Last August, Anthropic ran a small experiment with a robot dog. Two teams of non-expert employees raced to program an off-the-shelf quadruped to use its sensors and cameras; one team had Claude to help, the other had only the internet. The AI-assisted team won. When researchers then asked the model to do the tasks entirely on its own, it failed immediately, stumbling at the very first step of connecting to the robot. Ten months later, according to a report in Tech Times on the second phase of the project, that same setup produced a startling result: the model finished all four tasks the human teams had managed, alone, in nine minutes and 35 seconds. The Claude-assisted humans had needed 181 minutes. The unassisted humans needed 361.
The numbers are eye-catching, roughly 19 times faster than the AI-helped team and nearly 38 times faster than the team working without it. But the more telling figure is the code. Anthropic says the model wrote 1,045 lines to do the job; the human team back in August had written 10,309. Humans explore by trial and error, trying one sensor library, failing, trying another, piling up dead ends. The model, the company reports, mostly found the direct path on the first attempt and wrote only what each task required.
What makes the story worth reading is not that an AI beat some amateurs at a programming puzzle. It is Anthropic's claim about why it happened. The company says it did no special robotics training between the two phases. The leap from "cannot connect to a robot" to "outpaces human teams" came, in its words, from the same "general scaling" that has been improving these models at everything else. If that is right, physical-world capability is not a separate mountain to climb but a side effect of the climb already underway.
Then comes the wall. The one task the model could not do was the actual fetching: nudging a ball back to a starting point. That requires what engineers call closed-loop control, continuously reading where the ball went and correcting faster than the situation changes. Writing a program and checking whether it worked is open-loop, and the model excelled at it. Reacting in real time, push by push, is a different kind of problem, and current model architectures are simply too slow in the loop. Humans, after a little practice with a controller, did it without thinking. The machine could line up perfectly behind the ball and then could not make the rapid corrections the task demanded.
That boundary is the honest center of the story. A researcher with more robotics experience did manage to write a working autonomous retrieval program, so the problem is not unsolvable, just beyond what general scaling has reached. Anthropic frames the whole thing as the start of "physical agentic AI," the same three-stage arc it claims to have watched in software and cybersecurity: tools first augment humans, then humans briefly steer the tools, then the tools work alone. It is a single data point from a deliberately constrained test, and worth treating as such. But the shape of it, fast progress on the open-loop tasks and a clear architectural wall on the real-time ones, tells you more about where robots are headed than any demo reel.