Discussion keeps the world turning.
You're tuned in to Roundtable.
I'm Steve Hatherly today with Yushan and Yushun.
Coming up, you've heard of gig work, but this is different.
No apps, no ratings, just a camera strapped to your head while you fold laundry or care for your kids or work on a sewing machine your life recorded, sold to the biggest tech companies on earth.
Today we'll talk about this the invisible workforce training, the machines that could replace them and the trade we may not realize we are making.
Our podcast listeners can find us at roundtable china on apple podcast.
While you are there, if it's not too much trouble, you could leave us a five star review.
We'd very much appreciate that.
And don't forget, you can contact us too.
Roundtable podcast at qqcom.
You can send your voice notes.
Roundtable podcast at qqcom.
What would you trade for a free house cleaning?
Sounds like a good deal, until the cleaner shows up wearing a camera on their head.
Every room, every dish, every corner of your home becomes data.
And that is just the beginning.
From factory floors in India to nursing homes here in China, thousands of workers are recording their everyday lives for tech giants, building the robots of tomorrow.
Today we're going to talk about this, the hidden world of what they call ghost work.
We'll find out what it is, who's doing it, and what happens when your life becomes a data factory.
All right, so we'll start our story today.
You should welcome to the show, by the way.
Good to see you today, too.
We'll start with a house cleaning service that's happening in various parts of the world.
Yes, this company that's making headlines is an AI training startup.
Actually, it's called Shift.
It's owned by the German data research lab Micro AGI, and it's unique in the sense that it's just as much a data collecting process as it is an actual cleaning service.
It's been occurring in the US, as well as in Germany, Turkey and other European countries for months.
So how does this work?
What are they doing?
They collect data, basically.
That's what they do.
But the thing is that these devices capture egocentric video.
That's a first-person perspective, that's that shows exactly how a person's hands interact with the objects in a home environment.
They call it egocentric?
Yes, the cleaners wear camera headsets and camera caps or hat-mounted devices to record their whole cleaning process.
So this process is free for the customer.
So the customer does not pay with money, but they pay, I suppose, with their data, by allowing the cleaning process to be filmed.
The cleaners, yeah.
Some sort of device is worn on their heads usually with a camera, and it's pretty simple.
Some of the people who did it said you just put the machine directly on your head and then you just work as you normally do.
You don't really do anything different.
So the reason why these data is being collected is that AI systems need a first person view on human behavior to understand physical tasks accurately.
So from an external angle, you can see what happens.
Like, for example, the cleaner wearing a headset of camera and recording how human hands grab a sponge or grab a mop.
And then those kind of image and data they transfer back to the AI system so that in the future a humanoid robots could be grabbing the same sponge and grabbing the same mop with the exact same gestures.
So it's not just a video of a room then.
It's about the behavioral... It's about very detailed behavioral data.
Yeah, it's about the ability to actually identify, recognize and actually do these kind of what we call chores.
It captures how a human hand grips a sponge, how water is turned on at different faucets and how dishes are stacked, how objects are moved around in real time and how tasks change depending on different home layouts.
This is the data that we say that is very precious and urgently need in the industry of these humanoid robots.
Because imagine the first person point of view, because you kind of need these kind of data instead of these third person point of view.
If you give or feed these robots these kind of videos, they need to analyze.
Okay, this is me and I need to do like this.
That's another step.
But if you gave them the firsthand or the first person perspective, they can analyze exactly what this video is about and they can actually learn from it.
This is how it looks like.
Yes.
And there's so much variation from home to home, right?
Not every home is the same.
So you're going to get different lighting.
You're going to get different size spaces.
You're going to have differently designed kitchens, for example, or bathrooms, for example.
And also the placement of objects. is going to be different from home to home to home to home.
Your microwave isn't always going to be in the same location in every single house, right?
So all of that data becomes training material for the AI systems, especially robots yes, that need to learn how to operate in these types of real-world physical environments.
So it's not just learning what a kitchen looks like.
It's learning how humans actually interact with a kitchen.
But that first example was just about the cleaning service that's happening in some parts of the world.
But it goes beyond that.
And it's happening here in China, too.
Definitely.
China is a player in the game.
And the big leader in that realm would be JDcom, who has launched a massive infrastructure project aiming to mobilize 600000 people over two years to collect 10 million hours of video data in the specific environments needed.
Such as well, for example, in nursing homes, where they expect to record details of caring for the elderly.
And also in farms, that's to record the physical movement of picking fruits and handling agricultural tools.
And also in garment factories, filming the specific actions of workers operating sewing machines.
So in these places, very detailed, very practical.
And it's happening in Suqian in East China's Jiangsu province already.
And funnily, I just learned this morning that Suqian is also where the hometown of JD's big boss is at.
Oh, really?
So it's probably he's got a big influence there and it's easier to mobilize a lot of the residents to work.
And indeed, the residents there are trained to record specialized tasks.
And these home collectors, they call it, typically earn between 3,000 and 3,500 yuan.
And that's around 410 to 480 U.S. dollars per month.
This is a way from this kind of perspective.
I think this is really a good way to create job opportunities for a lot of people in Suqian.
Of course, it is a city in Jiangsu province, but relatively, it is not that economically developed.
But I'm not saying that they are poor.
But But Jiangsu is a large area that incorporates different regions and every region has its own development speed.
And Suzhen, yeah, like you said, in this chance, it's a good way for its people to get a job.
Yeah, one of the advantages, for sure.
There's a story of one full-time mom who is employed in this fashion and she performs her household chores.
She picks vegetables in the garden.
She washes dishes.
She mops the floor.
She's wearing a head-mounted camera device, just like that cleaning service we talked about a few moments ago.
But she also records herself taking care of her children while wearing the equipment.
And perhaps that's a point we can come back to in just a minute.
Yeah, that's the China example.
There are plenty of examples.
There's one in India.
This is more of an industrial application, though.
Workers in India are working in clothing factories and they wear ring-shaped cameras on their heads.
Again, it's another device that goes on top of your head.
And they do that during their 12-hour shifts.
And that data is sold to major tech companies like Tesla, also Boston Dynamics.
Yeah, if you're curious to know why, Boston Dynamics, that's a robotics company.
So that kind of makes sense why they would want that data.
But Tesla, when I first saw Tesla, I thought, why?
Why?
But Tesla is also in the robot game, too.
Yeah, exactly.
Actually, this is their next big ambition for their whole future.
Actually, Tesla has discontinued some of the production of Model S and Model X They're their cars, their EV cars, in some markets and has also adjusted parts of its factory and project plans, converting some production capacity to actually support its robotics manufacturing lines.
There are robotics called Optimus, and they are planning to deploy these robots to their car making or other factory lines.
So this is kind of part of Tesla's broader strategic shift, with a greater focus on humanoid robots and autonomous driving technology.
And those 12 hour shifts are like a gold mine for data collection because if you compare that to the house cleaning service, that might be three or four hours worth of data, but this is a continuous 12 hour shift.
That allows the data collectors to train their AI models, their robots, on recognizing objects like fabric or zippers or buttons.
Also understanding hand-eye coordination in an unstructured environment or in a messy environment.
By messy, I just mean like where things are here and there and here and there and the robot would have to coordinate its way around there, right?
So yes, lots and lots and lots of data to be collected there.
I mean, there are so many different examples.
MicroOne is another one.
They're a data collecting company where they employ about 4000 gig workers in 71 different countries around the world.
These workers.
They make about 15 an hour and they record themselves performing common domestic tasks like folding laundry, making beds, taking out the trash in their own homes.
That's quite good, $15 an hour.
Now, I wanted to talk about that, and let's save that for a little bit later.
But yeah, that's the pay for that particular company.
But beyond domestic cleaning, there's other types of human activity that are being packetized.
Is that the word or data collected and then sold off?
Yeah, people are not leaving any chances alone just to teach AI or teach the machines to think like human, including not only footages and filming things at this point, even recording sounds.
The ambience, that's part of the AI training too, to my surprise.
That's in India, where students visit hotel lobbies to record environment sounds, earning over 100 US dollars a month just because of that.
Because recordings of everyday environments are used to train AI in sound recognition and acoustic modeling.
So this data teaches AI to understand and Navigate real-world audio environments, such as in the restaurant, where there's the clicking of plates and your knife and fork, and then hotel lobbies, the rolling of the wheels from your luggages, and city streets, the honking of cars, for example.
So that's all useful for developing smarter voice assistants, hearing aids and audio-based AI models.
So as post-producers, we are so familiar with ambience like that.
But now it's not only the humans who need it for putting that together as a part of a production, but for AI too.
Yeah, and they even want your conversations.
They want your audios of talking to people.
In Chicago, workers sell recordings of their phone calls with friends.
Yeah.
For 50 cents per hour per minute.
50 cents a minute.
Yeah.
Again, depending on how long you're talking.
Not bad.
So this is called ghost work.
This type of employment is called ghost work.
And to me, I understand the attraction for people to get involved with this because.
A, you choose when you work.
B, this is the point you wanted to make, Yusha, and you get paid pretty well in some instances.
You know, $15 an hour while you clean your house is pretty reasonable.
It is technically a bonus because you're doing your thing.
Yes, exactly.
But, you know, we've got these, and this is the but, right?
This is where we transition to the, am I comfortable with this part of the conversation?
Because we've got these big tech companies hiring thousands of people.
We just gave you a laundry list of examples.
Mm-hmm.
Privacy is definitely one of the things that we're tackling on.
Engaging in these kind of work often means sacrificing your intimate and also personal data.
In domestic settings, recording devices, capture how people move and also live in your private spaces, and creating data that could be used for various things, like personalized pricing, for example.
How?
Why personalized pricing?
That is more like the more data you offer to these kind of companies, they are getting into a mode of okay, you are a like return customer and we're giving you maybe less money per minute or per hour.
Yeah, the same as the.
You know some of the e-commerce platforms.
When you frequently buy on this certain platform and maybe when you're comparing to your friend um On your platform, you can get a higher price.
Yeah, I mean, when you record audio or video inside your house as part of this ghost work, you're not just capturing different types of isolated sounds or movements.
You're creating a detailed behavioral profile. of yourself.
So AI can analyze that data to infer things like how many people live in your house, or what time do you wake up, or how often do you cook, or How often do you order food.
What brands do you use?
What are your stress levels, even your health status?
So companies.
If that data ends up in the hands of a company, then they can use those traits to adjust their prices specifically for you.
Well, that is one thing.
The other thing, like I mentioned earlier, more and more people are willing to receive the money while teaching AI how to think and behave like us.
What's the problem with that?
I mean, isn't that teaching them the machines to replace us one day?
You are effectively teaching your own replacements.
Indeed.
That's what people are afraid of nowadays.
We talk about it all the time, how AI has the potential of replacing us.
And then we see day-to-day chores being done by normal day-to-day people.
Who is actively transmitting data back to the AI system just for that purpose?
Do you see it that way, Yusheng, as training your eventual replacement?
We kind of always have this kind of like questions when we are having maybe robots or AI in our lives.
Still, I'm playing the optimistic role in this conversation because...
First of all, in this scenario, I can't understand why people feel uneasy.
Recording everything in your home can feel like your privacy is completely exposed.
But to me it's also similar to what happened when digital payment or fingerprint authentication or facial recognition that has been widely used in in our like smartphones, or they are becoming the mainstream of a lot of things right, so take face id as an example.
Our phones are constantly using the camera to recognize us or recognize who you are when you try to unlock them.
Yet over time, most of us have become comfortable with that.
I think a big reason is that we trust the companies holding our data to keep it secure.
Of course, whether they actually do something, we can never know this completely certainty.
At least so far, though, the Pandora's box hasn't truly been opened.
That's another thing that I feel really scary.
It feels like we are the grass that's constantly being mown by these companies.
There's always real-time scenario and practical things that they can record and get from us.
Yeah.
It's like we are the sources.
Yeah.
I mean, data collection is one thing.
We've been talking about this for years, though.
But the scariest perhaps thing for workers is something called.
Have you ever heard of the Amazon warehouse effect?
Yes.
So this is what experts warn about, that the data collected...
From you can be used to create software that can have an impact on human workers in new ways.
So here's how this works.
So imagine you're a worker in.
And by the way, it doesn't have to be Amazon.
It's just what the Internet has kind of put as a title on this.
But imagine you're a worker in a fulfillment center, an Amazon fulfillment center as a person who's a picker,
So your job is to walk through the aisles and then grab items from shelves and then place them onto totes that go down a conveyor belt.
You wear a scanner on your finger or you carry a handheld device that tracks every single action meaning, how many seconds does it take you to walk to a bin?
How long do you spend locating an item?
How quickly do you scan it?
How fast can you drop it into that tote?
Now, this is where the warehouse effect kicks in.
The company doesn't just use the data to monitor you.
They use it to algorithmically set your performance targets.
So the system, they'll analyze not just you.
They'll analyze thousands of workers' performances and movements and then calculate the quote-unquote ideal time.
So let's say that the ideal time to pick an item is nine seconds.
But that ideal is based on the fastest 10% of workers, not the average worker.
So if you next week fall below that nine second target, guess what?
You're going to get an automated warning.
If you fall short too often over the course of a few weeks, You might lose your job.
Do you see what I mean?
Humans through the performance that they're doing in this data collection through the software, not just the robots themselves doing the jobs, but in the expectations of human performance on the job in the future.
That is a bit terrifying to me.
That sounds way too familiar.
We mentioned previously in one of your soapbox right Steve, where one of the burger chains I don't remember the name had an AI assistant called Patty.
It's recording everybody's politeness in their restaurant to use that data as a way of leveraging or standardizing their staff behavior.
I think that's another scenario then, when they're not only using data to train the robots, but also using it to standardize the work efficiency of these workers.
I don't think that is the scenario we should go to or the companies should use the data for.
Yeah, but that's potentially what they are going to use it for.
Those are the downsides, but I just wanted to take a few minutes to look at the things that we should be wary of.
But it's not all doom and gloom.
There are plenty of positives with what's happening now in data collection, too.
Yeah.
Yeah, the same technology that can exploit workers can also be used to empower them.
We're already seeing glimpses of a better path.
Some companies are using wearable data not to punish people but to prevent injuries, alerting workers when they're lifting ladders awkwardly or reaching dangerous fatigue levels, just to tell you that nope, it's time to stop.
Others are using AI to rotate tasks, so no one does the same repetitive motion for like 12 hours straight.
So something like that is protective.
It's a layer of insurance on the very people they're trying to mimic, using it for good and bad.
Yushin, are you comfortable inviting a cleaner into your home for free?
For free, but they're wearing a camera and they're recording everything they see inside your house.
Yeah, as long as I need to remind myself, you know, dress up well and do everything.
I know what people are going to do.
People are going to clean their house before the cleaner comes into their house.
I think people do that anyway, regardless of whether they have a camera on their head or not.
Fascinating story.
All right.
Well, that is it for today's roundtable.
Thank you very much for being a part of our show today.
Yushan Yushan, thank you to you.
I'm Steve Hatherly.
Have a good one.
And please do come back again next time.
