While Fable 5 sits dark on day 6, Anthropic quietly published Project Fetch Phase Two — a follow-up robotics experiment where Claude Opus 4.7 completed tasks 20x faster than the fastest human team, with 10x less code. Physical AI just took a step.
Anthropic gave Claude Opus 4.7 autonomous control of an off-the-shelf robotic quadruped (a robot dog). The tasks: connect to sensors, detect objects, navigate environments. Same tasks human teams performed — some with Claude assistance, some without.
1
Human team without Claude: 361 minutes per task
2
Human team with Claude: 181 minutes per task
3
Claude autonomous: 9.5 minutes per task — code worked on first try
The code quality finding is striking: Claude produced 1,045 lines that worked on the first try. The human-AI team produced 10,309 lines requiring iteration. Less code, better results, faster delivery.
Where Claude Failed
Claude couldn’t fetch a beach ball — the actual “fetching” task the project is named after. The model struggled with closed-loop precision control: real-time perception adjustments based on visual feedback. The kind of adaptive motor control that humans learn through practice.
The Framing Ladder applies here too. Claude excels at the “frame → execute” workflow (sensor setup, object detection, navigation planning). It fails at the embodied, real-time, adaptive loop — the kind of task that requires physical intuition, not planning. High-FRED tasks (structured, decomposable): 20x faster. Low-FRED tasks (adaptive, embodied): still human.
These improvements came from general modelscaling — not robotics-specific training. Opus 4.7 wasn’t fine-tuned for robots. It just got smarter overall, and that intelligence transferred to physical tasks. This is the Dynamo Doctrine’s deeper claim: anything with structure is convertible to tokens. Robot control is structured. Therefore it becomes a model capability — not a separate engineering discipline.
The same week the AI Supercycle predicted “physical AI crosses the chasm” as a near-term development, Anthropic published the data showing it’s happening — through general intelligence scaling, not robotics R&D.
Business Engineer
AI & The Dynamo Doctrine
Anything with structure is convertible to tokens. Robot control. Protein folding. Medical imaging. The substrate is wider than language.
Anthropic’s best model is dark for consumers. But in the lab, it’s controlling robots 20x faster than humans with 10x less code. The irony: the model the government pulled for being “too capable” at reading code is also too capable at writing robotics code. The Permission Layer paused deployment. It didn’t pause capability.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
