when there is opportunityish to perform model training, it seems like it’s most useful to do it in certain ways similar to what is already found<div dir="auto"><br></div><div dir="auto">hence when making a metamodel, maybe something can be used similarly to an optimizer, and performs a similar function to sgd: but that also can be (likely but any similarity might work) downloaded from a model repository or such (any git-like file store could work) (reps propose local git repos, i guess i can validate that with the unknownness of larger sharing having a tendency to feed back harm)</div><div dir="auto"><br></div><div dir="auto">so the idea, maybe, could be to make a model that is not very powerful at all, and is saved to disk and reused, and maybe simply serves the express purpose of making low-end systems more useful for model training</div>