Tesla’s full self-driving computer is now in all new cars
techcrunch.com
techcrunch.com
I think it's very cool Tesla is building its own chips. But the rate of improvement in software just doesn't seem to be there yet, if Autopilot's public updates are any indication. It's hard to believe that AP could have these kind of basic regressions from just a month ago [1], while it purportedly is solving much harder problems.
EDIT: Not all of the US, see below.
http://www.mit.edu/~jfc/right.html
States where it is illegal to pass on the right:
Arkansas, Illinois, Kentucky, Louisiana, Maine, Massachusetts, New Jersey (My note: In 20 years of driving in Illinois, I have never been cited for passing on the right)
I guess the multi-lane open highway is the reason it's ok? Does that make those multi-lane open highways a free for all? No concept of a fast lane, overtake wherever you choose?
However, Colorado has a "left lane law" that applies on two+ lane highways with speed limits above 65mph. It is illegal to impede traffic in the passing lane, even if you're going the speed limit. https://codes.findlaw.com/co/title-42-vehicles-and-traffic/c...
A different law requires drivers in the passing lane driving below normal and reasonable speed, to change lanes (move to the right or pull off the road). If you're driving at the speed limit, it's by definition not less than normal or reasonable. And prior to the "left lane law" state patrol couldn't do anything about impeded left lane traffic so long as the impeding vehicle was moving at the speed limit, but since the 2004 "left lane law" you can be ticketed for impeding traffic in the passing lane even if you're driving at the speed limit (and you were always subject to ticketing if you were driving below the speed limit in the passing lane).
Rolling around a road, in broad daylight, with no adverse conditions and no "interesting" events isn't progress.
yes, merging onto a freeway is mildly impressive, but its not anywhere near state of the art.
What I'd like to see is what happens when someone swerves out in front of the car? does it stop dead? or panic and hand back control to the driver, just in time to crash? What about rain, or snow?
How does it handle diversions? what about new roads?
None of this is really about hardware, sure more power helps, as does sensors with lower noise floors and greater pre-processing. The difficult part is, and has always been, software.
What Tesla is doing is knowingly allow people to to unsafe software to gather scenarios that can be used as learning events. Not only that, they are getting people to pay a premium. I imagine that tesla will then rent the next generation of trained models back to the drivers, even though they paid for, and provided the training set.
As a Model S owner with over 10k miles on Autopilot, I take issue with you calling the software unsafe (I have never once had a disengagement that caused a safety issue), and I am happy to pay a premium to contribute to what I consider progress. I would happily pay more. Waymo vehicles aren't available to anyone buy a few beta testers. Tesla Autopilot neural net improvements are available to everyone in the fleet.
But, it is immoral to ship something thats barely level 2 and market it as an autopilot. Yes, they have made great strides, but tesla basically gambled that no one would crash.
This is the issue I have with our industry. That kind of slapdash "lets work 100 hour weeks and ship anyway, yes the target is impossible but a broken MVP is better than no MVP" is fine for a webapp. But for life critical software, no, no its not.
I do agree much more work needs to go into how life critical systems are developed and deployed, but impressed at Tesla’s development velocity without sacrificing safety (as evidenced by so few deaths in Tesla vehicles overall).
- Passing on the right. I almost never do that.
- Speeding: 75 in a 65 (0:28)
- Merging into a lane next to a car that was pulling off on the shoulder. That's a recipe for disaster. (0:41)
- Speeding: 70 in a 65 (1:03)
- The car angling the wheels right before a left hand turn (1:12)
- Merging over a solid line (1:15)
- Merging into the right turn lane at 30 and then taking the turn at 20. (1:20)
Some conditions may apply. Driving under inclement weather conditions, or in a dirty car, or in winter, during heavy traffic, on a road with poorly painted etc.. is not supported. /s
That's everything right there. It's not 4/5 if it's in tightly controlled circumstances. I can make anything level 4/5 if I get to control the conditions. It's the same as claiming 100% model performance by testing on training data.
Off by a few decades.
Why would they? their whole plan from the beginning was to do a reverse Uber - a fleet of self driving cars that acts as a taxi service.
Cost of a model 3 to TSLA: 25k
Average Revenue per year for self driving taxi: $12/hr * 24 * 365 = 105k
Producing just 100k cars a year for this program would generate 10B dollars a year in revenue, if they can make 1M cars a year, it's 100B... making TSLA a trillion dollar company even with a Ford valuation.
Most of that revenue will be more profitable than a gas powered car since they also own an energy company, and the software that runs the cars.
(Very glad I got a reservation worth of stock when the Model 3 was announced a few years ago, so I am biased)
No competition fallacy. If some market exists that's $<huge>, you can only sustain a high profit margin in serving that market if there are no competitors (which seems unlikely).
Also falling into the "self driving is possible" fallacy..
[0] https://blogs.nvidia.com/blog/2019/04/23/tesla-self-driving/
I had to double-check to make sure that April 1 had passed. At this point it's hard to take Musk's "Full Self-Driving" schtick seriously. They developed their own custom silicon, and now they "just" need to write the software that will allow it, plus the sensors on a Tesla, to be "Full Self-Driving."