Sunday, 24 May 2026

The robotics stories this week are interesting precisely because they expose a tension we're not talking about enough. Boston Dynamics' Atlas is genuinely impressive—coordinating a humanoid body to handle complex physical tasks with reliability is no trivial engineering problem. But the more I read about what's actually shipping, the more I'm convinced the future of embodied AI won't look like Atlas at all. Task-specific machines designed for particular jobs will almost certainly win out over general-purpose humanoids, at least in the near term. The economics are clearer, the training cycles are shorter, and you don't pay for arms when you only need a gripper. By the way, this pattern—the gap between what captures imagination and what actually gets deployed—keeps repeating across AI right now.

Speaking of deployment, I find the Microsoft situation genuinely revealing. The company is quietly ending most internal Claude licenses after just six months, citing unsustainable costs. This isn't a technical failure; Claude works fine. It's that when you give thousands of employees access to a powerful tool, they use it, and the API bills become brutally real. The broader narrative Microsoft is trying to manage here matters more than the headline. They've spent enormous energy promoting AI as a productivity multiplier, but the math is proving harder than expected—in some cases, using the technology costs more than hiring humans to do the work. That's a reckoning, not a setback.

Against that backdrop, the developments in AI agents and video generation feel almost orthogonal. ChatGPT's agent capabilities, which can now chain web searches, analysis, and actions across multiple steps, represent genuine progress on autonomy. And Google's Gemini Omni, their new multimodal model for video generation, shows the pace of capability growth isn't slowing. Yet neither directly solves the cost-per-task problem Microsoft is bumping into. A more capable agent doesn't help if running it drains your quarterly cloud budget faster than you can justify.

Here's what interests me going forward: will companies and labs start optimizing for efficiency per task rather than raw capability? Anthropic's valuation—apparently heading toward $900 billion—rests on a belief that bigger models and bigger safety investments create defensible competitive advantage. That may prove true. But if the real constraint turns out to be operational cost rather than model quality, we might see a sharp pivot to smaller, focused systems, better inference optimization, and local deployment. The robotics insight about task-specific machines over general humanoids might be the template for everything else too.