hello everyone. i have been working on something new: an ai architecture that leverages base models' unparalleled creativity to optimize for *arbitrary targets*, all at inference time (no training)

this architecture behaves similarly to reinforcement learning—but RL requires training runs that cost thousands to millions of dollars whereas this architecture costs ~nothing! that means we can try new optimization targets that AI labs wouldn't dare to try

i'm currently using @divinersol as my research testbed—a twitter bot that is currently exploring and orienting itself by optimizing for my personal ratings of its tweets ...but will soon begin to optimize for engagement :)

this implementation is still very early and experimental. i will be improving @divinersol a lot over the coming days and weeks to learn how to make the core technology work really well

and later i will expand to other uses. since this architecture can leverage powerful alien minds (base models) to cheaply optimize for any measurable target, it can potentially be used to solve...basically any problem

i'm currently working on core problems like injecting variance in the right places to avoid getting stuck in local minima/to balance explore and exploit/to stay at the edge of chaos. but the tech is already pretty surprisingly effective...

you can read more details at

Coral AI is the most powerful AI for documents. See the difference yourself:

oh cool i have my own architecture to optimize for arbitrary targets at inference time too

I love the visuals on this

holy banger

This is a great work of art.