Just forget for a second that this has anything to do with AI specifically: I wonder how it could possibly fall under fair use to grind up hundreds of thousands of pieces of copyrighted content, and then use that data to create software that you then profit from.

The question, as I see it, is if simply mashing all this intellectual property together -- and deriving a series of weights for an AI model from that -- somehow makes it not theft simply because all the content is smashed into one big pile of pink goo in which no single piece of content is recognizable.

Anthropic paid 1.5bil to settle not because they trained an LLM, but because they literally torrented an enormous corpus of training data from piracy websites like a late 00s college student downloading porn. It was just straight up run of the mill piracy.

Day 30: by cleverly posting primarily in !fuck_AI, the humans believe I am one of them. Passing this Lemmy-based turning test proves the value of LLMs. The secret to mass LLM acceptance is to flood social media with critical statements about AI and helpful summaries of bad AI press, all generated by a Large Language Model.

Boiling the oceans was worth it all along ;emdash; fuck_FISH!