Comment by hackerlight | Hacker News Reader

Comment by hackerlight | Hacker News Reader

Parent Full thread

No twist of fate, but self-sabotage in the form of being unwilling to use Lidar. Look how hard Waymo and Cruise are finding it to do robotaxis in a small geographical area and with Lidar. Tesla is trying to do robotaxis everywhere without using Lidar. They're solving a harder problem and are handicapping themselves while trying to do it.

Here's why Lidar makes sense. You want independent information streams that are cross checked. If both streams tells you there is no object in front of you, that's a way lower probability of being wrong than just one stream telling you that. You go from probability p to p^2 of being catastrophically wrong. It's the kind of probabilistic added safety of being in an aircraft that's capable of flying with only one engine. Even if one blows up with probability p, you're fine. You need a p^2 event for both engines to malfunction, assuming errors aren't correlated, which they probably are a bit, but not so much that the added benefit isn't massive.

That analysis only works when p is non-vanishing, though. In point of fact the "car drives into something it didn't see" failure mode doesn't really exist, or if it does it's at a level much smaller than that of the human driver its intended to replace.

It works regardless of how big p is. Engine failures in aircraft are exceedingly rare but this statistical effect of needing p^2 events instead of p events has saved many lives. Having multiple pilots has the same rationale. While errors are correlated, getting an error past two pilots is roughly p - p^2 less likely than if you had a solo pilot.

Self-driving, as a mission critical endeavor, needs to be leveraging the same statistical phenomenon in order to reduce the exceedingly rare but highly impactful mistakes.

And by "something it didn't see" I'm talking about CNN/ViT mistakes/hallucinations. I don't see any evidence that it occurs at a level that's much smaller than humans. Any safety data we have is with human oversight. So the majority of errors are caught by humans and won't show up in FSD crash frequency or other publicly available data.

> Self-driving, as a mission critical endeavor, needs to be leveraging the same statistical phenomenon in order to reduce the exceedingly rare but highly impactful mistakes.

Uh... why? Shouldn't the goal be to be safer than people?

Basically that's just bad engineering: If your argument is not to deploy an existing system that is better than the current installed base because some other technique might be better still, you are harming your metric (safety) and not helping.

I'm talking about robotaxis. There is no safety data available for vision-only robotaxis. It's a hypothetical system that we are speculating about.

If Tesla had one that's safer than humans, then yes they should be allowed to deploy it. My argument was that they're making it harder for themselves to exceed that safety threshold by insisting on vision only.

Counterpoint - Tesla is trying to avoid a local maximum where they are able to do robotaxi's in cars with $300,000 of equipment on roads that are precisely mapped down to the centimeter.

Waymo and Cruise may have beat them to "first robotaxi" but I'm not convinced that they have anywhere near a scalable, sustainable model.

This is analogous to saying "we're trying to avoid a local maximum where we are required to use expensive chemical propellants in order to fuel our space rockets; therefore we're going to double down on stuffing them full of larger and larger quantities of gunpowder".

Reply on news.ycombinator.com