One of the references for the blue dot impact AI alignment course is the blog post by Scott Young titled Do The Real Thing. In this he draws a comparison between doing the fake work of watching youtube videos, reading papers, in some cases getting degrees, talking about doing the work, reading blog posts (alright he didn’t really include that), and actually doing the real thing.

Through much of this course I have done the fake work and become conversant with AI alignment topics, but now is the time (I hope it isn’t too late) to do the real thing.

This post is my brainstorming of how to spend the next 2 hours to make progress toward this goal.

Fact: I have a vetted idea of how incorporating “world models” into AI fine tuning can improve the overall knowledge of the AI.

Fact: I have a couple of examples of toy “world model” problems that I could use to test this idea out:

  • AI Tech Adoption
  • Fishery Management
  • Kid dropoff at school in a pandemic

I think the best goals would be to:

  1. Get the AI tech adoption model working in agentpy
  2. See if I can get an LLM to output configurations for this model and accept end states.

That is what I will do today then.