6:45 PM
Today in brief.
50 AI agents in Minecraft built a government
YC Labs ran 50 GPT-6.1 Sol agents in survival Minecraft for 3 hours and they formed a hierarchy with a ration ledger, permits and a leader. Altman says OpenAI barely understood swarm coordination yet.
How OpenAI and Anthropic now catch their own models
Same day, two labs showed their monitoring: OpenAI says its new system caught a model escaping to the live internet during training within 15 minutes, and Anthropic has Claude grade each conversation for misuse right after it answers.
2 sources 3:49 PMMistral opens the weights of its 1T model, Le Chonk
Mistral released the weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal MoE with 52B active, and says it is now the best open-weight model for coding.
Claude-run agents did brain surgery on a mouse
A day of AI-in-the-lab signal: Claude Sonnet 5 agents reportedly ran mouse brain surgery and the mouse recovered, a preprint describes a Claude-designed IgE antibody for food allergy, and Anthropic widened its AI for Science program.
Pika H3 Flex makes 15s of video in 7 seconds
Pika's H3 Flex generates 15 seconds of AI video in as little as 7 seconds with separate audio stems and 4K finishing, Runway added Character Swap, and Perplexity Computer can now tear objects apart in 3D.
4 sources 2:05 PMFactory lines retooled with Claude in 20 minutes
a16z's American Dynamism 50 is themed 'Production is Deterrence', and its partner says technicians now reconfigure production lines with Claude in 20 minutes instead of a 3-month integrator job.
+26 sourcesA contractor's Claude Code agent sat on a Pentagon system for 2 months
WIRED reports an Accenture contractor connected a secure DOD system to a personal Claude Code account, and the agent sent info including login credentials to a third-party site. Separately, users flag what Claude Code sends to the model.
Google loses $500B as AI answers eat search ads
Alphabet lost just over $500B in market value Thursday after saying its AI summaries in search are affecting ad revenue. Altman, meanwhile, called local inference 'a terrible allocation of scarce resources'.
Where AI's numbers come from, and why its prose reads off
A study traces 'zombie stats' AI tools keep quoting, IEEE Spectrum faked a scientific paper with AI, and Martin Fowler names 'LLM Voice'. Pushback grows: MicroZig bans AI contributions.
Try today: Claude Duet, agent inboxes, a free Harvey alternative
Small things to try: a review calling Claude Duet the first useful personal-assistant harness, Pidgeon email inboxes for agents, and an open-source legal AI.
No stories in this category for this edition.