If they followed the same approach, with just cameras, and no maps they would be in the same state they are now. Pretty much exactly. If they used maps they would perform much better, but still have many issues.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Analysis of the price-hike for FSD, and the options it allows Tesla
- Thread starter bradtem
- Start date
willow_hiller
Well-Known Member
Then it's pretty clear you hold strong biases against Tesla. Even though it's just a hypothetical situation, there is some evidence that Tesla can perform just as well as services like Waymo in limited areas.
If you recall, there was a test drive given to hundreds of investors after Autonomy Day 2019:
So if Tesla had proceeded from that test drive, carefully working street by street on the area around Fremont, they'd probably have a Waymo-like service in a Waymo-like service area. But instead they're focusing on a generalized solution.
LIDAR isn't a magic bullet that's given Waymo their apparent advantage. It's being able to tailor driving policy to small areas with favorable conditions.
Last edited:
Tam
Well-Known Member
A different 2016 video is among the reasons the DMV is taking Tesla to court to ensure viewers are not deceived.
The New York Times interviewed 19 former Tesla employees. Among what was hidden from the public is that the car was well prepared with the pre-charted route and 3D mapping, and still, it collided with Tesla's private property roadside barrier. The filming was interrupted while the car's damage had to be fixed.
It made the documentary shown on Hulu and F/X: "Elon Musk's Crash Course."
Tesla needs a better software team.
The New York Times: Tesla Hit a Concrete Barrier in Original Full Self-Driving Video
Former Tesla employees that worked on Full Self-Driving are speaking out anonymously, but is FSD safe in the meantime? It doesn't seem like these EVs should be on the road unmanned.
www.motorbiscuit.com
How well does Tesla evaluate itself? Listen to what it discloses on the FSD beta warning that it "May Do the Wrong Thing at the Worst Time."
That is not the standard to let your car drives itself. At least, that's not stable enough for an AutoSteer with a hands-free option.
Last edited:
willow_hiller
Well-Known Member
Nope, you're mistaking it for this video: Tesla Self-Driving Demonstration
The one I posted is from 2019, and it immediately followed Autonomy Day where they had a fleet of Model 3s and offered test drives to all the investors that attended.
Tam
Well-Known Member
I indeed was confused with two different videos. Thanks for the clarification.
It's possible that if Tesla did homework first by pre-charted route, 3D pre-mapped, mishaps could have been reduced such in cases as:
The 2018 Mountain View fatal accident: 2017 Model X entered the gore point and collided with the median.
2022: AI Addict hit the green bollard.
Last edited:
willow_hiller
Well-Known Member
This is exactly my point. It's clear that Tesla was only able to take investors on an FSD test drive around Fremont in 2019 because they had premapped and painstakingly tested every turn that would be taken. But that's not a scalable approach to level 5.
Take Chuck's UPL as an example that demonstrates the differences in approaches between Tesla and other companies. Tesla could have mapped that exact intersection, written instructions on how to navigate it, and performed the turn perfectly every time. But instead they took that class of turns, improved their occupancy network in order to better predict oncoming traffic, wrote a creep network so the vehicles can estimate how far out to crawl for visibility, wrote a median detecting network, and wrote driving policy that can handle split highway UPLs in the abstract. It's less than 100% reliable, but it can now attempt every UPL across a split highway regardless of the location. And the reliability will improve over time across all locations equally.
Nobody has no biases, but mine are not strong, and they are not biases, which are preconceptions, but rather judgments.
Many don't understand the difference between "doing a successful drive without mistakes" and "driving for a whole human lifetime without mistakes. Waymo has about 20 human lifetimes, with a couple mistakes. Perhaps there is something hard about Silicon Valley but I, a single individual, can't get FSD to complete a single day of driving without some serious intervention being needed. And I live in the most Tesla-dense area, where most of the team is, where the former HQ is.
People don't comprehend the immense gulf this is. I will agree with you in one way. If Tesla worked on training their networks on just one town, their networks would effectively contain a map, so they would improve and get rid of the frequent errors they make because they lack maps. So it was wrong to say they would be exactly the same. However, it is unlikely that one could encapsulate all the data of a map of the world in the neural network.
What would be the same would be their perception and their path planner. It is true that testing those in different types of roads and conditions is valuable. The question is, how much more complete is what Tesla sees in the USA to what Waymo sees in the many cities it tests in (not just Phoenix and San Francisco and Silicon Valley.) More complete yes, but with little evidence Tesla is learning from its miles than the other teams. Right now Tesla is at a very low quality state. At a low quality state, interventions and error reports come in far faster than any team could handle them. They probably get tens of thousands per day. That seems good, but not if you can't use them. The teams that are getting one serious issue per week are having teams look at each one.
I want everybody to succeed, including Tesla. I feel Tesla is taking the wrong course and it is slowing down their progress. I own a Tesla with the FSD package. I own both TSLA stock and GOOG. My bias is I want this to happen as fast as it can.
willow_hiller
Well-Known Member
I just honestly think it's way too early to compare progress between Tesla and Waymo.
It's like having a race between someone hand-assembling vehicles and someone who puts together an assembly line. The person hand-assembling cars will say "Ha! I've already got an entire road-legal vehicle assembled, and I'm already starting on my second one! I'm way ahead of you because you've only got 10,000 bodies made, and you're barely just getting the wheels attached!"
But the point of the race isn't to see who can make two vehicles first. It's to see who can make two million first.
That is true, but it doesn't mean you can't examine the progress. It has always been possible that Tesla has picked the better path. Their path depends on a computer vision breakthrough which nobody can predict the date of. It might happen next month. It might take years. Most people don't want to bet on that, and they know that LIDAR works, and is available now, and its price is dropping fast. Betting that digital/electronics tech will drop in price with time and volume is a very safe bet. Betting on a research breakthrough is a risky bet.
Many experts in the field will say they believe that vision only will work some day. But they don't think it's the way to get there first. There is more debate about whether more heavily machine learning approaches will be superior in the end -- and even those who think they will also wonder if they will get there first, to meet that goal of 2 million first. (It's not 2 million but it's not a dozen either.)
On the other hand, Tesla's decision against maps is just silly. A car that can create its map on the fly is a car that can build you maps for free if you just bother to remember what the cars learned. To not have a memory is to cripple yourself for no real gain.
Just about everybody questions Tesla's approach of using the cameras from a 2016 Model S. There is no reason to build a robocar in 2022 with anything but 2022 hardware. Tesla's only reason is they want to keep a promise that their software will work on those older cars. That is a limiting factor they voluntarily took on, but it will slow down their project.
The other leading teams all know what Tesla knows, and more. Other companies have better understandings of machine learning and neural nets. They have better processors. They have better sims. They have better hardware. Tesla has a larger fleet than all but MobilEye, I will give them that. Only if you think that the large testing fleet is the sole key edge would you be likely to bet on Tesla. It is useful, but Tesla is missing so many other edges that it's not credible to call them the leader, not when their software performs so horribly even with that fleet to exploit. If Tesla is the Tortoise, it should still have crossed the starting line.
willow_hiller
Well-Known Member
Not sure how closely you're following the development of 10.69.X, but they may have already achieved it.
I don't know any other vision-only systems that can recognize the outer-bounds of an arbitrary unknown obstacle (temporary outdoor dining set up during Covid) from 400 feet away: FSD Beta 10.69
The breakthrough would not be recognizing a thing -- though that is always good -- it's spotting all the things, all the time.
Last edited:
willow_hiller
Well-Known Member
I think it's much more likely that this is proof that the occupancy network can detect the size and shape of any arbitrary object in the road from a reasonable distance, than Tesla training their occupancy network on outdoor dining spaces constructed out of miscellaneous materials.
GlockWorkOrnge
'20 Model Y Performance
In the article you suggest Tesla pulls FSD for it's used cars and then charges for it again.
That is not the case.
I bought a used MYP and FSD was included.
In fact, it you watch the used car pages at Tesla you'll see they almost ALWAYS come with FSD (and it might be always. Every model I looked at came with it--every model, and I was shopping for anything other than a model 3, which I was trading in).
If an FSD equipped car ends up going back to Tesla for resale or thru an auction FSD will be removed.
Indeed (and I think this is new) most cars on Tesla used cars have FSD. In fact, too many of them, which suggests Tesla is adding it as a way to get more money from the cars during this period that new cars have a long wait.
This is different from the question of whether Tesla pays you more on your trade-in if it has FSD or not. That I don't know. Whether a car for sale used from Tesla has little to do with whether the car had FSD when it came in, other than perhaps the question of whether it had HW3 or not when it came in.
We would need more data to figure out what's going on.
willow_hiller
Well-Known Member
Prior to this year, Tesla did not pay extra for trade-ins with FSD. But there was some push-back on social media, so since the beginning of this year they've started valuing FSD at $6,000 (presumably 50%) for trade-ins.
It is good they are doing that, but a sign that they themselves realize the feature is worth far less than they charge for it in the open market. It's software, it doesn't depreciate (in fact in Tesla's view it keeps getting better) and right now the cars are not depreciating at all, they are selling for above retail even with lots of miles on them.
Because it seems you do get $3K to $4K extra in private resale markets for it, Tesla has to offer you something for it, or you would be silly to trade in to Tesla. Not that people don't do many silly trade-ins.
Since Tesla can add the feature to any car at no cost to them (if it has HW3) the only reason for them to pay $6K on trade-in is to make sure they get the trade-ins, and also to not be seen as devaluing the feature too much. The money paid to the seller is otherwise wasted as they can enable the function on sale. It does look like they are not doing the "Sell the car without FSD and make them pay $12K to enable it" approach, probably because take rate on it wasn't great.
Last edited:
Tam
Well-Known Member
Tesla removes fSD before a non-Tesla dealer gets the car. The press covered when that procedure went wrong as the next owner saw clearly that the FSD was not removed:
Tesla Remotely Removes Autopilot Features From Customer's Used Tesla Without Any Notice [Updated]
One of the less-considered side effects of car features moving from hardware to software is that important features and abilities of a car can now be removed without any actual contact with a given car. Where once de-contenting involved at least a screwdriver (or, if you were in a hurry, a...
jalopnik.com
Tam
Well-Known Member
The older the FSD gets, the resell value should get higher toward the $100,000 tweet estimate:
Tesla to hike FSD price but Musk says worth $100K
Tesla's Full Self Driving package will rise by $US1,000 come July 1, but Musk says it will be worth more than $100K in time.
thedriven.io
Similar threads
