Hacker News Reader

Comment by brianbreslin | Hacker News Reader

Something seems fishy to me. Cruise which GM bought for $1B a few months back was working on similar tech. Having only raised $3M he could have EASILY sold his tech/team to Chrysler or Ford or Hyundai, or any other car company for $100M and made his investors super happy and himself very rich.

So why "quit" when the feds ask you to deal with regulations that involve passenger safety. Its not like he had to spend $20M on clinical trials.

View on HN

It is not "similar tech" at all. Yes, both approaches involve some machine learning (which is a huge field). But Cruise's approach is similar to Google/Uber's approach in that it relies on LIDAR. I'd assume (but might be wrong) in that it also relies heavily on premapping, as does the Google/Uber approach.

Comma's approach was complete end-to-end learning with just a single camera (to give you some context, research problems involving two cameras are already considered distinct (but related) research problems from those involving one camera), which is an extremely different approach (the Google/Uber approach can be thought of as more "handcrafted"). From the deep learning researchers I know of there was heavy skepticism of Comma's approach. End to end learning makes for nice demos but is still an active research topic.

From what I can tell, Mobileeye is somewhere in between an end-to-end approach and a handcrafted approach, so I would guess Tesla is similar to Mobileeye on that spectrum.

There are more details in this lecture Hotz gave at Berkeley this September: https://www.youtube.com/watch?v=Hxoke1lDJ9w

One of the things that seemed missing from the lecture when I watched it is how data was going to be collected for handling exceptional situations/collision avoidance. Hotz also glossed over the question in this interview from July: https://www.youtube.com/watch?v=2zy_07g2IrM#t=35m

I suspect they ran into trouble with edge cases. Also suspicious is this job posting on their website right now:

"Localization Lead Engineer DESCRIPTION

We have over 50,000 miles of video data from cars, and will have millions by the end of the year. Looking for someone to build a SLAM algorithm capable of scaling to the world.

Imagine an API that, given a picture, returns the exact location of the camera with cm accuracy.

Build this

REQUIREMENTS

    Strong math background.
    Ability to write concise, reliable, and readable code quickly.
    Github with stars is nice.
    Computer vision experience a plus.

Basically someone who could have written https://github.com/mapillary/OpenSfM or https://github.com/raulmur/ORB_SLAM2"

http://www.comma.ai/positions.html

Hotz has been claiming that they would not need mapping and SLAM like Google's self-driving car. The whole "Looking for someone to build a SLAM algorithm capable of scaling to the world" sure sounds like "I guess this time he couldn't find anyone on IRC to finish the hard parts for him and let him take the credit" that lawnchair_larry posted elsewhere in this discussion (https://news.ycombinator.com/item?id=12818219).

Is that IRC jab related to the PlayStation cryptosystem breaking ?

Yeah. Those are all really hard problems. Visual SLAM is far from solved, especially if since this is only monocular SLAM. And even if he did solve that to a pretty decent level (which is already multiple top-PhD theses worth of research), that would only get him to the point of where he would be is he had just mounted LIDAR on cars and said "now what?"

> Looking for someone to build > a SLAM algorithm capable of > scaling to the world.

This has got to be a joke. Text like this proves he was out of his depth

Anyone who thinks you can easily up and sell a small pre-product company for $100M lives in a different reality than I do...

Yeah this seems like an easy out that allows him to save face.

Unless he tried this already and anyone who took a serious look saw that he had nothing of value.

Every major player in the automotive industry is probably further along in development than this product is. They've probably been running cars on the road with similar capabilities for about a year or more now.

Suspicious especially given Hotz's comments about Cruise: https://www.youtube.com/watch?v=2zy_07g2IrM#t=40m10s

He posted the answer to this on Twitter: "Would much rather spend my life building amazing tech than dealing with regulators and lawyers. It isn't worth it."

Money isn't everything to some people. He has the true hacker ethos.

For me, hacking is about understanding systems and altering them in a way they behave favourably to you. The right thing to do when a regulator fires letters to you is to build a firewall of managers and/or lawayers to ensure they don't penetrate the hacking activity.

True hackers ship.

Reply on news.ycombinator.com