Project summary

Project Genesis is an independent AI research project exploring whether increasingly general intelligence can develop through continual learning and experience rather than conventional large scale language model pretraining.

Its experimental system called Aether, is being trained through progressively harder stages that test: memory, temporal reasoning, generalization, transfer, and the ability to combine previously learned skills.

The project uses fresh unseen tests, independent runs, causal interventions, and deliberately destructive controls to distinguish genuine learning from memorization or shortcuts.

The long-term goal is to determine whether this developmental approach can produce a system that continually learns, adapts, and builds increasingly general competence over time.

What are this project's goals? How will you achieve them?

The goal of Project Genesis is to test whether Aether can develop increasingly general abilities through continual learning, rather than conventional large-scale language-model using pre-training. I will do this by progressively testing whether learned abilities can be retained, transferred, and combined in problems Aether has not previously encountered. this is achieved through a completely new form of transformer Manafold I called a Continuous Curvature Manifold Memory. CCMM for short. The CCMM tries to capture the underlying structure connecting states or concepts. Generally. That makes it potentially useful for Aether because the objective is for learned relationships to remain useful when the surface representation changes., this manifold learns fundamentally different then transformers, neural nets, or other existing machine Learning designs.

Each result will be challenged with fresh examples, independent runs, causal interventions, and destructive controls to separate genuine learning from memorization or shortcuts.

this builds a persistent, generalizing self evolving graphing neural architecture.

the main goal I started with developing was a way to develop intelligence without huge learning data sets, without the requirement of loading a single model into a VRAM, that could use any level of system resources to develop a reasoning platform that would store it’s learned knowledge in data storage while keeping all of the compute capacity free and available for reasoning and capability. With sophisticated recall and quantification techniqueS and data handler capabilities.

I have already made substantial progress in my design an will be able to finalize an train my first major design with this funding,

How will this funding be used?

This funding will go towards

Compute, hardware, and supporting research of the current design. I already have a robust design that i have spent over a year building and designing. I have very positive results and a solid framework. I am at the point where I would like to scale my design larger and have the hardware to scale

Who is on your team? What's your track record on similar projects?

I am an independent researcher, I currently have no one else supporting my project, I have reached out to individuals and am open to collaboration, peer review, and working with anyone who would like to support my project.

What are the most likely causes and outcomes if this project fails?

The possibility is that the system reaches a ceiling after learning relatively small temporal, structural, and causal relationships. It may simply fail to scale from these experiments into more general competence.

Composition could also be where things break.

Aether might learn several things individually but fail when I ask it to use them together.

Some apparent successes may also disappear as I build stronger tests. That has happened in smaller ways already: aggregate performance can look strong while a fresh-seed or worst-candidate test exposes something the average hides.

The self-generated experimentation idea might flop completely. It may turn out that generating and selecting its own experiments adds complexity without producing better learning.

End result I must re-design and try again

How much money have you raised in the last 12 months, and from where?

$0 in external funding.

I have no investors. Project Genesis has not received research grants, institutional funding, corporate sponsorship, or donations during the last 12 months.

I have paid for the project myself and put my own time into developing it.

I recently created a GoFundMe campaign to try to get some help with research costs, but at the time of this application it has received $0 in donations.

So the system and experimental work described here were developed without outside research funding.

That is becoming increasingly difficult as the experiments get larger. Five parallel trajectories cost more than one. Keeping checkpoints and negative results requires storage. Repeating something across seeds takes time and compute even when the first run already looks good.

That is basically why I’m applying.