Robot Orchestration Gets Modular. AI Gets Expensive.

The robotics world is thinking bigger this week. Open Robotics just published a roadmap for next-generation Open-RMF-a framework for orchestrating multiple robots in shared spaces-and it's not just a polish job. They're redesigning the whole thing as modular pieces (traffic planning, task planning, execution, dashboards) that can be swapped out independently. That matters because it means a hospital deploying robots doesn't have to commit to one vendor's traffic system; teams can mix and match implementations. The staged rollout starts with traffic management, but the vision is clear: this framework wants to work at factory scale, airport scale, even city-scale operations.

The Real Cost of Cheap Tokens

Meanwhile, the AI world is hitting an uncomfortable wall. Token prices have never been lower-OpenAI's API costs dropped dramatically-but demand is climbing 2.5x faster than supply can scale. That's the Jevons Paradox applied to intelligence: cheaper compute creates more demand, which creates a bottleneck, which shows up not in price but in availability. Anthropic is cutting session limits for Pro users. Google's ancient TPUs (7-8 years old) are running at full utilization. OpenAI is turning down opportunities. The economics look good on paper-more volume-but the squeeze is real, and it's hitting users hardest.

Tools That Actually Work (And Honest Takes When They Don't)

On the builder side, there's healthy skepticism emerging. Theo called out Claude Code as unusable, and rather than generic cheerleading, we're seeing detailed technical critiques. Meanwhile, practical workflows are shipping: screenshot automation that stays in sync with your codebase; RAG systems designed for production (not demos) with reranking layers and semantic chunking; intent-based frontend design that reduces clicks in favour of understanding what users actually want. Abacus's new CoWork tool points somewhere real: taking messy, mixed-format data (receipts, PDFs, transcripts, Jira tickets) and turning it into finished work-audits, timelines, compliance answers-with structure and reasoning.

The common thread isn't hype. It's specificity. Automation that solves a real problem (eliminate screenshot drudgery). Systems designed for production constraints (reranking, chunking strategy, monitoring degradation). Tools that do one thing well and integrate with what you're already using. That's where the momentum is.