Tesla Lays Off Over 200 Autopilot Data-Labelling Employees

Today, Bloomberg reported that Tesla just laid off around 200 data-labelling employees from the Autopilot team and closed a branch in San Mateo, California. This lines up with Tesla's goal to make all autopilot data auto-labelled, and hopefully this shows that significant progress has been made.

This is promising news for FSD's Neural Networks being trained to label data.

Andrej Karpathy, Director of AI at Tesla, stated last year that around Tesla hired 1,500 employees to auto-label data, making this about a 13% cut in employees in that sector.

In addition, this past month, Elon Musk stated that he would cut 10% of salaried jobs at Tesla, while many in the data-labelling team were reported as hourly workers. Regardless, Tesla appears to be making appropriate cuts to make their workforce as efficient as possible.

(Featured Image Courtesy of Tesla, Inc)

Source: Bloomberg, Drive Tesla Canada

Bloomberg - Are you a robot?

 

[PerBB]

Today, Bloomberg reported that Tesla just laid off around 200 data-labelling employees from the Autopilot team and closed a branch in San Mateo, California. This lines up with Tesla's goal to make all autopilot data auto-labelled, and hopefully this shows that significant progress has been made.

This is promising news for FSD's Neural Networks being trained to label data.

Andrej Karpathy, Director of AI at Tesla, stated last year that around Tesla hired 1,500 employees to auto-label data, making this about a 13% cut in employees in that sector.

In addition, this past month, Elon Musk stated that he would cut 10% of salaried jobs at Tesla, while many in the data-labelling team were reported as hourly workers. Regardless, Tesla appears to be making appropriate cuts to make their workforce as efficient as possible.

View attachment 822466
(Featured Image Courtesy of Tesla, Inc)

Source: Bloomberg, Drive Tesla Canada

Bloomberg - Are you a robot?

Well I can for sure say, that whatever this means, no progress on Autopilot functionality can be detected. And Musk-promised autopilot update in march?

 

[joe.smith]

Well I can for sure say, that whatever this means, no progress on Autopilot functionality can be detected. And Musk-promised autopilot update in march?

Second that. I haven't seen significant progress in Autopilot in quite some time.

 

[Kognos]

I don't know what country Elon really lives in, but to me the labeling is bad to the level of unusable. Practical, real-world, L5 ready labeling would include emergency vehicles, line markings, school zones, crosswalks, sign reading, languages, flags for construction zones, etc. Very little of that is in place, and would be required for a human to pass a DMV written test, not to mention the actual drive. Baseline stuff isn't in place at all.

I realize I'm not part of that team at Tesla, but the DMV tests aren't really different state-to-state, so really -- what's the wait? Does Tesla not care about school zones, for pete's sake? Some of this is just nuts.

What is in place is impressive, but a glorified party trick for what needs to be in place for anything beyond fully-aware human oversight.

This is why I'm fundamentally mad to pay for something that is really a glorified beta test that adds stress and risk to my driving experience.

 

[EVEnergy]

They laid off people who do labeling - a job that the dojo computers should be able to do. It looks like they are just automating the process which should yield faster improvements in FSD since the computer can do it faster than a human.

 

[Tam]

They laid off people who do labeling - a job that the dojo computers should be able to do. It looks like they are just automating the process which should yield faster improvements in FSD since the computer can do it faster than a human.

A deletion doesn't necesarily mean advancement.

Elon Musk didn't want a steering wheel on 2020 Model Y but that didn’t mean the improvements were any better to let that happened.

Elon Musk didn't want Tesla's Model Y to have a steering wheel so engineers developed the car behind his back, according to a new book

Then-Tesla engineering lead Doug Field didn't want Musk to intervene with Model Y development, according to journalist Tim Higgins' new book.

www.businessinsider.com

Same with the deletion of radar: It's an advancement for bugs, software recalls, NHTSA phantom brakes investigation...

 

[mrbulk]

Together with their new 4D radar patent plus reintroducing Enhanced AP, does anyone else see a potential connect-the-dots situation where FSD is now prioritized far behind doing a full-on robo taxi product first?

 

[E90alex]

Together with their new 4D radar patent plus reintroducing Enhanced AP, does anyone else see a potential connect-the-dots situation where FSD is now prioritized far behind doing a full-on robo taxi product first?

A robotaxi cannot exist without mature FSD software…

 

[ItsNotAboutTheMoney]

It's just part of the cost-cutting they need to deal with the China lockdown supply chain disruptions, the slow ramp of the 4680 and associated technology, and resulting effect on ramps of the factories in Austin and Berlin, as well as potential reduction in sales from the recession.

Right now is the time to find the least damaging cuts. Giving they don't even yet have combined stack, and that they're working on auto-labeling, cutting labeling is a reasonable one. They can always use some cheap labeling in Africa or other places to replace them later.

 

[abasile]

The need for human labelers isn’t going away anytime soon, but it may be reduced. By all accounts, letting go of 13% of these workers is not a huge deal. (In the Bay Area, they’ll have no trouble finding other jobs.)

Also, my experience is that Autopilot is improving. I am seeing fewer phantom braking events, automatic lane transitions are better, and the system does impressively well in conditions where the lighting is challenging and the lane markings are difficult to follow. FSD Beta is improving as well.

 

[Big Dog]

Andrej Karpathy, Director of AI at Tesla, stated last year that around Tesla hired 1,500 employees to auto-label data, making this about a 13% cut in employees in that sector.

A 13% cut is not necessarily true (but could be). Karparthy could have hired 100+ more employees since his statement from "last year" so the cut might not be much of one at all. Plus, these folks could be rehired in Austin in the future.

 

[JohnB007]

I think I read somewhere that there's more people working on this based in Buffalo office....

As usual, I forgot where...

 

[NoGasNoBrakes]

A 13% cut is not necessarily true (but could be). Karparthy could have hired 100+ more employees since his statement from "last year" so the cut might not be much of one at all. Plus, these folks could be rehired in Austin in the future.

True, I think this job in particular is pretty expendable so they quite possibly could have hired more autopilot-labeling employees since.

Right now is the time to find the least damaging cuts. Giving they don't even yet have combined stack, and that they're working on auto-labeling, cutting labeling is a reasonable one.

Also true, pretty expendable so cutting some of their labeling team definitely is the least damaging option. I do hope that’s partially because of improvements with auto-labeling.

I don't know what country Elon really lives in, but to me the labeling is bad to the level of unusable. Practical, real-world, L5 ready labeling would include emergency vehicles, line markings, school zones, crosswalks, sign reading, languages, flags for construction zones, etc. Very little of that is in place, and would be required for a human to pass a DMV written test, not to mention the actual drive. Baseline stuff isn't in place at all.

I realize I'm not part of that team at Tesla, but the DMV tests aren't really different state-to-state, so really -- what's the wait? Does Tesla not care about school zones, for pete's sake? Some of this is just nuts.

What is in place is impressive, but a glorified party trick for what needs to be in place for anything beyond fully-aware human oversight.

Assuming that their auto-labeling computers aren’t taking things like school zones into account yet, yeah. But I’m not claiming that FSD is going to be ready soon, because I highly doubt that. As you said, baseline stuff isn’t yet in place, but I think with more accurate auto-labeling (partially in thanks to all of the manual data labeling that’s already been done), their Neural Nets will certainly have a quicker time learning how to identify all of these things, which will be a step towards true full self-driving.

Tesla certainly isn’t claiming you can go to a DMV and FSD will pass a driving test, but perhaps in 5-10 years, it will.

 

[alangold]

“I just got laid off from Tesla without warning, and my life is uprooted” I just read an article on a Tesla recruiter laid off - no notice - no help finding a new position - Mr musk treats his employees like they are “robots”

 

[alangold]

“I just got laid off from Tesla without warning, and my life is uprooted” I just read an article on a Tesla recruiter laid off - no notice - no help finding a new position - Mr musk treats his employees like they are “robot”

 

[Tam]

...no notice...

Since Tesla is sued for the lack of 60 day layoff notice, Tesla is now paying laid-off workers for 60 days from the date of notice.

 

[sleepydoc]

So exactly what do these ‘data labelers’ do? (I also find it a bit contradictory that they hire people to do ‘automated data labeling.’ Isn’t the whole point of automation to remove humans?

This is a story that makes a good headline But without more information or context you can’t say much. Sure you can make statments like “Tesla is gutting their FSD team!” but it all depends. How many of these data labelers do they still have? Exactly what are they doing? How many do they need? Why did they lay them off - was it to cut costs or because they were no longer needed? How important is data labeling to the FSD project?

“I just got laid off from Tesla without warning, and my life is uprooted” I just read an article on a Tesla recruiter laid off - no notice - no help finding a new position - Mr musk treats his employees like they are “robots”

This is how employment works. Assuming Tesla followed employment laws as mentioned by @Tam, they’re not under any obligation to help with a job search or give them warning that they’re Going to be laid off. Yes, it sucks, but plenty of corporations downsize and reorganize every day. That’s just the reality.

 

[NoGasNoBrakes]

So exactly what do these ‘data labelers’ do? (I also find it a bit contradictory that they hire people to do ‘automated data labeling.’ Isn’t the whole point of automation to remove humans?

This is a story that makes a good headline But without more information or context you can’t say much. Sure you can make statments like “Tesla is gutting their FSD team!” but it all depends. How many of these data labelers do they still have? Exactly what are they doing? How many do they need? Why did they lay them off - was it to cut costs or because they were no longer needed? How important is data labeling to the FSD project?

I think Tesla’s AI Day 2021 answers a lot of these questions, but in short, data labeling is currently being partially done by humans. In order to have Tesla’s Neural Net start labelling all data, it must have existing, previously labeled data that it can refer to. Hence where data-labelers come in. They manually label autopilot data.

From the last time Tesla provided info on how many of these people there are, they probably cut about 10% of their data labelling team. Exactly what they do is, at least I can assume, label photos and videos taken by a Tesla’s cameras on autopilot and probably deal with collision/disengagement reports as well.

We don’t know if it’s to cut costs or because they’re not needed. I’d assume somewhere in-between, because at least in my opinion, they are the most expendable and their job can easily be replaced, compared those like engineers or factory line workers.

Data labelling is pretty vital to FSD, hence why Tesla Dojo made an extremely high-powered computer just so it could run millions of instances of driving simulations. And the only way those simulations can be accurate and useful is if everything in them is appropriately labeled. For example, the computer should know the difference between a human and a garbage bin, or a large truck and a sedan.

 

[doc5339]

Together with their new 4D radar patent plus reintroducing Enhanced AP, does anyone else see a potential connect-the-dots situation where FSD is now prioritized far behind doing a full-on robo taxi product first?

No

 

[Tam]

...So exactly what do these ‘data labelers’ do? (I also find it a bit contradictory that they hire people to do ‘automated data labeling.’ Isn’t the whole point of automation to remove humans?...

It's the same principle as Tesla Driving Automation system: You just can't rely on the system to do its job. There must be humans in the loop to catch its mistakes.

On the 2021 AI day, Andrej Karpathy said there were 1,000.

Laid-off labelers stated that they had to work overtime for the past 2 weeks. Too many weeks with 16-hour days.

The purpose of laying off is to save money which might not have anything to do with there's no need for labor hours.

Higher-paid workers are let go and new postings invite lower-paid workers to join for the same job.

Tesla is still posting labeler jobs publically:

Jobs