The trouble with anyone trying to copyright their model weights is that it's likely to make legal arguments their model isn't a derivative work of copyrighted i

The trouble with anyone trying to copyright their model weights is that it's likely to make legal arguments their model isn't a derivative work of copyrighted inputs harder to sustain. Successfully copyrighting model weights (like databases or maps or other non-creative representations of factual inputted data) would be a very favourable outcome for creators of NNs who exclusively use licensed or public domain content, but not so much for a lot of the existing state of the art ML models...