Plus, Elon’s latest post is causing a stir.
The Download
The day the robotaxis froze
Baidu
TL;DR: A few days ago, over 100 Baidu robotaxis halted on highways in Wuhan, China—the second mass robotaxi failure in a little under four months. These incidents show how centralized cloud systems make robotaxis especially vulnerable to shutdowns—and reveal the lack of a clear contingency plan when an army of driverless cars goes dark at once.
What happened: On Tuesday, over 100 Baidu Apollo Go robotaxis suddenly came to a stop in the middle of Wuhan’s busy, winding expressways. Though there were some collisions, no injuries have been reported, and Chinese authorities attributed the incident to a “system malfunction.”
These driverless cars didn’t pull over to an emergency lane—they just halted where they were, some in the fast lane with traffic streaming past on both sides. Robotaxi systems are generally designed to pull over to a safe location when something goes wrong. In this case, they didn’t—and it’s not yet clear why. Some passengers reported that in-car SOS buttons didn’t work, and one college student told Wired it took 30 minutes to even connect to a customer service rep—and help never came.
What makes robotaxis vulnerable: Robotaxi companies run their fleet on a centralized cloud with shared systems for routing, navigation, and remote assistance. When there’s a major hiccup in that infrastructure, it can take down a huge number of connected vehicles in the fleet at the same time. When the system fails, there’s not always a human driver who can grab the wheel, especially as many commercial robotaxi fleets no longer place human safety drivers (as they did during testing) in the vehicle.
Robo-car crash pileup: This wasn’t the first time robotaxis have come to a standstill. Last December, when a massive blackout in San Francisco knocked out city traffic lights, Waymo’s fleet of robotaxis caused a giant traffic jam. The cars are supposed to handle dead traffic lights on their own—but they’re also programmed to occasionally ping Waymo’s remote assistance team for a “confirmation check” before proceeding. During the blackout, a lot of cars in the fleet were pinging at once, and the cellular network was down too, leading to the traffic buildup.
A glimpse of the future: What happened in Wuhan isn’t just an aberration; it’s a peek into a world where robotaxi adoption is a lot further along than it is in the US. Over 1,000 Apollo Go vehicles roam the city—Baidu’s largest fleet anywhere—serving a population of more than 12 million. China has been further ahead on robotaxi deployment and in discovering what can go wrong with self-driving more broadly: A fatal assisted-driving crash involving a Xiaomi car last year prompted regulators to tighten rules around cars with autonomous driving features and ban the term “autonomous driving” from car ads. But regulating individual cars is one thing. There doesn’t appear to be a widely adopted plan—whether in China or elsewhere—for what happens when a city’s worth of robotaxis go dark all at once.
Bottom line: In the US, Waymo is now doing 500,000 paid rides a week and targeting 1 million by year’s end. But when a fleet serving so many rides fails at once, what’s supposed to happen? So far, the standard protocol seems to be “call customer service and hope someone picks up.” —WK
The Zeitbyte
Your niche expertise? ChatGPT’s next lesson plan
Morning Brew Design
OpenAI wants ChatGPT to know how to ranch cattle, fly a commercial plane, and identify soil composition—and it’s apparently paying real humans with those skills between $50 and $500 an hour to make it happen. The vehicle: data-labeling startup Handshake AI, which has quietly recruited 3,000 to 4,000 freelancers for a project internally called “Stagecraft,” Business Insider reported.
The job is essentially method acting for AI training. Contractors channel their own professional expertise to simulate work tasks—a nurse practitioner triaging patients, a sculptor sourcing materials—complete with real goals, context, and deliverables as if briefing an actual colleague. Handshake’s training guide specifies the focus is “knowledge work, not manual labor” and that contractors shouldn’t water things down “based on assumptions about what an AI ‘needs.’” No dumbing it down for the robots.
One contractor captured the existential math pretty cleanly: “We all were aware that we were basically training AI to replace us.”
It’s not just knowledge workers, either. Jack Dorsey just copublished a lengthy essay arguing AI should replace middle managers, a thesis he spent the first half building up to by citing the Roman army, the Prussian military, and the American railroad industry. The contractors, at least, are getting $500 an hour for the trouble. —SM
Open Tabs
- Apple issued a rare security patch for older iPhones to block the DarkSword hack.
- SpaceX is reportedly hosting its first-ever analyst day on April 21, kicking off a road to what could be the largest stock listing in history.
- OpenAI secretly bankrolled a coalition pushing for an age verification bill that could also benefit Sam Altman’s iris-scanning startup.
- Anthropic’s Claude Code DMCA blitz knocked out 8,100 GitHub repos while trying to claw back its accidentally leaked source code.
- The CEO of America’s biggest bank said it’s eyeing a move into prediction markets.
- Elon Musk has triggered Bitcoin price crash fears after pointing out an upside to Google moving Q-Day up to 2029: “If you forgot the password to your wallet, it will be accessible in the future.”
- United’s app now tells you how long the TSA line will be. As a bonus, you can also share your bag’s AirTag location directly with agents since you were already refreshing it every 30 seconds.
- Google just released Gemma 4, an “open weight,” downloadable version of its Gemini AI that can run locally on your device.
This time last week…
Readers’ most-clicked story was about a humanoid robot walking the red carpet at the White House. (You can watch the whole thing here.)
Written by Whizy Kim and Saira Mueller