Smaller Fine-Tuned Models Win

This points to data and workflow becoming more important than raw model size. For a company like Copy.ai, the valuable asset is not building a frontier model from scratch, it is capturing thousands of real user interactions, then using that feedback to train smaller models that handle narrow jobs like headline generation, rewriting, or classification with lower cost and better consistency inside a production workflow.

• Copy.ai had already moved from a single GPT-3 wrapper to 20 to 30 distinct fine-tuned models by late 2022. It was tracking whether users copied, saved, or rewrote outputs, then feeding those signals back into retraining. That means product usage itself becomes training data, and training data becomes product advantage.
  1 sacra 2 sacra
• Smaller fine-tuned models can win on repetitive tasks because they do not need to be good at everything. In practice, teams use a strong frontier model to generate or label examples, then train a smaller model for one job. That can cut inference cost by around 90% while improving speed and task reliability.
  3 sacra
• This is especially useful in enterprise workflows where correctness has to be exact, like mapping a webpage to the right CMS category ID or following a multi-step instruction every time. The market shift from prosumer writing tools to enterprise workflow software made that kind of reliability more important than having the broadest general model.
  4 sacra 5 sacra

Going forward, the winning application companies are likely to look less like simple AI wrappers and more like model routers plus data engines. They will use frontier models for broad reasoning, then push high volume, repeatable tasks onto cheaper custom models trained on their own customer workflows, which should widen margins and deepen product lock in over time.