Thinking Machines, the lab founded by former OpenAI CTO Mira Murati, published a mission sta...
Thinking Machines, the lab founded by former OpenAI CTO Mira Murati, published a mission statement this week arguing that the industry's default posture, renting one giant model to every customer, quietly devalues the specialized knowledge that makes an organization distinct. The post matters because it comes from an insider who helped shape the frontier lab playbook and is now publicly rejecting a piece of it.
Her counter thesis is that the winning pattern is customers owning and tailoring smaller models against their own data, workflows, and evaluations, rather than outsourcing cognition to a shared frontier endpoint. The framing treats a model less like a utility you dial into and more like infrastructure you accumulate, in the same way companies came to treat their data warehouses or search indexes as strategic assets rather than commodity rentals.
The technical implication is concrete. A rented frontier model absorbs your prompts, your corrections, your edge cases, and none of that compounds into an asset you control. A tailored smaller model, fine tuned or continually adapted on your proprietary traces, gets cheaper per token, faster to serve, easier to guardrail, and, critically, benchmarked against your tasks rather than MMLU. That last point is easy to skim past but does a lot of work. Public benchmarks reward general capability. Internal evals reward whatever actually pays your bills, which is usually a narrow slice of behavior that a smaller adapted model can master.
For builders, the shift shows up in the org chart of the stack. Frontier APIs still have a role, but only at the top of a routing tree. Below sits a portfolio of small, owned models per domain, each with its own eval set, retrieval index, and override log feeding the next training run. This is a very different operating model from the current default, where most teams treat the model as a single vendor dependency and pour engineering effort into prompt scaffolding rather than into training data curation. The Thinking Machines argument is that the prompt scaffolding is throwaway work, while the eval set and the override log are the real compounding assets.
The wider question worth watching is whether the tooling catches up to make this practical for teams that are not staffed like a research lab. Fine tuning pipelines, evaluation harnesses, and adaptation loops are still rougher than the API call they replace. If that gap closes over the next year, then the moat really does move from which model you call to the data flywheel you own around it, and the frontier labs will have to decide whether they sell access, sell training infrastructure, or both.
Originally posted on LinkedIn.