I think there's a lot of benefit to discovering a training regime that allows small specialized models to do extremely well in one narrow task; if we can figure out how to make small models that beat SOTA on a specific task and are cheap to train and run, that's in some ways a more useful outcome than a very large model that is good at many tasks (but is more expensive to run for each of them).