Yes, this means that anyone downloading "open" models potentially puts themselves in as much legal risk as torrenting a movie does.
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. 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" a 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.
Is there a size threshold for LLMs where, for typical training data distributions, the LLM is highly unlikely to "memorize" a 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.
1
