Tesla – Sleight of hand.

Tesla is still bad at autonomous driving.

• Tesla’s latest pronouncement on the performance of Autopilot suggests that Autopilot is safer than humans, but the reality remains that Tesla is very far away from real autonomy and I think well behind its rivals.
• Tesla needs to give the impression that it is good at autonomous driving because a good portion of the valuation of the company is attributed to the future robotaxi business that Tesla thinks will earn it 64% gross margins.
• Every quarter Tesla reports statistics regarding the distance that its cars are driving without an accident with and without the use of Autopilot.
• This quarter Tesla states that the miles driven per single accident were: 4.59m with Autopilot engaged, 2.72m with active safety features switched on and 1.79m with nothing.
• This gives the immediate impression that Tesla’s Autopilot is better than humans at driving, but the reality is quite different.
  • First, highway or city: As Brad Templeton (see here) points out 94% of Autopilot use is on the highway where the accident rate is much less in terms of miles driven.
  • For total miles driven by humans, 60% of them are on city streets where the accident rate is 3 times higher than on the highway.
  • When one corrects for this one finds that humans and the Autopilot are broadly equivalent to each other and have been that way for two years.
  • Second, Apples and oranges: This comparison between human and machine is not like for like.
  • This is because when humans make mistakes there is nothing to correct the mistake and prevent them from getting into accidents.
  • Whereas when machines drive there is always a human there to correct the machine.
  • This is why this measure is not a very good representation of reality.
  • Hence, a better measure of this would be disengagements where the human has taken over to prevent the Autopilot from doing something stupid.
  • Data from Lex Friedman (MIT) suggests that Tesla Autopilot disengages much more frequently which when placed in RFM’s analysis ranks Tesla 20^th out of 30 of the autonomous players (see here).
  • A better measure would be disengagements by Autopilot compared to every time the human makes a mistake.
  • Unfortunately, this is very difficult to measure but I suspect that because humans are allowed to drive and machines are not, that humans remain far superior at piloting vehicles.
• While I agree completely with Tesla’s philosophical approach to autonomous driving, I think it is still a long way away from making it work properly.
• Tesla says that because humans can drive with two cameras (eyes), two microphones (ears), no HD map, no infrared and no lidar that machines should be able to do the same.
• This will work as long as your cameras are as good at interpreting the surroundings as well as human eyes meaning that one needs to have solved the machine vision problem.
• Tesla makes this claim, but it is demonstrably false as there are many examples where Autopilot goes haywire because of a failure by machine vision to interpret the environment correctly.
• This is why the other systems use HD maps, lidar and infrared as they help to make up for the insufficient performance of the machine vision system.
• The limitations of AI and deep learning in particular are likely to prevent Tesla from reaching this goal before the costs of these additional systems has reached a point where they are ubiquitously deployed.
• Hence, I see no advantage for Tesla in the race to autonomous which combined with the large flaws in its business case for robotaxis (see here), keeps me very negative on the shares.
• Of all of the technology valuations, this is one of the most inflated.