I Trained Robots for Money

[00:00:01]Adjust your gear. This is it.

[00:00:03]>> Calibrating hands. My new job.

[00:00:06]>> Recording started.

[00:00:07]>> Training robots to do household tasks.

[00:00:11]Oh, it's real nice pressure on my forehead. Feels great. For the last few weeks, I've been strapping this extremely subtle contraption to my head.

[00:00:20]Yeah, I know. I look really good.

[00:00:21]>> Opening MicroAGI Shift app and recording my chores with the promise of earning up to $20 an hour. picking up some very warm dog poop. I don't want to talk about it. And other companies all over the world are doing the same kind of data collection. Also, we can teach the robots to do the same things we can do.

[00:00:41]How much data here do you need?

[00:00:43]>> At least like 20 million hours this year.

[00:00:45]>> It's all fun and games until you realize some people are doing this for money to help train the thing that could someday replace them. Let's talk about it.

[00:00:56]Today, Shift is launching free house cleaning services in New York City.

[00:01:00]>> Okay, so when I spotted this ad for Shift's free cleanings in New York, I was excited to have these tech bros clean my friend's apartment. The reality looked a bit different.

[00:01:11]>> Hi, good morning.

[00:01:12]>> Good morning. I'm Joanna.

[00:01:13]>> Here's how this works. You get a free apartment cleaning. MicroAGI, a German startup, pays the cleaning crew a flat rate. In exchange, the cleaners record video of the entire cleaning, which becomes training data for the robots.

[00:01:28]>> So, we're going to press start, and then we're in.

[00:01:30]>> Both cleaners wore Shift's signature camera hats, which connect to Android phones and record the entire cleaning from a firsterson perspective. For the next 3 hours, the women meticulously deep cleaned my friend's apartment, while I made a call to MicroAGI CEO to ask what exactly was happening here. I am in an apartment right now where we have two shift workers and they are cleaning. What will you do with the data they're collecting?

[00:01:58]>> The data specifically that you are collecting is used in pre-training of AI models.

[00:02:04]>> Translation pre-training is when an AI model learns general patterns from analyzing massive unlabeled data sets.

[00:02:12]In this case, lots of video of human hands doing stuff. So why do robots need this? to learn. Factories and warehouses are controlled environments. You can design them around robots. Homes are chaos. Take this unitry robot we met a few months ago. It couldn't do anything autonomously in my house or the 1X Neo I met last year that struggled to load the dishwasher. And that was with a human controlling it through a VR headset.

[00:02:40]See, humanoid robots have a data problem, a big one. Large language models like chat GBT have been created by swallowing well pretty much the entire internet. Robot models need their own version of that. Massive amounts of data showing the physical world. Some call this egocentric data. Microaggi collects all those videos, blurs any personally identifiable details, anonymizes the videos and AI analyzes the hand movements. Finally, all the accepted video is converted into code that can be used for training. In the case of what is happening here, they're cleaning the sink and they're cleaning the stove. Is this all to then train a robot to be able to clean the sink and clean the stove?

[00:03:27]>> Um, not directly. Your footage that has been recorded today is not necessarily being taught to do, you know, sink cleaning. It's only useful when you pair it with millions of other diverse examples where then something called generalization occurs and you can put it in anything.

[00:03:46]>> The goal here is bigger than just cleaning. It's a humanoid robot that could walk into a house it's never seen before and just know how to do things.

[00:03:55]At least that's the dream. And that's the dream that got me to wear this thing. Not to mention the chance to make hundreds of dollars doing chores I was already going to be doing. But well, yeah, you'll see. My chin is sweating in here.

[00:04:09]>> Recording is still running.

[00:04:11]>> Only 8% of what I have uploaded is useful.

[00:04:17]Adbreak and money mic time. Today, the money mic is sponsored by Mercury, our launch partner here at the new things and the main financial tool we use to run the business. So, no, I didn't make enough money training robots, but the bots are helping me manage our actual money. Mercury has a new AI helper called command. So, now I can command command to do things related to the business account, corporate cards, and more. Pay David back $15 for the sand which he bought me at lunch. Anything command wants to do, it asks for your approval first. No. AI is sending money without you clicking confirm. Beyond command, I just love how simple Mercury is. Managing accounts, corporate cards, invoices. It's all in one clean interface instead of digging through a maze of banking menus. Link in the description or go to mercury.com. Now, back to training the robots.

[00:05:14]>> Sometimes I like to talk to the robots.

[00:05:16]Robots, if you're listening, you want to really scrub the chocolate off the table.

[00:05:21]>> Shift doesn't just pay professionals. It encourages regular people to contribute data and pays individual contributors in the US $20 for every hour of accepted footage. Keyword accepted. Unlike the cleaning crews and other professional services that use the hat, Shift has individual contributors get this contraption. Inspired clearly by my personal style icon, Wayne Zalinski from Honey, I Shrunk the Kids.

[00:05:47]>> What's on your head?

[00:05:48]>> I was looking for the kids. So, for the last 3 weeks, I strapped it to my head and recorded myself doing chores around the house and even outside the house.

[00:05:57]But all I get are rejected. Rejections, rejections, rejections. Every uploaded clip is judged naturally by AI on three things: camera quality, hand visibility, and meaningful task. Hand visibility was my downfall. As the app kept telling me, hands were not visible enough. But when I reviewed the footage, my hands were in frame. Turns out it has to see both hands in frame for a continuous period of time, which explains why loading the dishwasher kept getting rejected while laundry and cooking were approved. And Lord knows why ski ball got a check mark. I can't aim with this thing in my face. I suck. But here's the other big thing. Even when your footage gets approved, you're only paid for the portion they decide is usable. Example, I recorded about 7 minutes of me folding laundry, but only 54% of it was usable.

[00:06:52]So, I only got paid for the three and a half usable minutes. For the first week, MicroAGI was nice. It paid me $20 an hour regardless of the quality of footage. But after, I only got paid for the usable footage. So, after 3 weeks, I recorded, let's see here, 3 hours and 51 minutes of footage, but only 57 minutes was deemed quality footage, or 24% of it was. So, I made a whopping $55 total, $40 of which was onboarding pity money.

[00:07:28]Am I only getting paid for that percentage that's accepted?

[00:07:32]>> Yep.

[00:07:33]>> And you think that's fair?

[00:07:35]Um, why would it be not fair?

[00:07:38]>> Because I'm still doing the work. I'm still recording for you 7 minutes.

[00:07:42]>> If robot cannot see your hands, it can't mimic the action. It's essentially like economically useless. I'll be honest, our global average of acceptance, like what we pay is 72%.

[00:07:57]>> So, you're saying I'm pretty bad at this job?

[00:07:59]>> I I think it can get better over time.

[00:08:02]Yes, you have to be comfortable with strapping Big Brother to your head. The recordings don't include audio, but this company now has hours of footage of inside my bedroom and my kitchen. Do I like the idea of a nice little side hustle where I can make money just doing my everyday chores? Sure, but this also isn't my full-time job. The two women who came to clean the apartment rely on it for actual income. Shift pays cleaning crews a fixed hourly rate regardless of acceptance of the video.

[00:08:31]The women who cleaned my friend's apartment each earned $100 for 3 hours of work. That comes out to a little over $33 an hour each. That's a bit more than my friend usually pays for a similar service to clean his space. MicroAGI says the company aims to pay people more than the local minimum wage or more than going rates. MicroAGI is expanding beyond cleaning in the US.

[00:08:53]>> Shift is launching free private chefs in San Francisco.

[00:08:56]>> Yep, the tech bros are back. They sent one to make me lunch. Not a tech bro, a real chef. The robots are going to love truffle. But through this all, I kept thinking, "Okay, I'm not making much money, but at least I'm helping train the robots for the betterment of society." But then I had another thought. What if I'm not even doing that? What if I'm just wearing a privacy invading dystopian science project on my head with no real evidence that this will improve the field? Is this a successful way of training robots? Am I helping by wearing this thing?

[00:09:30]>> Yes and no. Yes, you are helping with a really hard part of the robotics problem, which is diversity or coverage of possible physical experiences that you can have. My prediction is that data is not enough. You also need fundamental advances in how we model the data. So this might be for example new neural network architectures. The transformer was so powerful for natural language. We don't know it is the right architecture yet for physical intelligence. This point of view video training method is just one way of training robots. They can also be trained through tea operation simulations and more. Still, enough people believe this will work.

[00:10:07]Microaggently raised $55 million, and it says it's already paid out $5 million to 10,000 data collectors around the world in countries like Turkey, Germany, the UK, and more. And it's not just Shift.

[00:10:21]Other companies are providing money for your recording services. This is happening globally by other companies including in India where workers are recording. Which brings us to the darker side of this. As I watched those women clean that apartment, I kept thinking about the class divide behind this robotic future. Low- paid workers doing the physical labor today to help build what Goldman Sachs predicts could be a $ 38 billion humanoid robot market by 2035. And there are over 800,000 professional housekeepers in the US according to the Bureau of Labor Statistics. And that's the low-end estimate, not counting folks paid in cash under the table. There's something a little bit dystopian about people collecting the data that could one day automate their own jobs.

[00:11:08]>> There is already new jobs being created.

[00:11:10]There will be surely more jobs being created. What shift is literally as per its name is trying to suggest is that this shift is happening already so you would rather join it than be anxious and watch Tik Tok about it.

[00:11:26]>> He also floated an idea someday robotics companies could share their profits with the people providing all this training data. Yeah. Uh-huh. Because that worked out so well for the people whose work helped train large language models. So, for now, the choice is pretty simple.

[00:11:43]Strap in, do your chores, and train the robots and take a pretty tiny paycheck or let someone else do it. Training the robots to kill the robots is definitely what this company wants.

[00:11:56]I think our work is done.