isn't this the stuff llms are supposed to be good at? It seems like all of these "fixes" involve conventional code, not improvements in the models themselves.