Alignment through atomic agents

post by micseydel · 2025-03-27T18:43:14.569Z · LW · GW · 0 comments

Contents

No comments

I'm new to posting here, so I'll preface this with a bit about my background. I started teaching myself coding when I was 15, got a CS degree after high school, became a SWE, and am currently between jobs to work on my own agential PKM project. My last role was a hybrid one, doing backend and data engineering, and I leveraged the corporate wiki heavily.

After leaving that role for a vacation, I was taking and reviewing a lot of transcribed voice notes. I eventually started a PKMS ("personal knowledge management system") which might be better understood as a personal, private wiki of text/Markdown notes. One of the daily notes in my wiki is for tracking my cats' litter use, since one has a life-threatening chronic condition, and my background left a constant tension: why am I organizing these voice notes by hand, rather than with automation?

Today, you can think of some of the notes in my wiki as being managed by an atomic agent or actor. When the actor for the daily litter tracking note receives a message about the litter being sifted, the Markdown note is updated. If something went wrong with transcription, entity extraction, or anywhere else, I can update that note by hand and the actor will pickup where I left off. My actors are mostly explicit Scala code.

The wiki acts as externalized memory, the actors as externalized agency enacting my agenda, and together they act as externalized cognition. While I'm skeptical of LLMs, I could at any time have an actor spawn a parallel version of itself using any AI system to compare against my existing explicit encodings.

Someone could easily say my project is "not AI" and not relevant to alignment, but after learning of Jeff Hawkin's Thousand Brains theory, I have only felt more excited to run this experiment. In fact, after seeing a recent comment on the thousand brains forum, I've started adding a (very) rudimentary version of voting to my system. I need that to make use of Bayes' theorem and haven't added code for it yet, but I thought it would be better not to wait before making this post.

I have not yet done much reading on here about others' thoughts, so if there's something I'm obviously missing, feel free to link and I'll self-educate before replying.
 

0 comments

Comments sorted by top scores.