The actual research paper https://arxiv.org/abs/2603.18030
Also it pointed to where it is on github which I think means its possible to just tell chatgpt to tell me how to run this on my PC and then play with it, looks like you can just toss in anthropic/openai/etc api keys and go on your way with a POSIX controlled ai managed by any os https://github.com/kehao95/quine
Oooh I just realized you could probably schedule it to fetch content for yourself, if a web browser lets you use it via api. This would be like the chatgpt feature that is basically scheduled prompting but since this lets you mess with real hardware and software you can have a shell command that lets you fetch a “web surf internet rabbit hole”
This would be the way to avoid algorithmic nonsense because you can specify or have it infer your values based on content on your PC and it would do the job of an algorithm but with asuch or as little control as you want and not be beholden to any platforms
#!/usr/bin/env bash set -euo pipefail
mkdir -p “$HOME/digests”
cd “$HOME/agent-runs/web-rabbit-hole”
export QUINE_MAX_DEPTH=2 export QUINE_MAX_AGENTS=4 export QUINE_MAX_CONCURRENT=1 export QUINE_MAX_TURNS=20
quine " Fetch recent content from my chosen sources. Look for threads related to:
- interspecies communication
- lucid dreaming / dream yoga
- NixOS
- AI agents as OS processes
- federated media
Write:
- a short digest
- the best 5 links
- one rabbit-hole path worth following
Save it to $HOME/digests/$(date +%F)-rabbit-hole.md. "
Why do we need to know this?
It turns the useful parts of ai into normal boring computer scripting so it can be used more effectively as an agentic helper without the overhead of usual software tools.
The use case I laid out was basically a way to avoid getting algorithmically addicted and being able to use it broadly across multiple platforms which is a problem that is basically not solved rn unless you just don’t use some parts of social media.
But you can also use it for various other mouse and keyboard tasks to save some time for yourself that aren’t dependent on personal input like say checking some logs or reminders or organization tasks, so basically agentic bounded boring disposable task autocomplete for the people who don’t like ai being shoved into their personal thinking workflows.
You need to put a short explanation in the beginning of the body of the post per community rules.
I figured the 3rd paragraph does that but it is a fair criticism that it isnt the first paragraph.
I have it sort of reversed from normal where I have links and reasonigns first rather than after stating the premise.