Are neural net weights code or configuration settings? Are email connectivity parameters code or configuration settings?
How would a judge distinguish the two?
How would a judge distinguish the two?
The NN engine is just a substrate. It can't do anything on its own, and what it can do is vastly altered by the specific database of weights that are applied.
Additionally, as a proportion of what makes a NN system work, the weights are the vast, physical bulk of it.
I don't think any legal team employing a real ML expert witness would have any trouble arguing that the weights are the code.
Or like the bitstream to an fpga not the vhdl.
vhdl really is source, so it comes down to if nn weights really are more like vhdl or more like the bitstream.
It seems like the closest analog to a source for nn weights is training data and scripts.
In that case, it seems fine to license the weights seperately, just like the source to an image viewer or audio player, yet not some image that is displayed in it.
I'm not so sure about shipping them together under a single name as a single product. Maybe that act causes them to be obligated to provide the source to generate the nn weights?
I would love that, not because I value their data but because it would mean it was only their own attempt to mis-represent that now makes them have to give away their store. If they had simply packaged up their fork of Stockfish plus proprietary nn model honestly and in good faith, there would be no problem. Maybe they couldn't get $99 for that from very many people but, oh well!