I'm with you. I think we will need less "frontier" models in the future, but we'll reach for more specialized ones. A small, focused model trained to e.g. coding is within the realm of useful reality. JetBrains is making some moves in the field, and laudably, some of their work is open-weight.