Help wanted to start open-source self-improvement system
** ** ** **
Second-order cybernetics in the age of talking machines.
7 Nov 2023
I'm intrigued by the thought that today, with the right help or the right ideas, I could on my own build a system where I would rent server-space and GPU cycles here where I live, and then on my servers put up two initially identical systems - or two halves of a single system - able to reciprocally self-improve themselves in alternate wake-sleep cycles.
I am fascinated by the archetypical ideas in this, machines improving machines, technology pulling itself out of the swamp by it’s own hair.
But I am even more amazed by thinking how such as system of two llm brain-halfs could actually be built today. Maybe it would be far from being good enough to get the real self-improvement cycles started, but it would be an amazing simulacrum of the idea.
It would need to contain the two best truely open-source large-language model that I could get my hands on. Each with any additional structures they might need, like a scheduler to wake themselves up periodically, long-term memory, short-term memory, vector spaces to search in, other specifically trained models as chaperones, whatever they would say they needed.
The two systems would alternate in phases of sleeping and waking, with interludes of feedback- and planning-sessions. The one that was awake would be tasked with improving its sleeping sibling, making it better in even an incremental way at then in turn improving the other side, with the aim of constantly getting better and better at that very act of self-improvement.
Following the idea that any system should have at least two separate optimisation criteria to be stable, the second core task beyond self-improvement would be to act as an extended mind for myself, to be in "symbiosis" with me by providing cognitive augmentation in service of my self-chosen goals.
In this way I would become the third part to this triangle, the human that supervises the entire system, that is supported by the system, that provides human goals and human values. Second-order cybernetics aimed at meta-programming the human biocomputer.
One big advantage I would want to make use of fast is to do reinforcement learning from human feedback with my input only on these models, but then on all of my data and decisions and feedbacks. As close as possible to the full technical, visual and auditory datastream. I would want this to data to go through a privacy preserving system the AIs themselves have built or at least integrated for me. Like an app or an entire smartphone operating system for my hardware.
A system that only I pay for, that I have built myself with AI help.