driving in consumer vehicles at some point, but I don’t have any evidence that it is close to happening.
As Karpathy said, there are still several 9s to go through before it can be at 99.9999999%, which is needed for level4-5 autonomy, and each of those 9s represents years of work.
I think there’s a clear discrepancy between how Elon talks about self-driving at Tesla and what people who are actually building those systems, like Anderson and Karpathy from 2015 to 2022, are experiencing.
Elon has been lucky to find Ashok, Tesla’s current self-driving leader, who seems to be perfectly willing to endorse his consistently wrong FSD predictions.
It’s not really surprising when you know that Ashok is the one who produced the infamous FSD demo of 2016. As Karpathy pointed out, we should be doubtful of AI demos.
Looking at the prediction markets, people don’t really believe in what Elon is claiming. On Polymarket, people who have been betting on Tesla’s not delivering unsupervised self-driving this year have made a lot of money:
Elon reitereted the goal this week and the “no” answer still gained ground after his claim that Tesla was on track.
FTC: We use income earning auto affiliate links. More.