If a problem improves directly with raw model capability, the labs will take it

Schmidt’s test for whether the labs will eat your product: does the problem improve directly with raw model capability? Code generation, writing, and image creation sit on the “Yellow Brick Road” — “every dollar spent on pre-training and post-training improves product quality,” so the labs are structurally best-suited and will own them. The “rest of Oz” is where value comes less from the model’s raw capability than from the scaffolding that makes output trustworthy, compliant, and operational inside a specific industry.

His strongest evidence is a revealed preference: OpenAI and Anthropic have announced forward-deployed joint ventures to configure their models for enterprises — “you don’t pour billions into those programs if you think the next model release is going to take care of it.” This sharpens Frontier companies absorb every useful agentic pattern into their products — the labs absorb what improves with capability, not what requires vertical scaffolding — and gives a concrete discriminator for Sand vs Stone — if models double in capability tomorrow, what washes away and what remains?. It feeds directly into The system of work is the moat, not the model — the model is fungible underneath.