try.works/first-principles-of-model-routing

haha

that runs before the big model

so we need a tiny classifier

and cheaper

but faster

like nlp classification

and classify it

maybe we start with the query

idk

like in 100ms or less

but how do we do that

duh

understand the user’s intent

of model routing

like what’s the first principle

what’s the best way to start

okay but seriously

lol

impossible

but also accuracy

so we need to optimize for speed

and gets frustrated

the user just sees a spinner

if routing takes too long...

what about the user experience

everything does

but also adds complexity

that’s a good idea

hmm

that this model is the right one