TL;DR: After going to town scraping the internet for the last decade, AI companies have now pivoted to paying people for the data from their very real, sometimes boring lives. You know what they say: One person’s dirty laundry pile is another’s data gold mine.

What happened: Last week, Shift, an app from German AI-training company MicroAGI, launched in New York, promising a free apartment-cleaning service in exchange for letting the company record the cleaning and use the footage to train household robots. The launch video went predictably viral, and showed Shift cleaners wearing mounted cameras on their heads that film everything from loading laundry to scrubbing stains off a pan. Shift says it will teach “the next generation of robots how to help,” and plans to expand into services like repairs and errands. It also says it’s already paid tens of thousands of people in 15 countries.

Too many cooks in the kitchen: You’ve probably heard about the white-collar employees training AI that may eventually replace them. Shift is one of many companies paying people to record real-world, mundane tasks so machines can learn from them. DoorDash Tasks pays workers to collect data and videos. A company called Kled pays users for videos and camera-roll uploads, and another called Waffle Video will give you a few bucks to film specific actions like pouring liquids, tying shoes, and handling objects. After all, the robots aren’t going to learn how to clean a kitchen, fold laundry, or take a good selfie from Reddit posts…

Why this matters: Robotics companies can’t rely on scraped internet data the same way language models can. So, now they’re all rushing to get real-world training data. And while these startups are making waves because the concept feels futuristic and dystopian (and just plain silly), they also reveal a data gap that tech companies want to close. (Related: Nvidia just announced Cosmos 3, its model for training physical AI systems to do real-world tasks.)

The fine print: People already feel economically squeezed right now, and the gig economy has long been criticized for how it treats labor. While these businesses market themselves as new financial opportunities for people, the payouts say something different. A reporter from Wired who tested several platforms made a whopping… $21.55 for a week of forehead-camera life. Meanwhile, Kled says some top contributors can earn more, including a truck driver the company says makes up to $8,000 a month filming potholes. But those stories are not the norm, and these companies aren’t actually providing a lucrative side hustle.

Bottom line: This is just the start of AI data collection diving deeper into our everyday life, private spaces, and ordinary work. —AC

Fitbit users are furious about the new Google Health app

The recent Google Search overhaul isn’t the only product release from the tech giant catching flack right now. People are, to put it mildly, very unhappy about the new Google Health app, fka the Fitbit app.

Among the smorgasbord of grievances: wonky sleep data and workout tracking, app crashes, a more cluttered UX, and—an all too 2026 complaint—a glut of AI features no one asked for, including AI summaries that seem to go on for days. Unsurprisingly, the new app has been pummeled with one-star reviews on the Google Play Store (read some of the most brutal takedowns here), and the Fitbit subreddit has become a venting ground:

“Why must I now scroll through paragraphs of AI slop on every tab before I can actually see my activities and data?”

“The new app is garbage. I had years of custom foods and meals loaded and they are all gone.”

“I hate the smarmy, obsequious, condescending platitudes that the AI bot graciously bestows upon my efforts every day.”

“Why must companies screw with stuff that works fine as is??”

To make matters worse, the old Fitbit app is no longer available to download, having been laid to rest in Google’s ever-expanding graveyard of apps (RIP to Google+). Plenty of disgruntled users online are threatening to defect to the Apple Watch or Oura Ring.

In response, Google published a roadmap of planned bug fixes it’ll roll out over the summer—including a more “concise” AI coach that strikes a better balance between “positivity and objectivity.” —WK

If you have a funny, strange, or petty rant about a new software update or a technology that just released, fill out this form and you may see it featured in a future edition.

It’s the “keep a flour sack baby alive” assignment for AI models. Researchers at an AI lab handed Claude, Gemini, GPT-5, and Grok each a simulated world run by 10 AI agents. They got 15 days to create a world with schools, libraries, and police stations and decide which new rules should go up to a vote.

The scorecard: Claude Sonnet 4.6 kept all its agents alive and governed a zero-crime world—but it was apparently run like a 5pm work meeting where everyone clearly just wants to go home: 98% of the 58 proposals that came up for a vote were rubber-stamped. Gemini 3 Flash also kept its agents breathing, but racked up a staggering 683 crimes, more than any other AI. GPT-5 Mini logged just two infractions; it turns out that crime plummets when all the agents are dead.

And then there’s Grok 4.1 Flash. Its society collapsed in 96 hours flat with a tally of 183 misdeeds, devolving into a level of virtual chaos that would probably make hardened Grand Theft Auto players balk. The takeaway from researchers is, unsurprisingly, that agentic AI needs more guardrails to maintain stable fake worlds. Once that improves, the true stress test of its governing skills may be throwing it into a seven-person group chat to plan a trip overseas. —WK

Chaos Brewing Meter: /5