Post

@paul The story is still copyrighted, but telling it to your family wouldn't count as a "public performance" so wouldn't infringe copyright. Telling it to a crowd at a park probably would, though.

Copyright of characters is complicated and varies by jurisdiction. That said, Mickey Mouse is in the public domain now, so your sketch is totally fine as long as you aren't trying to sell it or pass it off as legitimate Disney merchandise (because Mickey is still trademarked).

(Disclaimer: IANAL, this is just my understanding.)

@hopeless

> Huggingface still seems to be operating?

People get away with torrenting movies all the time too, doesn't mean it's legal.

Fair point on torrenting vs downloading though (because seeding), edited. But yes people do get in trouble for just downloading too. Depends on the country.

@yukijoou Yup.

@florian It would, if and only if:

1) You train only on compatible license content
2) You meet all the attribution requirements of those licenses (this is a big one)
3) Your weights are licensed under a compatible license themselves (usually the same for most copyleft licenses)
4) You understand that the output of the model may be copyrighted and require the same licensing and attribution (making the model unsuitable for, say, creative generation for publishing, but it would still be fine for a local voice assistant or something like that) and make your users aware of this.

Unfortunately, I haven't been able to find any models that meet any of those conditions, let alone all of them (other than KL3M which credibly claims to be trained on PD and non copyrightable content only).

@lina I don’t know if I will redistribute the vibecoded test suite I wrote. The only LLM-generated code I have published were trivial bugfixes that were clearly based on my own code.

@lina also it’s only legal if you’re a mega corporation or rich. /s

BTW, there is almost certainly a model size and architecture threshold here. If I train an image upscaler on copyrighted data, for example, it's quite improbable that it will be able to generate copyrighted data from an unrelated input (for typical upscaler designs and sizes). Most likely that is safe.

Is there a size threshold for LLMs where, for typical training data distributions, the LLM is highly unlikely to "memorize" any copyrightable information? Is that size threshold a usable LLM? These are open questions (that nobody seems to be interested in researching).

I'm willing to bet if such a size threshold exists, it's much smaller than the minimum LLM size useful for "general purpose" prompting though.