Anthropic says AI models will move into the physical world. To test that idea, the company asked Claude to program a quadruped robot.
As robots begin appearing in warehouses, offices, and even people’s homes, the notion that large language models could reach into complex systems sounds like something out of a science-fiction nightmare. Researchers at Anthropic were eager to observe what happened when Claude attempted to take control of a robot — in this case, a robot dog.
The team monitored the model’s outputs and the machine’s responses, and engineers maintained oversight. The run highlighted practical and ethical questions about letting language models influence hardware, and it offered a glimpse at how quickly software and moving parts can intersect. Researchers said the experiment was controlled, with humans supervising each action and ready to intervene if anything unsafe occurred.