On Wed, May 28, 2025 at 5:28 PM Oliver Hunt wrote:

>
> The other is: “A person loads a paper into my AI and gets a description":
> This is functionally equivalent to the user puts a paper they have downloaded
> (and have the authors and original doc available) and your program runs sed
> and provides searching, etc.



My intention is even more simple than that. In the program, you would
type a question into a text box, a question which can be answered with
'yes / no'. So the question might be:

    "Does this paper make any mention of changing the behaviour of
R-value references if the type is const?"

or perhaps something like:

    "Does this paper make any mention of introducing a feature which
would replace 'std::forward' or otherwise make 'std::forward'
redundant?"

My program will load each of the 6000 papers, one by one, and ask this
question to each paper. So then the program will have 6000 answers
something like:

    no, no, no, no, no, no, no, no, no, YES, no, no, no, no, no, YES, YES, no

which of course I can just represent in a bitset:

    std:bitset< 6000u > all_the_answers;

Then the program simply populates a list box on screen with all the
papers that answered 'yes', and so you can click on one of the papers
and view it. The title and author of the paper will be right there on
the screen -- nothing redacted.

I'm not intending at this time to do anything more fancy than this
with the AI. I don't intend to reduce papers to a small summary or
anything along those lines. Nor do I intend to compare papers.
Although in the future I might the entertain the possibility of being
able to ask it "Looking at paper P1234, please point out the top 5
papers that are most similar to it", or maybe even let you upload your
own paper and then ask it "Pick out 5 papers that are most similar to
my own proposal paper that I uploaded just now". Maybe.

Six thousand papers won't fit inside the AI's context window which is
why I'll load in one paper at a time, ask the question, get the
answer, wipe the AI's memory, then load in the next paper, ask the
question, get the answer, and so on.

But now that we've been chatting about this for a while, I can see
that some people are uneasy about the whole thing. Version 1 of my
program, which I will probably release in a week or two -- definitely
by the end of June, will only use semantic search libraries like
Xapian and Manticore. All the AI stuff is on the back burner for now.