Hector Martin on Nostr: You keep using the word "copyright" to talk about patents. You can't do that. The ...

You keep using the word "copyright" to talk about patents. You can't do that. The ideas in patents aren't copyrighted. They're patented. They are different things.

The question the whole AI/ML debacle revolves around is whether the "training" process erases copyright status. IMO it does not, because that also doesn't work for humans - you can memorize a book (which is nothing but "training" your neural circuitry) and write it out again and the result is still copyright infringement. The reason why humans can produce creative works that aren't infringing as a result of training themselves on copyrighted input is that we have the ability to reason about what elements are copyrightable and what elements aren't, and we train at a much higher level of reasoning and abstraction than state of the art GenAI.

Until AI becomes able to do that, there is a strong argument to be made that AI output is a derivative work of the training data. We also have a definition of copyright that requires "fixing" into a medium and human brains don't count, so memorizing something by itself is OK as long as you don't reproduce it. That doesn't apply to ML weights, which are "fixed" in the storage medium, making the model itself infringing.

And none of this has anything to do with my original post, because it's all about things that *are* copyrightable to begin with, and hardware isn't (nor facts in general), and patents aren't the same, and conflating them is meaningless.