Hacker News Reader

Comment by Bjorkbat | Hacker News Reader

Oh, it’s very much possible to create AI models with copyrighted materials. The “impossible” part is turning said AI model into a cash cow without fairly compensating the original rights holder.

View on HN

Does that change with the stipulation of "with the same performance?"

I imagine it'll be the same performance whether they stole it or paid 10 billion dollars from all consenting parties. Is there some singular source of information that is so important to scape for the sake of humanity that is unwilling to license?

what if a new ai model retains any info fed into it, such that I ask it about a book that I post excerpts into, and it remembers that for future conversations with me, and others. I'm not talking about using vector search, I mean literally it's folded back into the next 'version' of a self-evolving llm. You can't make an intelligence that functions in the real world without it knowing something about the copyrighted works out there. That's like trying to create a human brain without knowing any copywritten material, and have them fit in society.

In that case you are feeding the LLM and that changes the problem space signifigantly towards how/if an LLM can accept copyright material. Pretty much what Youtube deals with today,

In that case there would be some separate DB used to identify copyright material that it cannot use when generating art. Hopefully with drastic sanctions if it arises that it is in the LLM.

Alternatively, we can go back to literal Youtube and rely on DMCA yet again. I don't think anyone desires that.

Reply on news.ycombinator.com