Just started watching this so I have no idea what’s in it, but I imagine it’s going to be our best resource for what Tesla is doing…
Ok, just finished. Really wished they went into more detail. But if you’re not gonna get more detail with an interview between Lex Fridman and Elon, then it makes me think that this is as deep as it’ll ever get with Elon. It does sound like they’re using imitation learning though for path planning, ya?
Better hope that Karpathy is able to speak at length at the Autonomy Investor Day
I actually gasped out loud!
But now I’m disappointed because you say that Elon didn’t go into much technical detail. I’ll have to listen and hear for myself.
Bad machine transcript:
Not sure. At 11:16, Elon says:
Well, there’s a lot of things that are learnt. There are certainly edge cases where say somebody’s on Autopilot and they take over. And then, okay, that’s a trigger that goes into our system that says, okay, did they take over for convenience, or did they take over because the Autopilot wasn’t working properly.
There’s also like, let’s say we’re trying to figure out what is the optimal spline for traversing an intersection. Then, the ones where there are no interventions are the right ones. So you then say okay, when it looks like this, do the following. And then you get the optimal spine for navigating a complex intersection.
Lex seemed more interested in talking about driver assistance than full autonomy (whereas Elon was more interested in talking about full autonomy). Maybe if the interview had been longer and if Lex had asked more questions about this sort of thing, we would have gotten more information about Tesla’s technical approach.
I really hope Autonomy Investor Day is a deep dive. Even if the webcast is targeted toward a general audience, maybe off-mic Tesla will answer technical questions from investors at the event. That’s why I wrote my Seeking Alpha article: in the hopes that someone might read it and ask Tesla these questions.
That’s why I wrote my Seeking Alpha article: in the hopes that someone might read it and ask Tesla these questions.
I have a friend who is attending, but I don’t think he has much interest/knowledge in the intricacies. I’ll see if I can get him to ask a question…
To me, the most interesting thing Elon said — besides the thing about learning the optimal spline for traversing an intersection — was (at 24:40) “the rate of improvement is exponential.”
If that’s indeed true — if the performance metrics for Tesla’s neural networks are improving at an exponential rate — then I can see how Elon could be extrapolating a trend out 6 months where it hits Tesla’s estimate for human-level capability. I hope Elon’s confidence in the end of 2019 timeline comes from quantitative evidence, as opposed to gut intuition or qualitative evidence.
If Elon is just observing a quantitative trend, then maybe he’s right. If he’s making a subjective estimate, then it’s hard to put much stock in it because he’s been so off with autonomy-related things before.
Beautiful, thanks for trying!
Ooh! Another little hint! At 14:53, Elon says:
…then refining the neural net and the control software
As I understand it, in a self-driving car context, control typically refers to:
how the vehicle uses steering, throttle, and brake to move through the world, executing a trajectory created by the path planning block.
The control problem is the problem of:
how to apply force to an object to control its movement.
The fact that Elon said “the neural net and the control software” makes me think the control software is regular ol’ Software 1.0.
The person in charge of the control software at Tesla is, I believe, Frank Havlak, whose title is Senior Staff Controls Engineer. He has a background in robotic control.
A small clarification from Elon:
We now know for sure that Tesla is using Autopilot interventions as a training signal for neural networks. Still not sure whether it’s IL, RL, or something else.
Impressive timelapse video of Autopilot driving on the freeway with no human input:
Lex Fridman just started a Patreon for the Artificial Intelligence podcast. He’s trying to get to 500 patrons:
Wanted to bring this back up again as I just noticed that it’s been about six months since this podcast aired. Elon stated on the podcast (around 18:09) that he thought that Tesla would still need to require hands on wheels for about six months from the time that was aired until they could drop that requirement (I think he meant more from a capability standpoint, barring any regulatory hurdles)
His trajectory was so aggressive. I’m curious if he still believes feature complete is coming before the end of 2019. I know the Tesla IR team is saying feature complete some time in 2020, so they are contradicting his statements a bit
I think he said 6 months after the feature complete version is pushed to customers. So if the feature complete version is pushed in December 2019, then customers can take their hands off the wheel in July 2020.
Correction: at 17:55, he says “at least six months”. “At least”, so it’s a minimum, not a maximum.
True, he did say at least
He says “at least six months from here”, which I interpreted as being from today (April 12)
The timeline he’s reiterated a few times is feature complete by the end of 2019, superhuman by the end of 2020 (possibly as early as July 2020). So I think that’s the one to go with.