[fire]

I don't understand why they aren't tagging data with license information and allowing users to use models that don't include certain licenses - seems like it would be the middle ground given the stance they've taken; like, "we don't think it's a problem, but if this makes you feel better you can use these other models that specifically don't train on gpl code, or whatever"

I would prefer to see full license attributions included in generated responses, though. Something that then also wouldn't be that difficult to generate a licenses file from?

Amazon's CodeWhisperer has a "reference tracker" that tells you the license of training data code if the generated response is within some similarity threshold, but that's still not good enough imo.

[JoshTriplett]

> I would prefer to see full license attributions included in generated responses, though. Something that then also wouldn't be that difficult to generate a licenses file from?

Exactly. By all means build tools like this, but build them to actually comply with Open Source licenses. Provide a list of the licenses you don't mind copying from, and get back attributions with your suggestions.

[danuker]

Suppose Copilot offered some pure-MIT licensed flavor.

Copilot could comply with MIT licenses by just outputting an MIT license with ALL the authors of code used in training.

[JoshTriplett]

That'd be a valid solution, if impractical. I doubt that people would be willing to copy hundreds of thousands of license notices into their project.

[danuker]

One perspective is that those authors actually contributed to the end result.

But sure, disk size could be a problem.

[JoshTriplett]

> One perspective is that those authors actually contributed to the end result.

They absolutely did, yes. The approach you're suggesting would work from a legal perspective, but the size might pose practical problems.

[LelouBil]

> Amazon's CodeWhisperer has a "reference tracker" that tells you the license of training data code if the generated response is within some similarity threshold, but that's still not good enough imo.

I don't think it's possible to do better than that with this technology.

[fire]

like I probably don't understand this in the right way, but I could have sworn we had the ability to probe latent space on models like these and make mappings based on them? Or was that only for diffusers?

[danuker]

Sure, you could build an index of attribution <> latent space coords, but it would not be clear whether a generated document near several index entries would require compliance.

I guess this is where the threshold comes from. Choose a generous margin and over-attribute rather than under-attribute.