--- image: "/images/notes/lowell-mills-verification-infrastructure.jpg" title: "Lowell's Mills and the Shape of Verification Infrastructure" date: 2026-03-19 url: "https://x.com/ccatalini/status/2034617918198284401" tags: ['ai-agi'] deck: "In 1842, Lowell's textile mills scaled looms faster than weavers could check them. The same bottleneck is now the binding constraint on the AI economy." tweetCount: 14 likes: 220 reposts: 28 --- This is a great description of what verification infrastructure looks like in practice. In our new paper we argue this is the binding constraint on the AI economy — the same bottleneck textile mills hit when they scaled looms faster than weavers could check them. [x.com](https://twitter.com/rohit4verse/status/2033945654377283643) In 1842, the managers at Lowell’s textile mills had an idea so obvious it barely counted as one. Their weavers each ran two power looms. The mills had just bought more machines. Give each weaver a third loom, output goes up 50%. It didn’t work. The problem wasn’t physical — the machines did the weaving. The weaver’s real job was watching cloth get made: scanning for broken threads, catching defects, intervening before a flaw ruined an entire bolt. Add a third loom and the verification load exceeded human bandwidth. The mills had to cut loom speeds 15% just to keep quality from collapsing. It took a full year of retraining before weavers could run three looms at full speed. They’d upgraded the machines, only to discover the machines weren’t the bottleneck. Economist [@JamesBessen](https://x.com/JamesBessen) reconstructed this history. Once the power loom was in place, 62% of subsequent productivity gains came not from better machines, but from better-skilled humans who could monitor more of them. “They were monitoring.” By 1902 a single weaver ran 18 looms and produced 50x the output of a century earlier. But mills had to triple training investment — from \$47 to \$162 per weaver. Capital and skilled labor weren’t substitutes. They were complements. Today the loom is an AI agent. The cost to generate is collapsing. The cost to verify — to confirm it’s correct, safe, and not hallucinated — stays anchored to human judgment. One curve is a rocket. The other is a bicycle. The gap between those two curves is where economic value goes to die. My co-authors [@XiangHui](https://x.com/XiangHui) [@JaneWu](https://x.com/JaneWu) and I wrote about why verification — not intelligence — is the binding constraint on the AI economy. The weavers’ story is comforting: capital and labor were complements. But here’s the question we wrestle with in the paper: what happens when the verification layer itself becomes software? The problem with AI checking AI is that the errors become correlated. Same training data, same blind spots, same confident mistakes. You don’t get independent verification. You get a hall of mirrors that looks like consensus. That’s the path to what we call the Hollow Economy. Output explodes. Everything looks efficient. But the system is quietly losing the ability to catch its own mistakes—because the humans who could spot them were never trained, and the AI verifiers share the same blind spots. Meanwhile the pipeline that produces those humans is already severed. Juniors aren’t getting hired for the entry-level work that builds experience. Seniors are codifying their expertise into the models. The stock of human verification capacity is drawing down. The alternative demands real investment: accelerated mastery through AI that compress years of experience, synthetic apprenticeships that simulate edge cases at density no traditional job provides, and many more entrepreneurial and R&D experiments. The goal isn’t to compete with AI at execution. It’s to keep human capacity for steering and verification high enough that we actually know when the machines have drifted. The weavers figured this out. The question is whether we will. [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/03/18/babysitting-the-slop/)