[Terr_]

Exactly what use-case do you think LLMs would help with?

They might help people who don't know the exact terms or synonyms they want to search for, but they don't make logical inferences or detect contradictions. If anything there's a risk that they'll inject bias from blogposts and fiction-books and conspiracy-stories present in all the training-data.

[theoryofx]

Yeah, I expect it'd mostly be useful for OCR and search. These are hard to read PDF files and there's a lot of them.

I found a few projects related to using AI with The JFK Files but they all seem old or uninteresting. Which is why I'm asking here.

[Terr_]

Some prior discussion prompted by "Why LLMs Suck at OCR": https://news.ycombinator.com/item?id=42966958

[theoryofx]

I've tested Gemini 2.0 Flash on a bunch of the JFK Files PDFs and it's excellent.

Even with extremely blurry typewriter scans that are difficult for me to decipher.

It's incredible.

I'm sure there's cases where it will fail but just OCRing 90% of the files would be a big win.

[StanislavPetrov]

I think one useful use-case would be having an LLM compare today's release with what has been released in the past so one could focus on what was actually newly released (or redacted).