When AI Meets Physical Robots-And When It Should Stop

Friday afternoon brings three interconnected stories that reveal where AI is heading, and where some companies are drawing lines. There's genuine tension here between capability and responsibility, between what's possible and what's wise. Let's untangle it.

The Physical AI Wave Is Real

Intrinsic-the robotics startup that spun out of Alphabet five years ago-is officially joining Google. This isn't a small acquisition buried in a press release. This is Google signalling that physical AI (robots that can actually think and adapt in real environments) is a core business bet, not a research lab curiosity. Intrinsic's Flowstate platform lets developers build intelligent robotics applications without needing to be deep AI specialists. You can assemble robots from building blocks of behaviour, simulate them, then push to production. The work with Foxconn to build AI-driven factories of the future gives you a sense of scale here.

Separately, we're seeing drones deployed for something beautifully grounded: tracking plastic pollution on beaches. A team at the University of Limerick trained computer vision models to spot bottle caps from 30 metres up, distinguishing them from driftwood and weathered rocks. They've equipped community groups across Ireland and Europe with the technology. Volunteers get GPS coordinates, head straight to hotspots, and clean with purpose instead of guessing. It's robotics solving a real problem, not a demo.

The Line in the Sand: Anthropic vs. the Pentagon

But here's where it gets serious. Anthropic's CEO Dario Amodei published a statement this week saying the US Department of War pressured the company to provide unrestricted military access to Claude-including for mass domestic surveillance and fully autonomous weapons without human control. Anthropic said no. They drew explicit red lines.

What makes this matter isn't just the confrontation. It's that Google and OpenAI staff have reportedly signed a petition supporting Anthropic's stance. The industry is coordinating around shared values before external pressure can splinter everyone apart. There's a race-to-the-bottom risk here: if one AI lab caves, the others will face impossible pressure to follow. Anthropic's refusal to fold might have just prevented that cascade.

A retired US Air Force General put it plainly: "No LLM, anywhere, in its current form, should be considered for use in a fully lethal autonomous weapon system. It's ludicrous even to suggest it." The capability exists to build these systems. The question is whether we should. Anthropic is betting we shouldn't.

Building in Reality

For developers actually building things, there's serious work happening. Google dropped Gemini 3.1 Flash Image Preview-marketed as "Nano Banana 2"-and it's the fastest image generation model available, priced at half what competitors charge. Meanwhile, builders are grappling with how to structure agent skills properly, how to reverse-engineer how AI platforms actually search the web (spoiler: ChatGPT generates 8+ hidden search queries per question), and how to make distributed systems hold up in production.

The through-line across all three categories is the same: capability is accelerating, but the harder problems are about governance, reliability, and what happens when these systems touch the real world. Build faster, yes. But build responsibly. That's the afternoon's lesson.