[AINews] Open Models, Model Labs vs Agent Labs, and What's Untrainable — Sarah Guo

Sarah Guo's latest Substack piece, 'The Untrainable,' builds a framework around legibility to identify what AI cannot commoditize. The core argument: applications earn durable competitive position by doing unglamorous integration work, arranging private company data so a model can act on it, then maintaining that translation indefinitely. Guo, a Cognition investor, puts it plainly: 'Integration and maintenance run as long as the relationship does, won by teams that put domain-specialized engineers and tools next to the customer.'

Three threads run through the piece that matter to anyone tracking the AI stack. First, open model adoption has flipped from bearish to real, validated by deployments at companies like Cursor and Notion. Second, benchmark scores are a trap: 'The most cited benchmark score of the year is a map of territory about to be worthless.' Labs that race to top leaderboards are signaling the exact capabilities that will be trained away next cycle. Third, the Agent Labs versus Model Labs distinction sharpens here. The value is not in the model. It is in the workflow, the tooling, and the customer relationship built around it.

The closing argument is the reason to read the full piece. Guo separates defense, knowing what to protect, from offense, knowing what to build in the first place. She finds the latter maybe three times a year. A model cannot help with that choice, benchmarks cannot measure it, and therefore it cannot be trained away. That is the actual moat. The original is short, precise, and worth your time.