JupiterMan
Active Member
the auto labeller is also getting trained. I guess there comes a point where for many clips manual labellers are no longer needed. If this is the case it is huge and potentially exponential in nature.
-
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
This August 2021 article calls out Karpathy saying somewhere over 1,000
How Tesla's Autopilot Team Refines An Unfathomable Amount Of Data Is Pretty Cool
Labeling in general probably doesn't need tight coupling. Training and testing data sets do.
This Week in Tesla: Gigafactory Buffalo is Hiring, Tesla Increased Their Workforce by 20,000
StealthP3D
Well-Known Member
It's not whether Telsa currently competes solely on price, apparently, right now they don't have to in order to take increasingly large chunks out of ICE sales. Consumers would not be paying more if the ICE car was comparable. We know ICE SUV's are inferior in just about every metric except for range. Even in that metric the Tesla wins if measured by the average range available each morning when you wake up.
Face it, the Tesla SUV's offer superior value or they would not be wiping out big swaths of the market when those ICE SUV's cost less. Car buyers are not stupid. Value does not always equal the lowest price, it's how much you get for your money. You will never convince me that in 2022 an ICE SUV is a better value for most people compared to an EV. Resale value on those ICE SUV's is really going to be in the dumpster, once the chip shortage has eased.
You can pay now, or you can pay later. There is no free lunch. ICE cars are over-priced crap. And don't be fooled by the archaic safety ratings. If ICE SUV's were as safe as you say, they would not be crashing, rolling over, catching fire, etc. and killing their occupants by the thousands every year. That could be your family. Tesla engineers go beyond a 5-star safety rating that is gamed by greedy and short-sighted car makers. Tesla has reams of real-world data to constantly improve the actual safety while other manufacturer's engineer to the synthetic safety ratings.
Maybe. But a little caution.This is extremely bullish.
Means autolabelling is solved to a large degree.
Means FSD is coming out of the beta phase soon.
Lots of moving parts on Tesla's org chart and losing a few people in labeling in one place doesn't mean it's necessarily declining. They might just be cutting low performers or shifting the jobs to another region.
davidlittle
Member
Tesla Update - Your Model Y collection is scheduled for 9:00 PM on Thursday 30 June at Horizon Cruise Terminal, Southampton, SO15 1AW. You must bring a valid driving licence and ensure appropriate insurance cover is in place. If trading in a vehicle, please bring the V5c and both keys, and remove all personal belongings from the vehicle before entering the terminal.
Arriving by car? Enter at Dock Gate 10 and turn right at the roundabout and follow the left-hand lane to Horizon Cruise Terminal. Arriving by train? The closest train station is Southampton Central Station - 12 mins by foot and 4 mins by taxi to Horizon Terminal.
We cannot guarantee the level of charge on your car. Your collection will be self-service, so download the Tesla app at ts.la/app to learn about your Model Y. Not ready for your Model Y or need help? Reply here.
Xepa777
Banned
Okay I tried not to get involved but I feel like people are speaking on different wavelengths. I, as a mediocre machine learning scientist, mediocre TSLA investor, and mediocre meme generator have a well balanced analysis of how Tesla is positioned in the autonomy space.
Where your friend is right:
Your friend is right that camera-based depth perception takes a lot more processing power. Yes the voxel depth resolution output currently is likely worse than what Waymo / Cruise are getting from lidar. But these are nothing new and no one should be surprised by this. I've wrote about this years ago on TMC, like here
Is current sensor suite enough to reach FSD? (out of main)
Tesla won’t have to wait for FSD to solve all weather conditions in everywhere in the world before rolling out TN and generate significant value. With current FSD computer they should be able to drive better than human under most conditions, that would enable TN in enough markets to consume...teslamotorsclub.com
Maybe Tesla has already used up all of HW 3.0 processing & memory, this is hard to tell because of course they are going to use as much as possible before optimization (induced demand)
He is right that sensor fusion gives higher accuracy than individual sensors (more on this later).
I would agree that its quite unlikely that Tesla will be able get to to say 10x safer than humans with current hardware.
Where you friend is wrong:
Everything else, and every conclusion he makes from those observations above. He basically discounts any of the data Tesla collects because he doesn't think the camera resolution is sufficient to be worth anything. This is inane and myopic.
The Big Picture
Tesla is focused on developing a profitable approach to autonomy
Tesla is starting with a L2 assist system that will probably soon make the entire Tesla AI team cash flow positive. Oh and they'll be able to sell excellent upgraded passive safety features (all Tesla's produced from 2019 onward will be able to detect pedestrians / cyclists / running a red light/ into a curb etc... soon). Let's assume the path to L4/L5 robotaxi autonomy takes 5 years, well in the meantime Tesla is still making money off their development with the best in class driver assist features.
Meanwhile, Waymo / Cruise were essentially forced from the beginning to rely on sensor fusion w/ LIDAR / RADAR. Sure it's more accurate, but they had to chose it because they rely on VC funding. And you aren't getting the next round of funding if you can't demo clear progress. That means focusing on the easy way to get something working (lots of expensive sensors and focusing on limited geographical areas). And still no clear path to profitability.
If Tesla need to add sensors, improve cameras, or improve processing power, they can do that later. What, excactly have they lost in the meantime?
Whose path do you think is most stable in the current moment?
Tesla is focused on developing a scalable approach to autonomy
By starting out with trying to make FSD work almost everywhere, Tesla is forcing itself to make a scalable solution. There won't need to be as much "going back to the drawing board" as there might be at Waymo / Cruise etc who barely have gotten things to work outside of one city. Of course a downside of that is going to be performance of the general algorithm will be much worse at first. The data shows most of the disengagements are currently due to mapping issues. Tesla clearly hasn't finished whatever their generalized mapping solution is, but I'm confident they have the diverse data to figure it out. It's hilarious to think people believe Tesla has flatlined progress when even that one issue clearly has a path to improve.
Meanwhile Waymo has "solved" perception, but nary a peep about global deployment? If it was scalable, they would be touting their scaling prowless, demo in 50 cities and IPO for a trillion dollar valuation. Oh, not happening?
Tesla is doing the best compromise between what a data scientist and an accountant would do.
If you told the data scientist you had unlimited funds to solve FSD, he/she would load 100,000 cars with cameras / radars, lidars and have them driving around every area of the U.S. or world, collecting the data and working to develop a robust, scalable solution. But that's not cost effective, so no one is doing that. Tesla is chosing to collect more diverse data while competitors are focused on collecting less diverse, but better resolution data.
By developing a "cheap" perception engine (cameras + AI), Tesla is allowing their AI team to move on to focus on how to solve all the other problems in full-scale autonomy. Planning, prediction, whatever else. There is no need to wait for centimeter level depth & perception precision in order to work on all the other areas of the tech stack. This is where your friend is totally wrong. And any good data scientist knows you need a wealth of diverse data in order to make a complicated algorithm generalize well. That's something that Tesla has and competitors absolutely do not.
So Tesla is going what a good data scientist would do given a constraint on funding. And guess what? Say that the camera only stack isn't good enough in a few years for robotaxi level precision? Is Tesla farked? Um no, they can simply add high-res radar or lidar in a few years and fuse the data then.
TLDR Tesla is making a cheap, scalable, and profitable autonomy software that they can upgrade with better sensors and hardware in a few years if they need to.
Appreciate the real reply dude
exactly what I'm looking for! And yah, it's a compromise, just a matter of whether this compromise will work. If not, it certainly can erode Tesla's lead.And local maximums, you don't know what you'll hit those. Much harder to find those with an expensive sensor suite given the difference in data quality. I'm rooting for Tesla to achieve FSD on its current sensor suite but we'll see if it's possible. I'm not smart enough to know how much of their training set they can bring over if they have to re-vamp to a completely new sensor suite, and have heard conflicting things both here and on non-Tesla communities.
For investors, it isn't really earth shattering if the current hardware can't handle it either. Yeah it's not ideal, but let's say Tesla had to pay for new cameras and a new computer for all cars with FSD sold. What's that cost, maybe $2k per car? The revenue recognition alone would offset that.Appreciate the real reply dude
And local maximums, you don't know what you'll hit those. Much harder to find those with an expensive sensor suite given the difference in data quality. I'm rooting for Tesla to achieve FSD on its current sensor suite but we'll see if it's possible. I'm not smart enough to know how much of their training set they can bring over if they have to re-vamp to a completely new sensor suite, and have heard conflicting things both here and on non-Tesla communities.
But more importantly, that would happen when FSD was truly complete. I don't see how the market would not brush off a few billion in expenses vs that enormous success.
bkp_duke
Well-Known Member
This is EXTREMELY SUBSTANTIAL for Starlink.
Basically, if DISH gets an FCC license for that 12GHz spectrum (to use on cell phone towers), Starlink in the USA is done for. One tower using this would interfere with Starlink reception for a 13 mile radius.
DISH has access to plenty of other spectrum bands to make their 5G cell phone business viable. I smell a rat, someone (*cough* viastat or other rocket companies *cough*) is funding this to trying to materially shut down Starlink.
Hyundai and Kia have been selling EVs for 6+ years. ??
Maybe author needs to do a little more research? Didn't read the source, but... why bother chasing garbage.
Nothing magically changed coincidentally at the same time when Elon is voicing concerns about the economy and his intent to lay off a portion of the workforce. Even if Autolabelling does greatly reduce human effort, the plan would obviously not be to cruise along at current manpower numbers and then suddenly cut hundreds of people and close an office.
We can argue that Tesla believes the reduced need for manpower is coming and it makes sense to cut the jobs in the current economic environment, but I think it's far fetched to suggest this is just caused by a sudden exponential improvement in Autolabelling capability and reduced need for humans.
You are likely to be disappointed then. From the linked blog post about the interview:
Tesla delivers first Model Y Long Range vehicles built at Giga Texas [Update]
Since the Model Ys launch in early 2020, all of the electric SUVs destined for the North American market have been made at Tesla’s Fremont factory in California. Today that has changed as the automaker […]
driveteslacanada.ca
In the "Elon Musk Part 3 Bonus Material" interview from a few weeks ago, he guessed that 1500 human labelers were now amplified 100x (potentially even 1000x) with auto-labeling. For example, maybe auto-labeling amplification quality previously was "only" 50x needing 1500 humans for an equivalent of 75k labelers, and a doubling of quality to 100x while reducing headcount to say 1000 is still an increased 100k equivalent labelers.
Also, I believe Buffalo has headcount requirements probably part of the reason why a labeling team was set up there even though it's further from Autopilot engineering. This could be a shift in headcount location while also saving on costs by closing the San Mateo facility (while shifting some labelers to Palo Alto to be even closer to engineering).
Why do you think this? Just about every Tesla ever made is a robust data collection capable vehicle.
Current Mobileye equipped vehicles are not capable of collecting, storing or distributing/sending data outside the vehicle.
Not only did I work on labeling at Tesla (I set up the first team), but also during my time at Google Maps and yes, they need tight coupling. Quick pivots are key in labeling as is mass updating when code changes are made to look at new features in new ways.
Filled it out as Starlink is like water for my house and especially when I get my first CT which I'll be taking to just about every cool 4x4 site and camping.
Like this one I did in a heavily modified Jeep (yes, attack angles and managing acceleration is key to not flipping which I saw a Land Rover do)
Nailed it! This is huge and is most likely, IMO, a strong signal that they are now fully creating ground truth from auto-labeling (which is most likely a set of trained AI models predicting, marking and classifying semantic objects) at full scale for Dojo training. I'm not going to say that they've solved unified vector space yet or single stack, but wow, this seems like a huge milestone.
I think you have no idea what you are talking about. Can't sugar coat it more than that.
Edit: I'm in anti-FUD mode today
Had a great run this morning!Similar threads
- Locked
- Locked
- Locked
