> But for most tasks, a fleet of tailor-made smaller models being called on by an agent seems like a solidly-precedented (albeit not singularity-triggering) bet.

intuitively it sounds akin to the unix model...

But then again the main selling point of using LLMs as part of some code that solves a certain business need is that you don't have to finetune a usecase-specific model (like in the mid 2010s), you just prompt engineer a bit and it often magically works.