UP | HOME

Landing Page

Table of Contents

1. Welcome 🔗

A digital garden is an online space at the intersection of a notebook and a blog, where digital gardeners share seeds of thoughts to be cultivated in public. Contrary to a blog, where articles and essays have a publication date and start decaying as soon as they are published, a digital garden is evergreen: digital gardeners keep on editing and refining their notes.

– Anne-Laure Le Cunff

Welcome to this digital garden about making machines do what you want. Currently my goal is to build a system that is powerful and aligned enough such that it allows us to perform a pivotal act. More on this digital garden.

This document is the best writeup of the "Understand Intelligence" agenda.

The site contains anything from polished writeups, drafts, inqueries abandoned halfway through, raw written stream of thought, to todo lists.

2. Website Features 🔗

2.1. Embedded Games using WebGPU 🔗

This website uses WebGPU to create interactables. Here is an example:

Above you should see something like this:

2025-01-04_12-32-28_screenshot.png

If you don't see anything your browser likely doesn't support WebGPU. To make sure my website isn't just broken, I recommend to also check if you see a red triangle here.

See the implementation status for what you need to do to make it work. Most of this site, and most of the pages that use WebGPU should work without WebGPU support.

See Motivation for learning Computer Graphics for why I am using WebGPU.

For more on the above example see Game of Live Codelab.

2.2. Dark Mode: Using Dark Reader on this Site 🔗

Instructions to make all the math on this site render properly when using darkreader in dark mode on this site.

Open darkreader -> more -> all settings -> Advanced -> Dev tools -> Dynamic Theme Editor -> Per Site Editor, and then enter the domain of this website in the single line text field, press create new fix and make the large textbox contain the following text: 

DOMAIN-NAME

INVERT
img[src$=".svg"]

3. "Automata Align" About this Name 🔗

Automaton means "self acting". It's a thing that automatically acts on it's own. It some sense it's the opposite of a puppet.

"Automata Align" is the magic spell that I want to be able to cast. More specifically, I want to figure out how to make automata do what I want. The holy grail is to understand how to make an artificial superintelligence running on modern computer hardware be alinged, but it is not limited to this. Infact it's not even limited to AI.

In order to solve hard ploblems you need to align your brain. To be effective you can't feel aversion towards doing the things. You can't hate what you do. And if you want to solve hard problem you must recruit your subconscious to work on them. Conceptualizing the brain as an automata (which it actually is) helps resolve issues like this.

How good are you at controlling computers? To me it feels unreasonable to expect to be able to conjure a superintelligence that does what you want when you're not able to conjure a simple spirit that sorts a list. Conceptualize an aligned superintelligence as an algorithm to understand. I find this is a useful framing. The general skill of understanding and manipulating automata of any kind is useful.

Everything in technical AI alignment ultimately needs to cache out in algorithmic knowledge.

I use the term automata to broaden the scope. Learning how to align yourself, or learning how to write simple programs often can teach you something relevant for aligning a superintelligence. How could it not?

Automata is the plural of automaton. A spell that allows multiple superintelligent automata to coexist and do what I want seems more powerful than one that works for a single automata.

If "automata align" is the spell, it is a spell on spells. An algorithm that can make an algorithm be aligned, which may include creating or modifying a automatas. I may run this algorithm on my brain or execute it on a computer. Each time I manage to make a computer do what I want am running some version of this spell. Though usually a much weaker one than the one I am trying to find.

4. About Cossposts on LessWrong 🔗

Some of the content on this site are crossposted to LessWrong. This is indicated at the top of a page or section. Consider commenting and voting on the LessWrong post. The Lesswrong crosspost may only be a stub. In this case especially, consider quoting the relevant parts of the document on this website when commenting on LessWrong. Ideally include a link to the relevant heading. Example quote:

Some of the content on this site are crossposted to LessWrong. – From About Cossposts on LessWrong

5. See Also 🔗

5.2. Older Research Artefacts 🔗

Previously instead of the Research Session Notepad and Research Session Log pages we used typst:

Before that we used this google doc:

  • The research artefacts listed here.

Before that we used the eraser boards listed here:

6. Sitemap 🔗

A list of all documents on this site in anti-chronological ordering (based on last editing time).