Comment by agumonkey | Hacker News Reader

Comment by agumonkey | Hacker News Reader

Parent Full thread

I have yet to see the simplest proof that comma ai work is just a cute trained NN over a camera, piped to basic controls, with some potential creative ideas (in a conf, hotz mentioned working on time based depth inference instead of the traditional stereo camera setup).

I'm biased for conservative companies, I have no idea what they're doing and how much resources. Maybe they're doing something at the same level. I believe they have more refined physics model to compute surroundings and car physics in a way to ensure sane and safe behaviour. Exception made of Tesla AP, which I always found borderline lying.

The medias and the crowd seeems to believe wild hacking and current year tech is giving you better products, but physics are physics. I'm still waiting for a proof of solid math, system and safety engineering there.

psedit; I forgot, I peaked at commaai github repo, the NN code is way above my head and what I could imagine, so I have to give hotz' team that.

Some people really have magic auras.

Actually all my money was on google, i expected them to drive into the scrapfield, after the car companys and Tesla pushed each other into a early market situation and upped Legislation, kicking themselves out of the market, with a carnaggedon. But they backed out, because their managers want the chicken to exist before the egg. So, why not bet on the outsider?

I liked Google's approach a lot. It seemed justly dimensioned, came out of Thrun DARPA success, had years and millions of miles of testing. What's not to love. Also G has the lever to accelerate and change tech if their old approach is obsolete.

Sad they went silent since a year or so.

I said it in other place, including HN, I don't consider Hotz an outsider, but mostly glorified hacker (his attitude with institutions and media don't help me clear this impression); and I don't think hacker mentality bodes well with this task.

Actually, that is a very humanistic point of view. There are third-world countrys where a margin of life-loss is considered "acceptable".

And sorry, but Teslas approach, seems to be a similar hacky "Wing-it with PR-Guns blazing".

About the first point, do countries where such acceptable risk still exist really need self driving vehicles that early ? Could be delayed a bit.

About the second point, see my above comment https://news.ycombinator.com/item?id=12841767

Isn't Google's approach dependent on expensive LIDARs and high resolution 3D maps [0], or have they changed that? I don't see that kind of system scale well.

[0] https://www.youtube.com/watch?v=YXylqtEQ0tk

I've read LIDAR cost was dropping fast, it was expensive at the time and as a short production item; nowadays it's even small.

About the 3D mapping you're probably right, I somehow expected it not to be a deep problem.

> in a conf, hotz mentioned working on time based depth inference instead of the traditional stereo camera setup

This sounds like structure from motion, SfM, which isn't new at all.

https://en.wikipedia.org/wiki/Structure_from_motion

Sure, nothing is new in that field. Space did all that decades ago, except their habits are just not "idiomatic silicon valley".

Reply on news.ycombinator.com