Comment by throwaway314155 | Hacker News Reader

Comment by throwaway314155 | Hacker News Reader

Parent Full thread

throwaway314155

That's not usually the argument. It's that LiDar has historically been working more reliably than tech that doesn't use it - and reliability is pretty important in that industry.

In any case I'm curious what gives you such blind faith that Tesla's vision-only tech will work out. There's a lot of really practical and frankly obvious reasons why it hasn't been working for a very long time and similarly lot's of videos (which are aggressively taken down by Tesla or can't even be posted due to shady customer NDA's) of their self-driving equipment getting dangerously close to murdering people. And of course, there are the _actual_ cases of Tesla's self-driving tech just you know, straight up murdering people.

You can shift blame to the humans not being attentive enough to take control when self-driving mistakes come up, but then you've lost the broader argument that self-driving Tesla's are "ready", even if you're right about the former.

Edit because fuck-you-HN-comment-limit-soft-ban:

(response to child comment below):

> Obviously, only vision and hearing are required for driving

Human vision is drastically better than even the most advanced computer vision systems we have or are likely to have any time soon - especially when there are realtime and hardware constraints. That's before you introduce the relatively (!) shoddy approximations that even SOTA vision models learn via SGD.

> I'm curious what gives you such blind faith that Tesla's vision-only tech will work out.

Obviously, only vision and hearing are required for driving. The question isn't "can it work," the question is "which approach will be economically viable on what time scale, and is there a winner-takes-all dynamic that excludes the approach that comes second."

I've never understood the argument "humans only use vision and hearing so self driving cars should too". We don't constrain what we do with machines to what humans do.

Humans don't have wheels either. Is Tesla planning to replace wheels in their cars with legs, Flintstones style?

That's not the argument and never was. Humans aren't proof that you can't do better than vision, we are proof that the problem can be solved with vision.

This isn't a question of max-sensors vs min-sensors, though. It's a question of max-sensors vs max-data under economics constraints that forbade "why not both?". Tesla chose max-data, and they now have a 40x data advantage thanks to their choice, and probably more once you consider the improved diversity of data collected from private cars in every corner of the country and world.

If you were training a self-driving car, would you rather have 40x the data or would you rather have point clouds that are only a smidgen better than what you can get from modern monocular depth reconstruction on the vision data?

> Humans aren't proof that you can't do better than vision, we are proof that the problem can be solved with vision.

It's proof that the problem can be solved with vision when your driver is a human. That doesn't necessarily prove it for non-human drivers implemented on currently known AI architectures.

Humans can learn from a lot less data then current non-human AIs require. We also learn in the field, often recognizing when we've made a mistake and consciously updating ourselves to handle similar situations better in the future. It is possible that this is necessary to make vision only driving work well.

You dodged the question about whether you would choose 40x data or redundant depth maps.

I think you dodged the question because you know that 40x data is the clear winner.

I didn't dodge the question. I didn't even consider it enough to form an opinion.

I responded to the part of your comment that interested me, which concerned what we can infer from the ability of humans to do vision based driving.

You dodged it again lol.

I didn’t say nobody should try to solve the problem using lidar, I just said humans are an existence proof that it can be done. Lidar gives you more data, but maybe that data will turn out to be overkill. We’ll find out.

Reply on news.ycombinator.com