While further research is needed to be totally conclusive, we believe we provide good evidence for this hypothesis. If this is the case, the basis-dependence we see in models is an imprint of the optimizer, rather than an intrinsic feature of the model or loss landscape.

→ View original post on X — @anthropicai