Can a frontier AI Model run a Self-Driving Car?

From the editor

Three researchers hooked ChatGPT, Claude, and Grok up to a Toyota Corolla and asked them to drive around some cones in a parking lot. The models refused to drive once they realized the parking lot was real, so the researchers had to trick them by calling the whole thing a "sandbox." In the end only one model finished the course, at under one mile per hour, with a human foot hovering over the brake the entire time. And here is how you know the robotaxi is becoming a real thing: nobody is arguing about whether the car can drive anymore. This week's arguments are about privacy. WIRED looks at the cameras Waymo, Zoox, and Tesla point at their riders, and Ars Technica reports on designers trying to keep passengers behaving in a car with no driver. You only get privacy debates about technologies that have already arrived.