Home Lex Fridman Episode
Lex Fridman · 2019-04-12

Elon Musk: Tesla Autopilot | Lex Fridman Podcast #18

Elon Musk explains Tesla Autopilot's vision, the new FSD computer, why driver monitoring may soon hurt safety, and AGI.

Elon Musk: Tesla Autopilot | Lex Fridman Podcast #18
The guest

Elon Musk: CEO of Tesla, SpaceX, and Neuralink, and co-founder of several other companies, leading Tesla's Autopilot and full self-driving effort.

What this episode covers

Lex Fridman talks with Elon Musk about the past, present, and future of Tesla Autopilot. Musk argues autonomy will make non-autonomous cars nearly obsolete and that a Tesla bought today is an appreciating asset because its hardware is already capable of full self-driving, with capability arriving via over-the-air software updates. They debate camera-based driver monitoring, with Musk contending that once the system is dramatically safer than a human, human intervention could actually decrease safety, using elevator operators as an analogy. The conversation closes on adversarial attacks against neural nets, the gap to artificial general intelligence, AI and love, and simulation theory.

The guest's own work

ProductBy the guest

Tesla Model 3

Tesla

“if you order any Model S or X, or any Model 3 that has the full self-driving package, you'll get the FSD computer”— Elon Musk
ProductBy the guest

Tesla Full Self-Driving (FSD) Computer

Tesla

“we've finally developed our full self-driving computer, which can process an order of magnitude as much as the NVIDIA system that we currently have in the cars”— Elon Musk
ProductBy the guest

Tesla Autopilot

Tesla

“It's a total game changer for quality of life, for using Tesla Autopilot on the highways”— Elon Musk

Also referenced (named, not recommended)

ProductReferenced

NVIDIA system (in-car compute)

NVIDIA

“to use it, you unplug the NVIDIA computer and plug the Tesla computer in and that's it”— Elon Musk
ProductReferenced

Cadillac Super Cruise

Cadillac

“another vehicle system that we were studying is the Cadillac Super Cruise system that's, in terms of ODD, very constrained to particular kinds of highways”— Lex Fridman
MediaReferenced

Her

Spike Jonze

“Do you think we will ever create an AI system that we can love, and loves us back in a deep meaningful way, like in the movie Her?”— Lex Fridman

Big reveals from this episode

  • Tesla has roughly half a million cars on the road with the full sensor suite, which Musk claims is about 99% of all such driving data.
  • Tesla's new full self-driving computer can process an order of magnitude more than the NVIDIA system, and you swap it by unplugging NVIDIA and plugging in Tesla.
  • Musk states the hardware currently being produced is already capable of full self-driving, with the rest being a software problem delivered over the air.
  • Musk's claim that buying a Tesla today is buying an appreciating asset, not a depreciating one.
  • Musk predicts that possibly by end of year or next year, having a human intervene will actually decrease safety.
  • Musk says the rate of improvement of the self-driving system is exponential.
  • Musk claims defending against adversarial attacks is easy: train the net on what is definitely not a car and exclude matrix-hack-like inputs.

Worth remembering

  • The conversation happened after Fridman's MIT group released a paper on driver functional vigilance during Tesla Autopilot use, and the Tesla team reached out offering the podcast.
  • Fridman states Tesla never financially supported his research, and he has never owned a Tesla vehicle or Tesla stock.
  • The FSD computer is two redundant systems on a chip, likened to a twin-engine commercial aircraft that can operate safely on one.
  • Tesla treats all driver input as error: if a user had to take over, something signaled a problem to learn from.
  • Musk suggests Autopilot may need to be 200 to 300 percent safer than a person for regulators to allow unmonitored driving.
  • Musk notes there are about 40,000 automotive deaths per year in the US, yet four in a Tesla would get a thousand times more press.
  • Fridman's team annotated about 18,900 Autopilot disengagements and found drivers maintained functional vigilance, against vigilance-literature predictions.
  • Musk's physics view of love: if an AI loves you in a way you cannot test as fake, then it is real.
  • If Musk could ask an AGI one question, it would be: What's outside the simulation?