Dave Lee invested in Tesla when it was effectively only a few dollars a share and now has the YouTube channel “Dave Lee on Investing.” James Douma is cofounder of Nitobi Software and an expert on machine learning. They teamed up to talk about their thoughts on Tesla, Tesla Full Self Driving (FSD), and 2022 in the video below.
It’s a bit technical, but some key points covered Elon Musk’s comments on AI, the possibility of FSD reaching Level 5 by the end of 2022, Tesla Dojo computer chip progress, and the outlook for 2022.
Also, Dave has a big heart for those in need and is raising funds to help orphans in Afghanistan. The funds go to a nonprofit organization based in California that has been working inside Afghanistan. You can learn more and donate here.
Elon Musk’s Involvement In AI
Elon Musk recently shared thoughts on AI and other topics in a third podcast with Lex Fridman, and Dave noted that his impression is that Elon is a CEO who is doing more work than a typical CEO would do.
“He’s pretty familiar with everything that’s going on — the major issues and how they’re trying to solve them.”
In the interview with Lex Fridman, Elon spoke about Tesla’s Autopilot, Tesla self-driving, neural networks, and when Tesla could solve self-driving. In his podcast with Elon, Lex spoke of the natural skepticism he had when he first drove a Tesla with the initial system based on Mobileye tech.
“At first when I got in, I thought, ‘there’s no way this car could maintain, like, stay in the lane and create a comfortable experience.’ So my intuition initially was that the lane-keeping problem is way too difficult to solve.”
Elon explained that they had to wrap a lot of code around the Mobileye thing and that it didn’t just work by itself. The deep, detailed conversation is what Dave Lee was referring to when he pointed out how familiar Elon was with these issues and trying to solve them. Douma agreed and added that Steve Jobs knew a lot about his products but not on the level of technical detail that Elon does with Tesla’s FSD and AI.
Douma pointed out that machine learning is incredibly specialized and has several subdivisions of people who are specialized. His general sense is that Elon has a much more holistic understanding, like a deep understanding of all of the specifics that go into the metallurgy of the raptor engine.
FSD Level 4 By The End Of 2022?
Dave asked Douma what his take was on Elon’s prediction of when Tesla would achieve Level 4 autonomy. In the podcast with Lex Fridman, Elon Musk answered a question about when Tesla could achieve Level 4 autonomy. It could be this year, or by the end of this year, he said. Elon pointed out that the rate of disengagement has been dropping rapidly in FSD Beta, and is likely to achieve a significantly lower rate of accidents than an average human next year.
After achieving this, Tesla would still have to prove its capabilities to regulators. This will most likely take some time. Dave asked Douma for his thoughts and whether or not he thought it was realistic, and what exactly Level 4 is?
“Broadly speaking, I think he’s becoming more specific, you know, less vague about his assertions, and less over the top optimistic about how soon things are going to happen. What I heard was a fairly specific narrow assertion that the intervention rate for serious, serious interventions; that on the current improvement curve; that in the next year; that it looks like it will cross underneath the human accident rate.
“Those are kind of two really gross ways of sort of trying to compare FSD to where it needs to be. If where FSD needs to be is a benchmark in terms of how humans drive — like, once it’s twice as good as people or four times or whatever; like, if that’s our benchmark for when we move to the next phase of letting the car drive without or not requiring people to pay as close attention or turning it loose on a city as a robotaxi — if that’s your benchmark, you need some human metric and then you need some approximately equivalent FSD metric that you track over time so you can sort of make a prediction about when they cross. And one way you could look at this — and this is what I’m hearing, but I’m extrapolating a bit from what he’s saying — is that they have this sort of synthetic metric which is human accident rate, right.”
Douma explained that if the number of interventions that one would have to make to avoid an accident while driving on FSD comes down far enough, it will eventually cross under that threshold. However, he added that at some point, the human starts causing accidents for FSD when intervening, and that this is something that has been seen.
“Humans intervening with self-driving systems have caused some of the worst accidents that the self-driving cars have had, so that’s a real thing.”
He further explained that when those two numbers begin to cross, a decision at the program level whether or not to allow the car to override the human has to be made.
“When that becomes the only way for the system to get safer, now you’re taking the human out of the loop because that’s safer. You’re crossing that threshold.
“I think they’ve got this approximate metric that they think that they need to get to and they’re tracking the internal metric of FSD.”
Tesla, Douma noted, is watching the intervention rate continue to go down. To him, it sounded as if Elon was saying that, in the third or fourth quarter of 2022, those lines could possibly cross. However, it’s hard to draw a straight line to make a prediction when the progress line is curlier than Arby’s curly fries.
You can watch the full conversation with Dave Lee and James Douma here.