Anthropic's AI agent outperformed humans by 20 times in controlling a robot dog: breakthrough or temporary success?
The artificial intelligence industry continues to surprise us with unexpected applications. The Anthropic team has released the results of the second phase of the Project Fetch experiment, and they are impressive: their latest model, Claude Opus 4.7, handled programming and controlling a four-legged robot 20 times faster than the best human engineering team working last year.
To remind you, in August 2024, the experiment looked different: employees with no robotics experience used AI as an auxiliary tool to find solutions. Now, the situation has changed dramatically. Claude Opus 4.7 operated almost entirely autonomously — under minimal researcher supervision. The neural network independently completed a whole range of tasks: connected to video sensors and LiDAR, wrote a manual control program, created a robot trajectory monitoring system, and configured an object recognition algorithm.
Superiority in Speed and Code Efficiency
The numbers speak for themselves. Opus 4.7 proved to be 18 times faster than the team using older AI versions and 37 times faster than humans working without chatbot assistance. But the key point is not just speed, but also quality. The code generated by the neural network turned out to be 10 times more compact than that of human teams. This directly indicates that AI is capable of finding more elegant and optimal solutions, avoiding the excessive complexity inherent in human programming.
Particular attention should be paid to Anthropic's statement that progress in robotics has become a "side effect" of the general scaling of language models. The company did not implement specialized algorithms for controlling hardware — this means that skills for managing physical objects emerge as an emergent property of more powerful base models.
Physical Limits: Where AI Still Falls Short
However, one should not rush to conclusions about completely replacing humans. The experiment also revealed a significant limitation: Claude successfully guided the robot to the target but failed the task of gently pushing a ball to a specific point. This operation requires complex real-time feedback — an area where humans still outperform AI. The physical world, with its nuances and unpredictability, remains a serious challenge for neural networks.
Anthropic believes we are entering the era of "physical AI agents." My analysis confirms this trend: we are witnessing the convergence of language models and robotics. However, the path to fully autonomous physical agents capable of performing delicate manipulations is still far off. Interestingly, amid these successes, on June 13, Anthropic was forced to halt access to the Fable 5 and Mythos 5 models due to a directive from the U.S. government on export controls — a reminder that geopolitics goes hand in hand with technological progress.
Expert opinion: Anthropic's breakthrough demonstrates that the physical world is becoming the next frontier for LLMs. However, the 20-fold speed advantage should not be misleading: while AI surpasses us in planning and code writing, it falls short in tactile sensitivity and adaptation to micro-changes in the environment. A true breakthrough will occur when neural networks learn to combine these two skills.