The week's most eyebrow-raising development: Anthropic has apparently set up a physical biology lab where Claude can run actual experiments. That's not a metaphor for agentic AI — it's a wet lab. Separately, the company partnered with Novo Nordisk on drug discovery workflows. The pattern here is deliberate: Anthropic is positioning Claude not just as a reasoning tool but as an active participant in scientific production, which is either the most ambitious enterprise pivot in AI or the setup for a very interesting liability question. Meanwhile, a research team used AI to narrow four candidate muscle-saving compounds down to one, and small AI models are finding a niche in longevity biology's notoriously messy datasets — signs that domain-specific, smaller-footprint models are doing real work where frontier models would be overkill.

On the more speculative end of the biology-AI frontier: Wired reports that organoid-based AI models — systems using biological neural tissue from rat brains — are closer to practical deployment than most people realize. That sits in a different category from Claude pipetting reagents, but both point toward a week where the boundary between AI as software and AI as physical agent got considerably blurrier.

AstroForge is handing autonomous AI command authority over its next spacecraft — not as a backup system, but as the primary decision-maker for a mission where round-trip communication latency makes human-in-the-loop control impractical. That's a consequential real-world deployment of autonomous AI, and it's happening with almost none of the governance discourse that surrounds far less consequential chatbot use cases. At TechCrunch Disrupt, Hello Robot brought its Stretch 4 robot onstage in what the company is framing as a milestone for accessible, affordable manipulator hardware — a category that matters more as physical AI moves from warehouse to elder care. Meta's work with the University of Pittsburgh on AI-assisted assistive robotics adds a third data point: the robots are arriving, and they're showing up first where the use case is clearest and the stakes are personal.

Trump has now proposed an 'AI Force' — a new military branch — on top of earlier proposals to rename the field entirely. The substantive policy action this week came from a more credible direction: the FDA has contracted Cognita Imaging to test whether LLMs can evaluate radiology AI systems, which is quietly significant as a framework for how AI gets used to audit other AI. New York's seizure of twelve celebrity deepfake websites remains one of the only concrete state enforcement actions using existing law in the absence of federal rules. The MIT 'HardFlow' algorithm — designed to make generative AI outputs comply with strict safety constraints for safety-critical applications — is the research answer to the governance question, though it will take longer to matter than the policy window requires. The gap between what regulators are discussing and what engineers are deploying has not narrowed this week.