News for hackers. Stuff that matters.
Re:AI Is Already Killing Books by macilacilove on Saturday January 06, 2024 @09:49AM (#38891835)
Current LLMs can not summarize books even if they are in their training set. It is only possible if the training set also includes human summaries.
Re:AI Is Already Killing Books by apwell23 on Saturday January 06, 2024 @10:37AM (#38892263)
> Current LLMs can not summarize books even if they are in their training set

Why is this? due to legal reasons?

Re:AI Is Already Killing Books by macilacilove on Saturday January 06, 2024 @11:30AM (#38892757)
Most importantly due to the context length not being long enough. If the context length was long enough, it is possible that they could do it with clever training. I only trained much smaller language models though.
Re:AI Is Already Killing Books by avandekleut on Saturday January 06, 2024 @11:13AM (#38892602)
The context windows aren’t large enough, as I understand it. It might be possible via a chain-of-summarization, though.
Re:AI Is Already Killing Books by dTal on Saturday January 06, 2024 @01:16PM (#38893701)
"Even if" is an odd phrasing. Why would you expect that to be possible? LLMs don't have direct access to their training set. Having a book in the training set just means they emit text in the style of the book - it's not like a database they can query. But they can summarize very well, if the text is front of them.
Re:AI Is Already Killing Books by macilacilove on Sunday January 07, 2024 @05:05AM (#38899967)
I mean, currently you have to wait for a human to summarize a book, before you can read an llm rephrasing of that summary. Or you have to buy the book and pass it to the llm employing some tricks to avoid reunning out of context length.