Sorry dude. You’re underestimating what it would take to retrofit VW Golf’s “from years ago” to turn into full self driving robo taxi’s. And everything that your trying to imply w/r/t your multiple comments on the matter are plain fantasy.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
I don't think that was the FSD demo. his tweet says that's whats in Tesla cars now and it was for a ride back to the airport. So I think that that probably is just a video of NOA
Causalien
Prime 8 ball Oracle
Ya, they mentioned they are creating test units as they add them. Don't know how much time that will take though for full retraining. If it is 3x better, potentially 3x the learning curve as there are more neurons?
Lots of assumption at this point.
Very happy with Tesla's Autonomy update.
Lots of great details on the hardware specs, neural nets, training and autonomy strategy.
Other than these details, was particularly interested in:
Lots of great details on the hardware specs, neural nets, training and autonomy strategy.
Other than these details, was particularly interested in:
- Project Dojo - Sounds like an end to end neural net self driving solution in development.
- New targets for safe enough to sleep in 2Q20 and Robotaxi regulatory approval in 2020.
- New battery pack due in 2020 targeted to last 1 million miles.
Last edited:
Anstandswauwau
Member
comma.ai begs to differ. Still ahead of pretty much anything BMW, MB and co. has to offer.
I think it’s probably important to address the idea of retrofitting FSD here. I work in Software and have worked specifically in computer vision(more inference side than training side) for the last 6 years.
Designing a retrofittable system isn’t as easy as some here seem to believe. The way networks learn often doesn’t work the way a person would expect. As Karpathy pointed out during the presentation, the network sees a bunch of numbers and uses those numbers to build up belief about the world. It will use any hint or shortcut that consistently appears in the data to get to the right answer. (A little anecdote: when Perceptron networks were first developed, there was what appeared to be an extremely impressive demo showing the ability to detect military tanks even when partially/mostly occluded by trees. But the network didn’t work at all when tested on other data. The reason? The pictures with tanks were taken on a cloudy day and the pictures without were taken on a clear day. The network learned that there are tanks when it’s cloudy and nothing else.)
If all/most of your training data comes from one vantage point and you try to run the network from a different vantage point, your results will be nonsense, since the network learned about the world only from that vantage point. You can get around that by either restricting inference to that single or small set of vantage points you trained on(what Tesla does) or train from many different ones to force the network to learn more generally. But if you choose the latter, all those corner cases Karpathy mentioned need to be represented by many of those sensor configurations.
The data requirements for a licensable solution explode, requiring many different Tesla-sized fleets all collecting data. It might be done at some point in the future, but I don’t see it happening in the short/medium term.
Designing a retrofittable system isn’t as easy as some here seem to believe. The way networks learn often doesn’t work the way a person would expect. As Karpathy pointed out during the presentation, the network sees a bunch of numbers and uses those numbers to build up belief about the world. It will use any hint or shortcut that consistently appears in the data to get to the right answer. (A little anecdote: when Perceptron networks were first developed, there was what appeared to be an extremely impressive demo showing the ability to detect military tanks even when partially/mostly occluded by trees. But the network didn’t work at all when tested on other data. The reason? The pictures with tanks were taken on a cloudy day and the pictures without were taken on a clear day. The network learned that there are tanks when it’s cloudy and nothing else.)
If all/most of your training data comes from one vantage point and you try to run the network from a different vantage point, your results will be nonsense, since the network learned about the world only from that vantage point. You can get around that by either restricting inference to that single or small set of vantage points you trained on(what Tesla does) or train from many different ones to force the network to learn more generally. But if you choose the latter, all those corner cases Karpathy mentioned need to be represented by many of those sensor configurations.
The data requirements for a licensable solution explode, requiring many different Tesla-sized fleets all collecting data. It might be done at some point in the future, but I don’t see it happening in the short/medium term.
BMW was doing it decades ago. This was to make the cars lap in the same way as a pro racing driver. The controls are the hard part to iron out and calibrate, but the main job is done once, not for every example being retrofitted. I'll go out on a limp guessing neither of us have tried to install 8 sensors and if needed added controls into a 2015 car in the year 2022. I prefer to not underestimate the #2 to achieve Level 5. They will find a way to get as many cars doing rides asap. Someone about having only one goal and none of the limitations.
For Tesla to really deal the competition a blow, they'd offer their own computer at a fair price into the market to dissuade anyone from developing their own. Bring other BEVs into the TN, money is money, displaced pollution is just that, right?
Edited to add: I have good faith in 2022 Level 5 capable engineers to get a grasp on sensor placement variation and I am by no means suggesting a one size fits all solution. Analyze the fleet on the road, select a candidate and develop the retrofit kit for that particular vehicle.
Tesla is having to do it for multiple cars also now.
No, there’s no hands on the wheel at all, even during lane changes...
EDIT: He updated and said it’s NoA... I think it must be some unreleased version. Current version I have requires hands on wheel before making a lane change.
He did answer it, just not in the way the questioner asked. Tesla's approach doesn't get better or worse with geofencing. If it works at all, it will work pretty much everywhere.
Just a thought,once Starlink is activated, a few years down the road Tesla's could talk to each other and provide real time input about local data/situations that can be further process by the chip/NN network on the go. ...
bummed that we didn't get to see the demo. But, conference was great ..
bummed that we didn't get to see the demo. But, conference was great ..
There are different ways. Again I was not in auto dealer service. What I liked best was that every request for service outside scheduled maintenance was logged into the quality system as a complaint. Every complaint requires a letter from the QA director to the customer specifying the resolution of the immediate complaint as well as a statistical analysis of the frequency of this failure mode. All this is in addition to the regular service event communications.
We sometimes got criticized for too much communications but there was little doubt that we were committed to quality.
Our customer satisfaction survey was designed with in conjunction with an outside company that did human being phone interviews and/or paper surveys worldwide. It was all done anonymously and covered all aspects of customer interaction from Sales and service all the way to the products themselves. Direct quotes from customers were included. Data was trended over years and presented directly to the board and then to the department heads from there.
Periodically we would present an overview of customer satisfaction feedback at national conventions/meetings where customers could attend.
A system like this is a bit painful to institute but it makes it very hard to ignore a recurring problem area.
Not a demo. That's what everyone has already... still looking for a real demo from today's rides.
Nio too got a car on fire just a couple hours after the Tesla in Shanghai :
Electric Car Fires in China Should Set Off Alarms
Electric Car Fires in China Should Set Off Alarms
Once all cars are autonomous and sharing information with each other, this and similar concerns should evaporate.
SpudLime
Active Member
powertoold
Active Member
I don't think lidar is useless. It would help with defining the "drivable space."
Todd Burch
14-Year Member
Go ahead. Wikipedia entry for photogrammetry probably has this info though.I said I'm a FSD skeptic, but I'm firmly against anyone that says "LIDAR IS NECESSARY FOR FSD." Do you mind if I share this on Twitter? Or if you're on there, share it yourself and I'll retweet
Causalien
Prime 8 ball Oracle
Something that most people missed. Oppenheimer asking the relevant question.
"Can you talk a little bit about what that look like, what your expectations are in terms of financing for the next 3 4 years to build up this fleet.
"blah blah cash flow neutral"
"We will make the right move. I think we will make the moves you think we should make."
So capital raise to build out the robotaxi fleet. Maybe end of 2020 is my prediction.
"Can you talk a little bit about what that look like, what your expectations are in terms of financing for the next 3 4 years to build up this fleet.
"blah blah cash flow neutral"
"We will make the right move. I think we will make the moves you think we should make."
So capital raise to build out the robotaxi fleet. Maybe end of 2020 is my prediction.
Similar threads
- Locked
- Locked
