Enterprise AI splits the work: knowledge in the prompt, behavior in the model — resolves 2027-07-14

The debate: to make a general model understand a business, you can look up the relevant documents at question time and hand them over (or use a very long input window), or you can retrain the model on your data (fine-tuning).

Why it matters: the two routes differ sharply in cost and upkeep, and because knowledge goes stale, keeping it in the prompt lets it be swapped out while writing it into the model means retraining.

Who is on each side: one camp expects one route to swallow the other; SecondSource reads the two as coexisting, split by whether what you want to change is knowledge or behavior.

SecondSource’s reading: enterprise customization will divide the work rather than have one approach replace the other. Knowledge goes on the prompt side, through retrieval-augmented generation (looking up relevant documents and handing them to the model) or long input windows, which keeps it updatable and its sources checkable.

Behavior and format go on the weights side, through fine-tuning, which makes them stable without restating them in every prompt.

Two independent 2026 papers each showed one side ahead in its own setting, which we read as a division of labor, not a win for either.

arXiv (Lewis/Perez/Piktus/.../Riedel/Kiela; FAIR/UCL)
“models which combine pre-trained parametric and…”

Check in July 2027 whether the division-of-labor story still holds: if mainstream enterprise customization is retrieval plus prompting, with fine-tuning limited to behavior and format, the call is a hit.

What would overturn it

Weight-level knowledge writing, of the Cartridges type, is reproduced in enterprise settings at a lower cost than retrieval.