Yeah, a lot of people have reported significant gains from "model fusion" lately, which I think the literature just calls ensembles. You just ask a bunch of mod

Yeah, a lot of people have reported significant gains from "model fusion" lately, which I think the literature just calls ensembles. You just ask a bunch of models to do a thing and then you combine the results. Research tasks benefit the most from this. Because such work benefits from having a large number of agents working on a problem in parallel (i.e. crawling the web), and the model size becomes less important past a certain minimum. I don't know about other categories of work, like programming. I imagine looking for bugs or security issues would benefit from it.