The AI company racing to build safe superintelligence just got caught copying homework from Meta's old version control system.

The Summary

The Signal

Anthropic published a technical post about redesigning their source control system to handle their growing AI development needs. They positioned Scion as a fresh approach to monorepo management, necessary because Git and Mercurial couldn't scale to their requirements. The post got significant traction: 473 points and 470 comments on Hacker News.

Then engineers who'd worked on Meta's infrastructure started reading more carefully. The first analysis pointed out that Scion's architecture mirrors Mononoke, the version control server Meta built to handle their massive monorepo. Both use a content-addressed storage layer. Both separate the working copy from the repository backend. Both solve the same scaling problems with similar technical approaches.

"The similarities aren't coincidental. They're architectural."

The second, deeper investigation reveals more overlap:

  • Both systems use a proxy server model to handle massive repositories
  • Both implement lazy fetching strategies for working copies
  • Both separate repository metadata from file content in similar ways
  • Both optimize for the exact same bottlenecks Git hits at scale

The issue isn't that Anthropic built on existing ideas. That's how engineering works. The issue is presenting Scion as novel infrastructure while downplaying how much it borrows from Mononoke's public architecture. Meta open-sourced Mononoke. The design patterns are documented. But Anthropic's announcement barely mentioned prior art.

This matters beyond version control drama. AI labs position themselves as frontier research organizations. They raise billions on the premise that they're inventing new paradigms. When their "novel" infrastructure turns out to be established patterns from Meta's 2018 playbook, it raises questions about what else is borrowed versus built.

The Implication

Watch how AI companies talk about their infrastructure breakthroughs. The actual engineering is often solid, sometimes even necessary. But the framing matters. If an AI lab can't accurately credit influences in their version control system, how confident should we be in their claims about novel AI safety research or unprecedented capabilities.

For engineers building in this space, the lesson is clearer: standing on giants' shoulders is fine. Just say whose shoulders they are.

Sources

Hacker News Best