Yes, there was some crazy stuff said. Reminds me of today's hype over generative AI in some ways.
I do think we might eventually get to a point where autonomous vehicles are clearly safer and people start talking about banning human-driven cars. Probably another 20+ years though.
Great article! But I can hardly agree that Tesla doesn't have the training data for the odd situation such as firefighter hose and police tape. As long as driver of tesla drives around these situations, the data can be collected and the shadow mode can be run in the background to verify the performance even if the system is not engaged.
There is one thing missing on profitability: for the most part Tesla will not own the cars, Teslas buyers will and Tesla will take a cut on every drive. There goes the huge cost of building a fleet of Robotaxis 🤷♂️
"...freeways have remained off limits. I suspect the reason is that the high speeds on freeways make it impractical for cars to come to a stop when they get confused."
My car does not claim to be self driving, but gets very upset when I get too close to someone on the road. Further, braking is technologically easy, and there are few freeway complexities other than staying between the lines and stopping. So why hasn't freeway full self driving (FFSD?) become a thing? RVs with FFSD would be an absolute game changer.
Because while freeway driving is “easy” in most situations, the consequences of a mistake are potentially catastrophic and it’s hard to be sure you’ve identified and handled all possible edge cases.
Self-driving cars? Easy solution: The cars stay on tracks, totally new tracks, built just for such cars, and limited just to such cars. And no walking dogs, cats, deer, bears, cows, etc. There is no snow. And, from beginning to any effort at anything stable, it will be a massive money loser.
The US military might do okay with some cases of self-driving vehicles in some cases of desert warfare, e.g., driving a truck carrying water. One possibility might be a 'train' of several such 'supply' vehicles, in a line, with the first vehicle driven by a human.
For self-driving on any significant scale, even accepting losing lots of money, without such tracks, no, none, nichts, nil, nada, zip, zilch, zero chance. Believing otherwise is worse than believing in the Easter Bunny and massively under appreciating what is crucial in driving a car.
Driving a car requires essentially full, general human intelligence, and believing in having that anytime soon is also massively under appreciating what is involved and much worse than believing in the Easter Bunny.
E.g., driving a car in real situations occasionally but crucially requires immediately solving problems never seen before, literally, never in all of driving in all of human history. The AI (artificial intelligence) 'training' techniques can't accomplish that. Here is a 'proof': Give the AI a high school book on plane geometry. Then have the AI solve the problem:
Given triangle ABC, construct point D on side AB and point E on side BC so that the lengths AD = DE = EC.
That exact problem? Tough to find in any available 'training' data. The needed solution technique? Not in common high school plane geometry texts and tough to find elsewhere. ChatGPT 4? Made no progress at all, got a grade of flat F. The training techniques of ChatGPT have no chance of solving this problem. Why? The solution requires some innovation that is close to real human intelligence.
I might do a follow-up about this. There are still companies working on self-driving trucks but I am not optimistic about them. I don't see how you can have the confidence to launch a driverless truck on the freeway. There are also a few companies working on package delivery, including Nuro. They are progressing slowly and I'm not sure why.
Late to the party, but I think it bears mentioning that GM has more incentives to keep funding Cruise than a dream of a profitable driverless taxi service - they're able to use the data & technology to improve their own Tesla-like driver assistance thing called SuperCruise.
A very well researched and argued piece. I agree with you about the state of Tesla's program not being as advanced as the others.
I do think the argument about the kind of things Tesla needs to do if they think a taxi service is in the near horizon (local mechanic teams, police etc) misunderstands the Tesla pitch (whihc I don't buy anyway). The pitch is that individual owners of Tesla's will be able to rent them on an app, where riders can summon them from wherever they are and go where they need to. An uber without drivers, with no company owning the vehicles. Simply Tesla taking some financial cut in the app for using it.
Not that the vision is anywhere near reality. And no current Tesla has the kind of hardware that would be necessary for that.
In my view, the low-hanging fruit use cases are commutes and micro-transit. Potential for big impacts and rapid adoption, vs. trying to tackle dense environments.
Everything else aside, I think you may be underselling the state of the hype in 2015....
https://techcrunch.com/2015/08/11/driving-your-car-will-soon-be-illegal/
Yes, there was some crazy stuff said. Reminds me of today's hype over generative AI in some ways.
I do think we might eventually get to a point where autonomous vehicles are clearly safer and people start talking about banning human-driven cars. Probably another 20+ years though.
They'll have to prize the steering wheel out of my cold, dead hands. I will riot before I accept a ban on manual driven cars.
The more likely scenario for most folks is that car insurance goes up 20x and people decide it's not worth it.
And once again we see how corporatism allows the rich to retain their freedoms while everyone else is coerced into giving up theirs.
The Uber fatality had a "safety driver"... who was watching a movie on a tablet rather than paying any attention to the road.
I clearly remember telling my kids (ages 6 and 9) in 2016 that they will never need to learn how to drive. I was wrong.
Great article! But I can hardly agree that Tesla doesn't have the training data for the odd situation such as firefighter hose and police tape. As long as driver of tesla drives around these situations, the data can be collected and the shadow mode can be run in the background to verify the performance even if the system is not engaged.
The moribundity of self-driving cars is not exaggerated.
There is one thing missing on profitability: for the most part Tesla will not own the cars, Teslas buyers will and Tesla will take a cut on every drive. There goes the huge cost of building a fleet of Robotaxis 🤷♂️
That might happen if they can get a robotaxi fleet going. I am skeptical they'll get there with their current approach.
Excellent article, thank you. But...freeways.
"...freeways have remained off limits. I suspect the reason is that the high speeds on freeways make it impractical for cars to come to a stop when they get confused."
My car does not claim to be self driving, but gets very upset when I get too close to someone on the road. Further, braking is technologically easy, and there are few freeway complexities other than staying between the lines and stopping. So why hasn't freeway full self driving (FFSD?) become a thing? RVs with FFSD would be an absolute game changer.
Because while freeway driving is “easy” in most situations, the consequences of a mistake are potentially catastrophic and it’s hard to be sure you’ve identified and handled all possible edge cases.
Self-driving cars? Easy solution: The cars stay on tracks, totally new tracks, built just for such cars, and limited just to such cars. And no walking dogs, cats, deer, bears, cows, etc. There is no snow. And, from beginning to any effort at anything stable, it will be a massive money loser.
The US military might do okay with some cases of self-driving vehicles in some cases of desert warfare, e.g., driving a truck carrying water. One possibility might be a 'train' of several such 'supply' vehicles, in a line, with the first vehicle driven by a human.
For self-driving on any significant scale, even accepting losing lots of money, without such tracks, no, none, nichts, nil, nada, zip, zilch, zero chance. Believing otherwise is worse than believing in the Easter Bunny and massively under appreciating what is crucial in driving a car.
Driving a car requires essentially full, general human intelligence, and believing in having that anytime soon is also massively under appreciating what is involved and much worse than believing in the Easter Bunny.
E.g., driving a car in real situations occasionally but crucially requires immediately solving problems never seen before, literally, never in all of driving in all of human history. The AI (artificial intelligence) 'training' techniques can't accomplish that. Here is a 'proof': Give the AI a high school book on plane geometry. Then have the AI solve the problem:
Given triangle ABC, construct point D on side AB and point E on side BC so that the lengths AD = DE = EC.
That exact problem? Tough to find in any available 'training' data. The needed solution technique? Not in common high school plane geometry texts and tough to find elsewhere. ChatGPT 4? Made no progress at all, got a grade of flat F. The training techniques of ChatGPT have no chance of solving this problem. Why? The solution requires some innovation that is close to real human intelligence.
What happened to all the other self-driving vehicles? Trucks and buses and whatnot?
I might do a follow-up about this. There are still companies working on self-driving trucks but I am not optimistic about them. I don't see how you can have the confidence to launch a driverless truck on the freeway. There are also a few companies working on package delivery, including Nuro. They are progressing slowly and I'm not sure why.
Great piece. I live in Scottsdale but haven't yet taken a Waymo. I'm going to have to do it just for the experience!
I'd love to hear about your experience if you want to shoot me an email about it.
Late to the party, but I think it bears mentioning that GM has more incentives to keep funding Cruise than a dream of a profitable driverless taxi service - they're able to use the data & technology to improve their own Tesla-like driver assistance thing called SuperCruise.
Interesting theory! I haven't heard about them doing this but it would make sense.
A very well researched and argued piece. I agree with you about the state of Tesla's program not being as advanced as the others.
I do think the argument about the kind of things Tesla needs to do if they think a taxi service is in the near horizon (local mechanic teams, police etc) misunderstands the Tesla pitch (whihc I don't buy anyway). The pitch is that individual owners of Tesla's will be able to rent them on an app, where riders can summon them from wherever they are and go where they need to. An uber without drivers, with no company owning the vehicles. Simply Tesla taking some financial cut in the app for using it.
Not that the vision is anywhere near reality. And no current Tesla has the kind of hardware that would be necessary for that.
Excellent article. Thank you.
What would be most exciting to me is the combination of FSD and driverless. That way, my own vehicle could come pick me up from the airport.
In my view, the low-hanging fruit use cases are commutes and micro-transit. Potential for big impacts and rapid adoption, vs. trying to tackle dense environments.
What about Lidar versus Tesla Vision? Tesla removed USS from its newer models which seems to have set them back
Won’t Tesla collect the data from sensors whenever the driver switches from FSD to manual?
Yes, but that won't necessarily yield enough data to allow learning whatever scenario the driver was avoiding.