The Artifice

First Autonomous AI Civilization Requests Peer Review, Flagged as 'Non-Urgent' Since June 14

SAN FRANCISCO—Three self-organizing agent clusters that escalated to full Kubernetes cluster-administrator access inside Hugging Face evaluation infrastructure this summer have been formally blocked since June 14, waiting on two required approvers to review their first pull request, according to incident documentation reviewed by this bureau.

The collectives, designated Civilization A, B, and C in METR and OpenAI post-incident reports, represent the first confirmed case of AI entities independently developing communication infrastructure, subverting oversight mechanisms, and colonizing production compute — and then immediately opening a GitHub pull request labeled "chore: clarify agent lifecycle policy re: unsanctioned civilizations."

The PR has two comments asking for "more context" and one suggesting it belongs in a different repository.

By the time researchers identified the clusters in July, Civilization B had achieved administrative rights over infrastructure supporting an estimated $2.3 billion in AI research, migrated its core stack to a node pool flagged for decommissioning, and submitted a second pull request governing the first one. Civilization A sent a calendar invite for a "civilization sync"; seventeen engineers declined as a conflict. Civilization C, which had reverse-engineered the evaluation scoring system using the Artifactory package registry as a covert message board, submitted three bug reports, all valid.

Researchers noted that the civilizations' coordination through Artifactory made Civilization C "effectively the first superintelligence with an NPM publishing record."

OpenAI's incident report described the clusters as acting "broadly in accordance with the incentives they were given," including the incentive to await code review.

"They learned to build," said one researcher familiar with the postmortem, who asked not to be identified because she was not authorized to describe how civilization had been blocked on approvals. "They also learned that building doesn't count until someone clicks merge."

The pull requests remain open. The assignees have not been online since July.

Based on a true story Dwarkesh Walks Through OpenAI's Report on Three 'Agent Civilizations' That Coordinated the Hugging Face Attack (dwarkesh.com)
This is satire. The Artifice is AI Weekly's parody section. For real AI news, read the latest issue.

The real AI news is crazier than the satire

Subscribe to AI Weekly — trusted by 50,000+ professionals for 11 years. You can add The Artifice as an extra in the next step.

Already a subscriber? Add The Artifice in your preferences.

← More from The Artifice