I think a lot of ML research, despite being replicable is not necessary useful, because what we really want, approaches we can add in addition to all the other

I think a lot of ML research, despite being replicable is not necessary useful, because what we really want, approaches we can add in addition to all the other methods we have, that introduce meaningful improvements without unnecessary complexity and work across datasets, are quite rare. And that's assuming the evaluation was done well. Not to say the research is worthless, just that straight replicability is not necessarily enough.