Project Fetch: Phase Two

5 hours ago

An experiment called Project Fetch tested how much Claude AI could assist non-expert humans in performing tasks with a robotic quadruped (robodog).
In August 2025, Anthropic employees using Claude Opus 4.1 outperformed a team relying only on the internet and ingenuity, completing tasks faster and more effectively.
A follow-up experiment showed that Claude Opus 4.7, operating autonomously without human help, was about 20 times faster than the fastest human team from the previous year on certain tasks.
Despite improvements, Claude still struggled with precise robotic control, such as fetching a beach ball autonomously, indicating that LLMs have not fully solved robotics challenges.
The progress observed is attributed to general AI scaling rather than targeted robotics improvements, suggesting rapid advancements in model capabilities.
The experiments highlight a pattern where AI models evolve from assisting humans to operating independently, with potential implications for physical agentic AI and the use of off-the-shelf tools.
Future research may focus on AI's ability to customize physical tools and develop tailored control policies, though barriers to generalized physically capable models remain.
The post draws parallels with advancements in cybersecurity and software tool development, suggesting similar trajectories could occur in hardware and robotics.

Hasty Briefsbeta