I agree. We don't need a nondeterministic 10quadrillion vector model. We need an deterministic expert on our narrow business. Something small, that can be run on the 2026 version of the spare PC under the CTOs desk.

It will always be worse compared to a centralized approach where hardware utilization can remain high. Except in case which demand low latency which most development things do not need. It's okay if it takes an extra 100ms for a code review to take place.