Some thoughts in no particular order, edited from an IRC discussion I just had on this:
The price is extremely low. If you drive just 20 minutes per day, and your time is worth at least minimum wage, it makes sense to buy this for just a year, and hopefully it lasts longer than that. But there is a monthly subscription cost ($300 per year) i didn't factor in. And wtf, since when does hardware have a subscription cost? Oh, and to purchase one you need to have lots of "points" with the company, which you can earn by using their app. This is really weird, why not just raise the price if you can't meet demand?
The price also doesn't seem high enough to cover the risk. If each accident costs 10 million in lawsuits (i have no idea how much they will actually cost, but that seems like a good made up number, and corporate legal fees are expensive anyway) then they better have less than 1 accident per 10,000 vehicles that use it, just to break even. honestly i hope they have 0 accidents because they could bring in regulators and ruin the fun for everyone, as well as the very delicate reputation of self driving cars (people are already predisposed to distrust machines)
I can't believe this works at all. Wouldn't you need like multiple gpus to process all this camera data in real time? How do you account for the variance between vehicles? How many cameras do they have and what is the resolution? That unit looks really small.
It looks like it's just trained by predicting what a human would do. Human drivers are imperfect, and so I imagine this machine will learn to imitate the imperfections of human drivers, combined with it's own imperfections. It's also unclear how they trained it, because the dash cam videos don't give information on when users tapped the brakes or moved the steering wheel. They must be doing some interesting semi-supervised learning.
Apparently the cars have limits on how hard it can slam on the brakes or turn the wheel. He advertises this as a good thing that will prevent the AI from doing anything too bad, but I don't see how. This limits what the AI can do in a real emergency. Worse, it suggests that the AI needs these training wheels. That without them it would cause accidents, and that's very concerning.
I just can't believe that this works. Deep learning is completely amazing and can get 99% accuracy, but this requires 99.9999%. Those extra 9's are super super hard. Fortunately this doesn't need to be perfect, it only needs to be better than humans. And while deep learning is starting to beat humans in some machine vision tasks, I'm not confident this is true in general. In the real world there will be weird edge cases, like a car painted in a very atypical pattern, or ice on the road, or a bug on the windshield, etc. Humans are more capable of handling new, out of sample situations, and machines can fail epically at them.