You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I'm working on a question that I've been thinking about for a while: what if AI systems could become better not just by using more compute, but by improving the way they actually reason?
That's what I'm trying to explore with Manaslu Labs. The idea is to build systems that can find weaknesses in their own architectures, propose changes, test them, and keep the ones that work. Eventually, I want to find out whether this process can become self-sustaining, where each generation becomes better at improving the next.
I've built an experimental system around Lean theorem proving, where I can test architectural modifications and measure whether they improve reasoning performance. Some experiments have shown improvements, others have failed to generalize. It's still early, and I haven't demonstrated accelerating self-improvement yet.
The biggest constraint right now is resources. I'm doing this with a personal laptop and limited access to compute. Funding would allow me to run larger experiments, test more architectures, and properly investigate whether this approach can scale.
I don't know if the original hypothesis will turn out to be correct. That's what I want to find out.
My goal is to build AI systems that can improve their own reasoning architectures, rather than relying entirely on more compute or human-designed improvements.
I'm starting with Lean 4 theorem proving, where architectural changes can be tested and verified objectively. We've already seen improvements in reasoning performance and generalization, although true autonomous self-improvement remains unproven.
Next, I want to close the loop: let the system identify its own weaknesses, propose architectural changes, implement them, and independently evaluate the results.
Ultimately, I want to test whether these improvements can compound across generations, and whether meaningful progress toward recursive self-improvement is possible with limited compute.
Most of the funding would go toward compute, running larger experiments, and developing Manaslu's self-improvement system.
Right now, I'm working with a personal laptop and limited free GPU access, which restricts how many experiments I can run. Additional resources would let me test more architectural modifications, train small specialized reasoning models, and evaluate improvements across successive generations.
I'd also use part of the funding for research infrastructure, storage, and independent benchmarking to make sure the improvements we observe actually generalize rather than just overfit our experiments.
I'm currently the founder and sole researcher at Manaslu Labs. I handle the research direction, architecture, implementation, and experiments, with AI coding agents assisting in development.
Previously, I built a Devanagari OCR system that achieved 2.19% character error rate on my clean evaluation dataset, outperforming several established OCR engines in my tests. I also developed ARTS, an experimental real-time settlement protocol, and worked on payment infrastructure through PayArk.
For Manaslu, I've built a Lean 4-based reasoning system and conducted multiple generations of architectural experiments, including improvements that transferred to held-out theorem families.
I'm 15 and largely self-taught. I don't have an academic research background or a large team, but I've been building and testing systems independently for years.
The biggest risk is that architectural improvements don't generalize or compound across generations. We've already encountered this in earlier experiments, where improvements on development benchmarks failed to transfer to unseen problems.
Another possibility is that the cost of discovering better architectures grows faster than the improvements themselves, making recursive self-improvement impractical with limited compute.
If the project fails, we'll still have a better understanding of the limits of architectural self-improvement, along with experimental results, evaluation methods, and potentially useful reasoning techniques.
Even a negative result would help answer an important question: how far can intelligence improve through architectural innovation without simply scaling compute?
Manaslu Labs hasn't raised any external funding in the last 12 months. The research has been entirely bootstrapped, using my own resources, free compute, and AI development tools. I've recently applied for research funding, but haven't received any investment or grants yet.
There are no bids on this project.