Open Source Self Driving Car guy and his battle with Big Tech and government regulators

[Tartun]

https://reason.com/video/george-hotz...-needs-to-die/

Entertaining interview and interesting look at where self driving could be headed in the near future.

[Halyanne]

I don't think its too far away. Ford has promised that by 2023 their cars will be able to. I work for a company who does a ton of Traffic infrastructure and they just opened up a lab in Atlanta to have all the players, car companies, communications players, etc to come and test out their stuff. The biggest road block that we've seen is infrastructure handling driverless cars. Sure sensors and so on can do a lot of things, but your car needs to be able to talk to a traffic light.

I've included a couple video links.

https://www.youtube.com/watch?v=2l-Bq1i3CBE

https://www.youtube.com/watch?v=f5jVpQcfD-4

[Tartun]

I don't think its too far away. Ford has promised that by 2023 their cars will be able to. I work for a company who does a ton of Traffic infrastructure and they just opened up a lab in Atlanta to have all the players, car companies, communications players, etc to come and test out their stuff. The biggest road block that we've seen is infrastructure handling driverless cars.

I can't help but think that approach is part of the problem. When the cars depend on self driving infrastructure, what is to prevent them from suffering the same pitfalls as rail systems? What solves the problem of dealing with the errors of human operated vehicles? Jaywalkers? Any time someone breaks a rule? I think better ai and not infrastructure is the solution.

[Aristotle]

I am so ready for 100% self driving cars on the road.

We humans have been driving for over a hundred years and we clearly suck at it. 40k people die every year in car accidents, yet we will obsess over 1 guy dying in a self driving car for 2 years.

We had our chance. We sucked. Time to turn it over to the machines.

[Xywalan]

I'll have to listen to the interview but I'm skeptical, a lot. The current hype in AI is massively overblown and they are the result of improving old ideas and combining them with better hardware and they all boil down to deep learning which has serious problems and limitations:
Basically, in deep learning you create a collection of "neutrons", a very misleading term BTW since it has nothing to do with biological neurons, and feeding it massive amounts of data and adjusting it's knobs and wheels until it works. It is great for pattern detection jobs where we have a lot of data, cat pictures, human pictures, games, etc but they are terrible at truly novel scenarios.
And that's why I am skeptical. Humans are great in reacting to unusual situations compared to the things and I think that's the big challenge of self driving cars. They are great for keeping the car on the road, keeping safe distance to the next car, or any other type of usual scenario but they can't handle anything else.

[Tartun]

I'll have to listen to the interview but I'm skeptical, a lot. The current hype in AI is massively overblown and they are the result of improving old ideas and combining them with better hardware and they all boil down to deep learning which has serious problems and limitations:
Basically, in deep learning you create a collection of "neutrons", a very misleading term BTW since it has nothing to do with biological neurons, and feeding it massive amounts of data and adjusting it's knobs and wheels until it works. It is great for pattern detection jobs where we have a lot of data, cat pictures, human pictures, games, etc but they are terrible at truly novel scenarios.
And that's why I am skeptical. Humans are great in reacting to unusual situations compared to the things and I think that's the big challenge of self driving cars. They are great for keeping the car on the road, keeping safe distance to the next car, or any other type of usual scenario but they can't handle anything else.

I've been following this guy's approach for a little while now. What he's doing is crowdsourcing driving meta-data from volunteer drivers who put his device into their car's onboard computer. The AI is then supposed to take all the data from the human drivers and learn how to drive by observing human drivers. So instead of trying to think of every scenario and program an AI to respond a certain way in the provided scenario, the hope is that the AI will develop a model for proper driving much quicker by analyzing the meta-data provided by the existing car cameras, radar and onboard computer records.

[Xywalan]

I've been following this guy's approach for a little while now. What he's doing is crowdsourcing driving meta-data from volunteer drivers who put his device into their car's onboard computer. The AI is then supposed to take all the data from the human drivers and learn how to drive by observing human drivers. So instead of trying to think of every scenario and program an AI to respond a certain way in the provided scenario, the hope is that the AI will develop a model for proper driving much quicker by analyzing the meta-data provided by the existing car cameras, radar and onboard computer records.

I am skeptical. It's likely deep learning and deep learning is not magic. Also, the big problem with deep learning is that nobody knows how it works. Nobody can guarantee that the AI won't randomly decide to do crazy shit for no reason. When your AI classifies a traffic light as a giraffe, it's funny but not when it's driving your car.

[Tartun]

I am skeptical. It's likely deep learning and deep learning is not magic. Also, the big problem with deep learning is that nobody knows how it works. Nobody can guarantee that the AI won't randomly decide to do crazy shit for no reason. When your AI classifies a traffic light as a giraffe, it's funny but not when it's driving your car.

I think that this guy is pretty realistic and acknowledges these issues. That's why he makes the claim that truly self-driving is mostly hype at this point. His position is that AI will get us there faster, more efficiently and more effectively than infrastructure based self driving, as well as keeping the price-tag attainable and delivering an actual product to consumers that they can own, as opposed to a glorified robot taxi. He admits that the current state of "self-driving" is actually driver assist. His product uses a modified smart-phone mounted on the dash and monitors driver engagement with the front facing camera. What he tries to avoid is 'hype'. He wants to make a product available that does a specific thing and avoid claims that it can do anything else, but believes that AI will get us to fully self driving in the future. I think he's right and I like his approach.

[Tartun]

I am skeptical. It's likely deep learning and deep learning is not magic. Also, the big problem with deep learning is that nobody knows how it works. Nobody can guarantee that the AI won't randomly decide to do crazy shit for no reason. When your AI classifies a traffic light as a giraffe, it's funny but not when it's driving your car.

I'd encourage you to watch the interview. You might be surprised to learn how much this guy would agree with you.