Match Model Size to Task
Dave Rogenmoser, CEO and co-founder of Jasper, on the generative AI opportunity
The key point is that application companies win by matching model size to the job, not by defaulting to the biggest model available. For narrow, repeatable tasks, a smaller or task specific model can be better because it is cheaper, faster, and easier to slot into a product workflow. That means fine tuning can add value early, especially when the job is concrete, like matching content to a brand voice, extracting fields, or classifying inputs inside a fixed business process.
-
Jasper’s product logic already pointed this way. It started with broad GPT-3 generation, then used user actions like save, favorite, and copy as training signals to improve outputs for specific writing workflows. That is less about building a giant general model, and more about making one job work reliably inside marketing software.
-
Copy.ai described the same pattern later in enterprise workflows. For jobs like classifying a website into exact CMS categories or extracting growth themes from earnings calls, the valuable thing is not maximum general intelligence. It is a model tuned to one repeated task with clear inputs and outputs.
-
The economics reinforce it. Large frontier models are still the best fit for broad, open ended reasoning, but providers now explicitly price smaller models as faster and cheaper for well defined tasks, and position fine tuning and distillation as ways to push those tasks onto lower cost models.
This heads toward a layered stack where the biggest models handle hard edge cases, while smaller customized models take over high volume production work. As AI products mature, the advantage shifts to companies that can turn user workflow data into many narrow, reliable models that cut latency and cost without giving up quality.