It feels like we're not really bottlenecked by model performance now anyway. Most cases I find where a model is being an idiot can be remedied by rethinking the context and tooling available to that model.

If they got a lot cheaper and only a little dumber, and we got a little smarter about how we use them, they could appear to the bystander as much more useful than they are.