Anthropic just released Phase 2 of Project Fetch. They gave their latest AI model a robotic dog and told it to figure out how to control it. No human help. No robotics training data. Just raw intelligence applied to a physical machine. Claude Opus 4.7 completed every task roughly 20x faster than the best human team that had AI assistance less than a year ago. It wrote 10x less code and got most of it right on the first try. None of this came from targeted robotics training. It emerged from general model scaling. The AI got better at controlling robots because it got better at thinking. That is a fundamentally different dynamic than anything we have seen in automation before. Anthropic keeps seeing the same pattern across domains. First AI helps humans do the task. Then humans help AI do the task. Then AI does the task alone. They saw it in cybersecurity. Now they are seeing it in physical robotics. The model still failed the final challenge of precisely nudging a beach ball to a target. Real-time perception and adjustment is still out of reach. But the gap is closing fast, and it is closing as a byproduct of general intelligence, not domain-specific engineering. Less than two years ago the AI could not even connect to the robot. Now it programs the robot 20x faster than the humans who built the experiment.