Post #3700713
2026-07-09 18:54 UTC
@scottgal@hachyderm.io Agree with the points in your post. Also, there's definitely a lot of subtlety to what a 'summary' really means.
I was mostly doing the naive approach you called out for a laugh. Would be interested to see what your better pipeline returns for a 10 (and 1) word summary of War and Peace :)
https://raw.githubusercontent.com/mmcky/nyu-econ-370/refs/heads/master/notebooks/data/book-war-and-peace.txt
Replies (1)
-
@scottgal@hachyderm.io 2026-07-09 19:02
@Lee_Holmes@infosec.exchange So not well; mine reduces to segments so it would find the most representative sentence of the content of the book....TON of work and problably just. a term like 'Russia'. The problem is narrative extraction isn't a solved problem. What an LLM would do is look at other people's summaries most likely so it wouldn't even really TRY wiuth a public document like a book. What HUMANS do is build internal graphs of locations, relationships, themes etc...it's massively complex really.