Hacker News Reader

Comment by gkbrk | Hacker News Reader

You can't extract content from a SHA-256(book), but you can extract some content from LLM(book).

Can I, though? In my tests with ChatGPT I haven't been able to get it to regurgitate books or even small portions of them.

Yep, you can. I tested by typing the first 13 words of the book Frankenstein and GPT3.5 happily completed the rest. Feel free to try.

Here are screenshots of the completion and the diff with the real book.

https://ibb.co/album/sJBB11

That's a public domain text. So, yes, it can do it. I stand corrected there. But can you coax a copyrighted text out of it?

When I ask it for anything from Game of Thrones it refuses beyond offering a summary.

But this is all begging the question.

How are you going to codify that? "You can't feed my text into an algorithm that has the ability to reproduce it"?

I also tested this with Harry Potter books (which are still not public domain AFAIK). It starts auto-completing the correct words quickly, and then hangs and eventually stops producing output. You can call the API to generate more tokens, but it stops again after producing a few more correct words.

I think for a few high-profile authors (for books like Harry Potter and Game of Thrones), OpenAI probably installed some output filters in order to not get sued too hard. Of course, I can't definitively check that without access to the raw model. Which OpenAI conveniently doesn't provide.

Reply on news.ycombinator.com