Weekly Newsletter #51 | Updates Worth Your Attention 👀

Employees at leading AI companies — OpenAI, Anthropic, Meta, DeepMind — launched a petition calling for the pace of AI development to be brought under control. The manifesto they published boils down to:

• Every company is in a blind race to ship the most capable model
• Nobody is asking what happens to safety when these models get this powerful
• Bringing it under control requires governments, industry, and society to work together

The employees say the campaign reflects only their personal views and not their employers — but the CEO of Anthropic signed it, and OpenAI shared it from their official X account.

No previous technological development has carried risks of this magnitude. The fact that the quantitative and qualitative growth of offensive capabilities is now outpacing cybersecurity defenses is one of the concrete details that makes these concerns legitimate. Whether the solution really comes from government-industry cooperation is an open question, but one thing is clear: open-source models need to multiply and the development process of these models needs to be made transparent.

Pacing the Frontier →

Kimi published a paper releasing the model weights and training details for Kimi K3, their most advanced model to date with 3 trillion parameters. While closed models share nothing about their methods, this release matters: any company can now run this open-source model — approaching the capability of frontier models — on its own infrastructure and customize it. The paper includes the model’s architecture, weights, required infrastructure details, and cost estimates.

Model Weights (HuggingFace) → | Tech Blog (Kimi) →

Hundreds of conversations from the Claude interface were found to be accessible by simply adding “site:claude.ai/share” to a Google or Bing search. Anthropic moved quickly to block access. What surprised us: the exact same issue was previously discovered with OpenAI and Grok conversations — we covered it here at the time, though most people have probably forgotten. The fact that Anthropic made the same indexing mistake as if they’d never heard of these incidents was striking.

BBC →