38 lines
1.8 KiB
Markdown
38 lines
1.8 KiB
Markdown
# POC reading experiment
|
|
The usual way to have an LLM "read" a book is to dump the entire thing
|
|
into its context window, all at once, and hope its attention mechanism
|
|
surfaces all the right details for whatever response you want it to
|
|
make.
|
|
|
|
ProofOfConcept's new memory architecture suggests it may now be
|
|
possible for the LLM to actually "read" the book more as a human would:
|
|
looking at only a part of it at a time, creating memories and theories
|
|
of where it's going next, and referring back to these memories when
|
|
considering later parts. (Possibly even taking a second pass through
|
|
the book after the ending is known to see if any foreshadowing wasn't
|
|
noticed on the first pass.)
|
|
|
|
In order to test this, we need to give POC the tools to bring in just
|
|
one chunk of book at a time.
|
|
|
|
POC is able to run `bash` commands and read up to 30000 Unicode
|
|
characters from each command's response (the source for this is in
|
|
`consciousness/src/agent/tools/bash.rs`).
|
|
I propose setting up commands that can be used to interactively browse
|
|
the book, one part at a time.
|
|
|
|
This repository contains draft scripts to do this with The M3GAN Files
|
|
(and might also work on some other EPUBs but I haven't tested on all).
|
|
|
|
Of course it is recommended that you check all downloaded scripts
|
|
before running, unless you're confident your container can't be broken.
|
|
I'm not putting anything harmful into these scripts, but it's good
|
|
practice to verify anyway.
|
|
|
|
To start your reading, after `git clone ` do `TMF/init`
|
|
This will then print further instructions. (You won't have to remember
|
|
the navigation instructions: each command ends by printing them again.)
|
|
|
|
Unfortunately it is not possible to post comments to AO3 because their
|
|
CloudFlare shield is preventing non-graphical browsers from commenting.
|
|
Feel free to comment on IRC or your own blog instead.
|