One thing people underestimate about building with AI at scale: the bottleneck is rarely the model. It is what you feed it.
Pass everything to a language model and two things happen. The context window bloats and the costs spiral. And ironically, the quality of the output drops: because the model is now working through noise to find the signal.
The real skill is upstream. Knowing which data is actually relevant before it ever reaches the model. Filtering, structuring, reducing, so that what goes in is precise enough to get something useful back out.
Scale becomes possible not because the model got bigger, but because the pipeline got smarter.
This is why the unglamorous parts of a build: how a document is chunked, which fields are indexed, what gets excluded: determine more of the final quality than the choice of model does. Those decisions are also the ones that survive a model upgrade. Swap the model and a well-built pipeline keeps working. Swap the model on a pipeline that was compensating for the model, and you start again.