You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
want people to run, adapt and own AI with the reasoning power of leading systems.
That could mean a research assistant running on your computer, learning useful approaches from working with you and carrying those discoveries into unfamiliar problems. Small teams could build capable AI around their own needs, keep sensitive work on their machines, and depend less on a handful of providers.
With Neural Core, we’re investigating whether a small neural system can learn to construct powerful reasoning methods from a limited set of building blocks. We call this a finite grammar of reasoning: operations and rules for combining them into different ways of solving problems.
A small set of rules can describe many procedures. The difficult part is learning which procedure to construct, when it applies, and how to improve it. We want the core to develop that ability through experience. A method it discovers should become something it can reuse, combine with other methods, or revise when new evidence challenges it.
The long-term ambition is for much smaller models to approach frontier-level reasoning. We would assess any advantage against the cost of the whole system, including memory, tools and any supporting model.
We’re seeking four months of support to develop and test this approach. We have working experimental software, completed training runs, and findings that guide the next experiments.
Project summary
We want people to run, adapt and own AI with the reasoning power of leading systems.
That could mean a research assistant running on your computer, learning useful approaches from working with you and carrying those discoveries into unfamiliar problems. Small teams could build capable AI around their own needs, keep sensitive work on their machines, and depend less on a handful of providers.
With Neural Core, we’re investigating whether a small neural system can learn to construct powerful reasoning methods from a limited set of building blocks. We call this a finite grammar of reasoning: operations and rules for combining them into different ways of solving problems.
A small set of rules can describe many procedures. The difficult part is learning which procedure to construct, when it applies, and how to improve it. We want the core to develop that ability through experience. A method it discovers should become something it can reuse, combine with other methods, or revise when new evidence challenges it.
The long-term ambition is for much smaller models to approach frontier-level reasoning. We would assess any advantage against the cost of the whole system, including memory, tools and any supporting model.
We’re seeking four months of support to develop and test this approach. We have working experimental software, completed training runs, and findings that guide the next experiments. We have not yet demonstrated the complete learning capability.
What are this project’s goals? How will you achieve them?
This project began with examining AI-generated solutions to mathematical problems. I became interested in the moves that make difficult problems tractable: finding a better representation, noticing a connection, introducing an abstraction, or asking a question that reveals what was missing.
I want to investigate whether a small learned core can acquire the ability to make those moves and build on them over time.
Our first target is a complete learning cycle. Given an unfamiliar problem, the core should gather useful evidence, construct a method, remember it, and use it successfully on new examples. If the problem changes, it should repair the affected method while preserving what still works.
We’ve made progress toward testing that question:
We have an experimental system that makes the core’s decisions inspectable. We can track what information it receives, what method it constructs, and whether that method changes the outcome.
We have a working comparison. A hand-written controller completed the learning-and-revision tasks in all eight held-out test worlds, showing that the available information and action limits permit a solution.
We have completed four learned-controller runs at two model sizes. None passed the full test. Increasing size alone did not resolve the failure under the training conditions we tested.
We have narrowed the next questions. Diagnostic tests found difficulties with stopping a generated method at the right point and identifying where it should apply. These give us specific explanations to test.
These experiments cover early prerequisites. They do not yet show that the core can invent a new way to investigate a problem. The next phase must establish reliable method construction and then test whether the system can go beyond reproducing a supplied strategy.
We propose the following four-month programme:
Month 1: investigate the current failures. Reproduce the results, address the identified training problems, and test whether the changes improve method construction on fresh examples. Preserve unsuccessful runs and distinguish problems with training from limitations of the proposed mechanism.
Month 2: test learning, reuse and revision together. Evaluate whether a learned method helps on new problems, whether the core knows when to use it, and whether it can repair the method after a change. Remove or replace learned methods to check whether they caused the improvement. Compare against conventional neural models and hand-written algorithms under comparable resource limits.
Month 3: test a different family of problems. If the mechanism passes the earlier tests, apply it to substantially different problems and test whether it can construct useful new combinations of operations. If it fails, use this period to test an alternative informed by the results.
Month 4: assess usefulness and publish the findings. If the mechanism works across problem families, test whether it can improve a small language model on selected reasoning tasks. Measure training and running costs, including supporting software and memory. Publish the code, reproduction instructions, results and an explanation of what the experiments establish.
Our four-month target is a convincing demonstration of learned method construction, reuse and revision. We would pursue broader transfer and small-model integration as the evidence permits.
We’re seeking $6,000–$15,000 for four months of research. Funding would allow me to work full-time on Neural Core, support my research partner Rishi V.S., and cover the experiments needed to test the approach.
At the $6,000 minimum, we would allocate:
$3,000 for my research support over four months.
$2,000 for Rishi’s research support.
$600 for model and API access used in development, research assistance and evaluations.
$300 for cloud compute.
$100 for storage and experiment backups.
This would fund a focused investigation using our existing equipment. We would concentrate on whether the core can learn, reuse and repair methods in controlled experiments, with fewer model comparisons and a narrower integration study.
I’m Raed, based in India. My academic background is in Mathematics and Computing at MSRUAS, with an 8.5 CGPA.
My engineering work includes Geowake(https://raed2180416.github.io/WakePoint/), and Phonebridge(https://github.com/Raed2180416/phonebridge). I also built Holt, a tool that helps developers inspect and preserve work across separate coding workspaces.
I contributed HP OMEN hardware support to the Linux kernel. The patch was reviewed and merged into the upstream repository. View the contribution.
Rishi V.S. is my friend and research partner and will work alongside me on Neural Core. His public projects include a photo-based attendance application and a neural-network chatbot. We will work together on implementation, experiments and analysis.
This is my first research program of this kind, and I have no prior research publications. The Neural Core work already includes implemented experiments, completed training runs, comparisons against a working hand-written solution, and records of unsuccessful results. Those experiments have identified specific weaknesses we can now investigate.
We use AI coding and research assistants extensively. We are responsible for the research decisions and claims, and we distinguish what the learned core does from capabilities supplied by existing models, tools and our experimental software.
The core may fail to learn how to construct useful methods. A small set of building blocks can express many possible solutions, but finding the right combination may require more training or computation than we expect.
It may work only in our controlled experiments. A method that succeeds on simple artificial problems might fail when the same idea appears in a different form. We will test unfamiliar problems and different environments before making broader claims.
An apparent improvement could also come from help supplied by the surrounding software. We will remove or replace the learned methods to check whether they caused the improvement.
The system might become capable but too expensive. Search, memory and assistance from larger models could outweigh any savings from a small core. We will measure those costs.
We may also find that existing approaches solve the problem more effectively. In that case, we would explain what the comparison shows and whether there is a reason to continue with our mechanism.
If the approach fails, the project would still produce reproducible experiments and a report showing which ideas were tested, what happened, and what remains unresolved. We would preserve unsuccessful results and document changes in direction. Matching frontier models remains the long-term ambition; the four-month programme tests the mechanisms that would have to support it.
We have raised $0 for Neural Core. We have received no grants or external investment for this project.