When robots need humans: safety failures and hidden costs

There's a particular kind of irony happening in robotics right now. We spend billions building machines to eliminate human error, then discover the humans we've asked to oversee them are the weak link. That's exactly what happened in Austin last month when a Waymo robotaxi encountered a stopped school bus. The vehicle did precisely what it was supposed to do: it stopped and asked its remote operator for permission to proceed. The operator said no, it wasn't a school bus. The vehicle obeyed. Six other cars followed. All of this happened at 7:55 on a Tuesday morning with children loading onto that bus.

This isn't a failure of autonomous driving technology. It's a failure of the entire system that's supposed to catch edge cases the AI can't handle. When you build a safety net that relies on a human operator half a world away making split-second decisions about what they're seeing through a camera feed, you've created a new problem instead of solving an old one. The NTSB is investigating. There will be recommendations. But the fundamental tension remains: how do you scale autonomous systems that still need human judgment, and how do you make that human judgment reliable?

The humanoid robot reality check

Meanwhile, humanoid robots are starting to arrive in homes. 1X's Neo bot costs $20,000, stands five foot six, and promises to do laundry and load dishwashers. What the marketing doesn't emphasise is that for most tasks, a remote operator wearing a VR headset is controlling the robot from somewhere else, and every movement is recorded. The company says these updates will happen less frequently over time, but right now, there's a human watching your house while your robot learns to fold your clothes. This creates a privacy frontier we haven't really grappled with yet: incredibly sophisticated machines collecting intimate data about how you live. And behind the scenes, remote workers in developing countries are doing the labour that makes this possible, often for low pay and with exposure to content that's disturbing.

The portability problem nobody's talking about

On a different but equally important front, a developer named Isabel Smith has written something that deserves your attention. It's about AI conversation history. Right now, if you've spent two years building context with ChatGPT-your projects, your preferences, how you think-exporting that data produces a 500MB JSON file that no other AI system can actually use. Claude has a different format. Gemini has another. There is no standard. This is vendor lock-in dressed up as data portability. It's what we accepted from cloud providers 15 years ago before we learned better. A startup called Phoenix Grove Systems built a tool called Memory Forge that converts these formats for $3.95 a month, proving it's technically solvable. The platforms simply have no incentive to solve it themselves.

These three stories-the school bus, the humanoid in your living room, and the formats nobody talks to each other-tell the same underlying story. We're building systems that work brilliantly in isolation, then discovering that connecting them creates new risks and problems we didn't anticipate. The good news is most of these problems are solvable. The hard part is that solving them often requires someone to give up a competitive advantage. And that's where progress tends to stall.