Building AI Systems That Actually Work; Robots Moving Into Production

There's a quiet revolution happening in how people are building AI systems, and it's nothing like the glossy product announcements. A farmer in South Korea with one Android phone and no computer has spent two years orchestrating multiple AI models into something he calls an agentic operating environment. His system-garlic-agent-doesn't just ask AI nicely to do things. It forces verification, handles failures gracefully, reverts when things go wrong, and logs everything. The core insight: don't trust what AI says it did; verify what it actually executed. He's not alone in this discovery. Across the industry, developers are learning that the real work isn't prompting better-it's building structure around unreliable components.

On the robotics side, money is moving fast. Oxa just closed a $103 million Series D for industrial mobility automation, with backing from the UK's National Wealth Fund and NVIDIA Ventures. They're putting self-driving software on 20+ different vehicle types-forklifts, tow vehicles, shuttles-turning factories into autonomous logistics networks. Meanwhile, Machina Labs is reimagining manufacturing with AI-powered robotic factories that switch products at the click of a button. These aren't experiments anymore. Companies like DHL and bp are already using these systems. The production floor is becoming genuinely flexible, and that's a bigger shift than most people realize.

The Open Source Reckoning

But there's a darker thread running through the week. Theo raised something that's been brewing under the surface: AI has fundamentally broken open source. When coding agents can ingest entire repositories and train on them without proper attribution, the economics of maintaining open source software collapse. Developers pour hours into projects that become training data for closed systems, with zero compensation and no control. It's not just a licensing problem-it's a motivation problem. If your work disappears into a black box the moment you publish it, why would you keep publishing? The counterargument exists: open source was always a gift, and AI just magnifies its value. But that gift only works if maintainers stay motivated to give it.

There's something underneath all of this worth noticing. The farmers and developers building real AI systems are discovering the same thing: structure matters more than capability. Anthropic shipped 1M context windows in GA this week-welcome to 2024, apparently-but the real conversation has moved past "how big is the model" to "how do I make it reliable, verifiable, and aligned with what I actually need." Skills are becoming standardized. Agents are getting persistent memory and self-improvement loops. Manufacturing is becoming software-defined. And open source is being reckoned with, finally, honestly.

The afternoon edition covers the practical work being done in the margins-the systems people are actually building and relying on-rather than the press releases from the front row.