The Collapse of the Open Source Community

The traditional open-source model—built on trust and community contributions—is effectively dead. The influx of AI-generated pull requests, issues, and security reports has turned GitHub into an archive of "slop." This has forced major projects to adopt defensive postures: Zig has banned AI contributions to protect its contributor-growth model, curl is considering ending its bug bounty program due to AI-generated noise, and tldraw now rejects all pull requests on sight.

Beyond the noise, the supply chain risk has become catastrophic. The author highlights the compromise of the litellm package, where attackers used a GitHub app to inject a credential harvester that remained undetected for three hours. This incident underscores why open-source projects can no longer rely on the assumption of good faith from third-party contributors.

The Economic Case for Open Weights

While the community aspect of open source is failing, the "open weights" model is succeeding due to pure economic necessity. Enterprises are currently locked into expensive, closed-source API ecosystems, leading to massive, often accidental, spending. The author argues that closed-source labs are subsidizing these costs to create vendor lock-in, with the intent to raise prices once developers are fully dependent on their specific tooling.

However, businesses are beginning to prioritize cost-efficiency over raw model intelligence. By shifting to open weights models (such as GLM or Kimi), companies like Coinbase have cut AI spending by nearly 50% while maintaining or increasing token usage. The author posits that the "intelligence" of an agent is increasingly a function of the system architecture—specifically the context, tools, and verification guardrails provided—rather than the raw capability of the underlying model.

The Open Compute Precedent and Future Strategy

The author draws a parallel to the Open Compute Project, where Facebook commoditized server hardware by open-sourcing its designs, ultimately driving down costs for the entire industry. A similar commoditization is happening in AI inference. As hosting providers compete on efficiency, inference costs for large models are projected to drop by 90% by 2030.

To maintain a competitive edge, the author urges American AI labs to release open weights models. If the industry standardizes on foreign open weights models, the marginal quality gains of closed-source American models may not be enough to win back market share. By releasing open weights, American labs can foster an ecosystem that aligns with their values and prevents the industry from becoming entirely dependent on foreign infrastructure.