I think one thing that would tip this in a positive way is easy re-sharing of learning. I have plenty of "source code" now for conversations where I've taken so

I think one thing that would tip this in a positive way is easy re-sharing of learning. I have plenty of "source code" now for conversations where I've taken something that was hard to find out (or code that was previously hard to write) through a non-LLM route. Publishing those conversations isn't as easy or re-adoptable into the commons as it could be. (I'll note that proprietary models would be first-order disincentivised to stop this because such conversations can be used to better train other models, but of course they will benefit more generally for maintaining a commons of knowledge to draw from.)