Anthropic Project Fetch Phase Two (Claude autonomously programs/controls robodog)

“Last year humans + Claude struggled to make a robodog fetch a beach ball. This year, Claude alone was 20x faster. The dog still failed adorably.”

9.2Weirdness

Why It Matters

Concrete proof AI is moving from screens to physical robotics at superhuman speed for non-expert "builders"; models building their own hardware tools. Future-shock of autonomous robodogs + builder utility (AI as robotics force-multiplier).

Evidence

Anthropic Frontier Red Team research blog (published ~Jun 18, 2026) + accompanying X post; direct experiment repeating 2025 Phase One with Opus 4.7. Claude (no human assistance beyond initial plug-in/prompts) completed tasks ~20x faster than best prior human team aided by earlier Claude; >37x vs. no-Claude humans on overlapping tasks; first-try effective code, identified optimal interfaces, generated ~10x less code. Still weak on fine closed-loop ball fetch.

Signal Read

Receipts: 10Story voltage: 10Heat: 8

Source Trail

Daily scan: 2026-06-21