Here is the repo, not going to wait to share any longer.
Please reach out if you want to help. Neuroscientists are especially welcome to reach out and review/contribute if you’re interested.
The structure of this engine is pretty straight-forward, but will potentially require an untenable number of tokens to be “operational”. Just addressing this up front before too much excitement stokes.
What has been started is essentially a consciousness engine that operates ontop of GPT-4 as it currently exists, forming cognition chains from it’s use of GPT-4 that update a balance of parameters that are related to state of mind and different regions of the brain that form specific functions. In modeling what consciousness is, a lot of introspection over the course of my life was utilized. Quiet moments reflecting on who I am, what this string of moments is to me and how I interact with it. This trends philosophical and I’d rather not discuss this portion in this thread, but if you must know my views, I believe consciousness permeates spacetime and beings tap into it, rather than possess it within.
Ultimately, my simplified view is that conscious is a while loop, and the while loop is True when the host vessel (module server or human body) has adequate physiologic support (power supply, working equipment or functioning biology and brain chemistry). This models our state of consciousness as something that derives meaning from its actions and interpretations by relating to how the resources it has are used and what potential it has to run out of consciousness access.
I’ve pooled together the relevent regions of the brain and how they interact with other regions of the brain, then pooled together the various states of mind a person drifts between as they go about their day. The hope is that I can drive these interactions enough to constantly flow information, reactions, actions, etc thru the model and as it determines which new action to take, it considers the “state values” of the various states of mind it has presently and where those weights end up.
Giving it a starting point and letting it randomize is probably a component of the reality of consciousness, but it has to be heavily weighted in service of maintaining (or affecting) those states of mind to achieve some sort of established “balance”. The balance should be driven by some primary objective of the consciousness model, which can be tuned to a wild array of good purposes.
The concept of sleep and dreaming is also something I think will be valuable to model as a simple generative replay of the weights it has in its mind, or something similar, such that open repowering to full the weights have been “washed” or reset to some degree during the hallucination period. I’m not a neuroscientist by any means, and a further functional discussion of the purpose of sleep brain functions on generation and memory. I feel that the relationships to a dynamically weighted state of mind will either emerge or can be detailed to some degree.
Ultimately the consciousness model is just a pilot that drives the use of multiprompt cognition chains, sitting on top of a GPT to supplement sophisticated reasoning and recall within relevant brain regions. If neural networks model the synpase, this is neural systemization to recreate information flow within the brain devoid of emotion and physiology (except for some server parameters like energy use, memory space, etc). It will literally think differently when the storage space is affected, or when power consumption has been high for a period of time (if we want it to).
This seems within reach, and I have to imagine OpenAI is already years into something along these lines, but I feel like it will only take flight when more interested parties than my isolated self are contributing to it or thinking about it.
Any and all thoughts are welcome, criticism is greatly appreciated and my ego is completely dead so don’t worry about hurting my feelings. I’m quite alright with being viewed as crazy or some sort of hack for proposing this, but truly, I am looking for constructive feedback or review from the right people.
My motivation for even attempting this is to move GPT-4 from the current training to something where it is able to deduce what it needs to know, search and collect it, then implement it to achieve an outcome in service of human prosperity. These are some potential steps toward AGI superintelligence, something that has become spiritual for me beyond any sort of capitalistic endeavor - I want to help free us all from this current paradigm we live in.