Tesla needs to come clean about HW3 before the word 'fraud' comes out
electrek.co
electrek.co
If they crack the human perception algorithm, they'd have the cheapest robotaxis on the road and have multi trillion valuation.
However cracking human perception seems much more complex than what was estimated and foretold.
I'm grateful we have multiple US companies trying different approaches. Competition is a wonderful gift.
Who was estimating that cracking how the human brain works, specifically the visual cortex together with our decision-making, would be the easier path to this endeavour?
It was clear it would be a very complex thing to anyone who actually knows about the human brain, the rest was just wishful thinking.
And a huge amount of tactile feedback.
It drives well in your city but we don't know anything about how has Tesla tailored that experience to your city. Did they need to run around mapping it with their own LIDARs and feed into some learning model to provide it in your city? Because that won't scale. Did they use only camera vision? Because there's also multiple videos showing FSD not driving well at all in some cities.
We cannot tell, and that's the whole problem, your anecdote that FSD works well for you is a N = 1, there are other reports on HN about others having the same experience as you, it's great it works for you, people! It doesn't mean it's been generalised and works in any city, we simply cannot know, and the longer time passes the less trust the general public has with Tesla.
As another anecdote: I know at least 2 people for which FSD doesn't work well in their cities, to the point where they're afraid of turning it on even when paying full attention, and both regret paying for it. This is not a good sample either, we just don't have enough data.
Spot on.
I am keen to see how FSD manages to execute hook turns to accommodate for the trams here in Melbourne, AU.
I Just came from 30h+ driving long weekend. FSD did 25h+ of that driving. It's definitely a value even in its current form, probably not as expensive as Tesla charges thought.
p.s. I was looking at Rivian, but it seems like their delivery of affordable cars is in limbo.
Does that mean I can claim I made a mistake if I do something wrong and get caught?
Not really. A judge recently ruled that it was 'corporate puffery' in a very similar case.
https://www.theverge.com/2024/10/1/24259588/tesla-lawsuit-au...
> ... if Tesla knows that it can’t deliver unsupervised self-driving on HW3, it needs to let owners know right now and stop selling the software package to HW3 owners without a clear plan to make things right. Otherwise, this quickly becomes fraudulent.
HW5 will feature analog optics and computing? How steampunk!
>At the minimum, it will have to reimburse that, but owners can even argue that they bought the car because Elon Musk told them it would become self-driving over time and become an “appreciating asset.”
>This could quickly become a very large liability for Tesla, and the way it handles it is also important.
It might appreciate, once, if Tesla ships FSD and creates a robotaxi market.
However, something in active production, providing a commodity value add, and competing with newer versions of itself is never going to continually appreciate.
It's also just a really weird attitude, that this will turn out to be a massive liability and that they'll need to think about retrofits. The truth is simple: They didn't sell that many of these cars, of those they did sell, the majority won't have bought the self driving package for the exorbitant price and most owners will have upgraded to a newer model anyway. So what's the plan? They'll do nothing, they won't admit it won't happen they'll just say they'll get to it one day, continue to offer relatively small incentives to upgrade away from these cars anyway. What are you going to do? Sue them. Good luck. You have literally shown that over the last 8 years he can just wave his arms and say "Well we'll get to it one day" and you believe he's fulfilled his obligations, so he'll continue doing that.
Senior people at Tesla are still repeating the claim, which freshens liability. And Tesla will still sell you FSD on HW3.
So they need 122,000 miles between disengagement? Really? That's a rather crazy standard.
"Unsupervised" sounds like you can take a nap in the back while the car takes you somewhere to me.
What's an appropriate safety standard the self-driving car must meet to make that safe enough? I'd say that the self-driving needs to be [at least] as reliable as a human driver.
If someone said that the average human driver has some kind of failure (loss of attention causing accident, feels unwell and needs to pull over, etc) once every 122,000 miles, that sounds about right to me.
(The data also suggest widespread self-driving cars should prompt adjusting the unsupervised driving age up to 18 to 20.)
[1] https://www.friedmansimon.com/faqs/how-common-are-car-accide...
Your source says a crash every 70,000 miles for 17 year olds, to my understanding. Number of unsafe actions that would warrant a disengagement, but didn't happen to cause a crash, would likely be significantly higher (e.g: the 17y/o driver speeds 100 times, and crashes once).
Correct.
> Number of unsafe actions that would warrant a disengagement, but didn't happen to cause a crash, would likely be significantly higher
Why? When I’m in a Waymo, I’m taking a nap or entirely uninvolved in the driver nearly 100% of the time. I am worse situated than a drunk, high teen because I’m in the back seat and not even pretending or trying to pay attention. And that’s Level 4. If you’re marketing Level 5, every intervention should be treated like an autopilot failure in aviation: more than we do for crashes.
If I’m in a car with a driver where I need to regularly grab the wheel to keep them from colliding, they’re an unsafe driver.
What percentage of speeding, for instance, do you think materialises into a crash? Even the most dangerous unsafe actions like running a red light won't cause a crash most of the time. So if a 17y/o crashes every 70,000 miles, they're doing something unsafe far more frequently than that.
> every intervention should be treated like an autopilot failure in aviation: more than we do for crashes
I think the benchmark for now should be whether it's safer than a human driver. We could compare human crashes caused to autonomous crashes caused, or unsafe human actions to unsafe autonomous actions, but I believe you're trying to conflate actual human crashes to unsafe autonomous actions.
For a Level 5 system, yes, a situation where the car requires intervention is equivalent to a crash or avoided crash. Particularly if it won’t have a steering wheel or pedals.