The Cost of Overfitting to a Model's Way of Working

The trend of OpenAI scaling back fine-tuning is worth watching. As large models increasingly solve more tasks on their own, the need for weight adjustment decreases, but big labs are also increasingly embedding harness design tailored to a small number of use cases directly into the model.

In fact, in the OSS Pi harness, Mario Zechner tried to elicit certain behaviors from GPT, but Claude kept interfering with the process. This reveals a case where a model conforms to its own camp's way of working.

If this trend continues, the behavior of the 1st party harness may already be built into the model, diminishing the value of 3rd party harness. There would no longer be an escape route to resolve this bias generically through fine-tuning.

Ultimately, frontier models may become more like dedicated devices than general-purpose platforms. For some companies, building applications will become easier, but in exchange for reliability and consistency, lock-in will grow. The footnote brings to mind John Siracusa's Naked Robotic Core analogy.