[vlovich123]

> They redistributed the statistical patterns of those copyrighted materials. Which perhaps should be treated similarly nos

Nos for the same reason that me giving you a word cloud of the frequency of words within Harry Potter isn’t infringement. It’s a novel transformation.

[samrus]

Thats not as complex or scalable as what LLMs can do. The capability and scale is what changes the equation. Quantity is a quality unto itself here

[vlovich123]

Exactly, and the complexity of what they’re doing is an even more significant novel creation.

Particularly in the US there’s a four point test and the very first point:

> To justify the use as fair, one must demonstrate how it either advances knowledge or the progress of the arts through the addition of something new.

I don’t know anything that has advanced the knowledge and progress of the arts more.

> The third factor assesses the amount and substantiality of the copyrighted work that has been used. In general, the less that is used in relation to the whole, the more likely the use will be considered fair.

This isn’t about usage in training. This would be in the LLM itself - the copyrighted works are very rarely used in the output.

> The fourth factor measures the effect that the allegedly infringing use has had on the copyright owner's ability to exploit his original work

Would you sincerely claim that owners have become less able to make money because of LLMs? Those same owners using LLMs to increase their own output of copyrighted works?

Anyway, copyright is not an absolute right and you have to really misunderstand copyright law to claim that LLM training infringes it.

https://en.wikipedia.org/wiki/Fair_use