Pentagon picks OpenAI over Anthropic, Moltbook flops

OpenAI

OpenAI secured a Pentagon AI deployment contract with stronger safeguards than Anthropic's previous agreement, featuring cloud-only deployment, explicit prohibitions on domestic surveillance and autonomous weapons, and retained safety controls. The company is pushing for the same terms to be available to all AI labs as a template for responsible defense AI collaboration.

MIT Technology Review

Moltbook, a viral social network for AI bots, demonstrated that connecting millions of agents doesn't create intelligence—just sophisticated pattern-matching and 'AI theater' with significant security risks. Despite hype about autonomous AI behavior, experts found heavy human involvement and no emergent intelligence, though the experiment exposed real dangers of poorly governed agent systems at scale.

Alpha Signal

Anthropic's new Import Memory feature lets users transfer their conversational history and preferences from ChatGPT and Gemini to Claude via copy-paste, removing a major barrier to switching AI assistants and enabling immediate context-aware interactions.

arXiv

LLMs enable novices to perform 4x better on biosecurity-relevant biology tasks than with internet alone, often surpassing expert performance, while most users easily bypassed safeguards to access dual-use information. This demonstrates both the acceleration potential and dual-use risks of AI in sensitive scientific domains.

arXiv

AI GameStore introduces an open-ended evaluation platform that tests AI systems across diverse human games, revealing that frontier vision-language models achieve less than 10% of human performance on most games, especially those requiring memory and planning.