LLMs contain all knowledge – I built way to mine deep meaning from them

https://github.com/andres-ulloa-de-la-torre/NoA

by scraper01 • 11 months ago

3 2 11 months ago

scraper01• 11 months ago

Hey everyone.

I've been looking into a fundamental problem in modern AI. We have these massive language models trained on a huge chunk of the internet—they "know" almost everything, but without novel techniques like DeepThink they can't truly think about a hard problem. If you ask a complex question, you get a flat, one-dimensional answer. The knowledge is in there, or may i say, potential knowledge, but it's latent. There's no step-by-step, multidimensional refinement process to allow a sophisticated solution to be conceptualized and emerge.

The big labs are tackling this with "deep think" approaches, essentially giving their giant models more time and resources to chew on a problem internally. That's good, but it feels like it's destined to stay locked behind a corporate API. I wanted to explore if we could achieve a similar effect on a smaller scale, on our own machines. So, I built a project called Network of Agents (NoA) to try and create the process that these models are missing.

The core idea is to stop treating the LLM as an answer machine and start using it as a cog in a larger reasoning engine. NoA simulates a society of AI agents that collaborate to mine a solution from the LLM's own latent knowledge.

It works through a cycle of thinking and refinement, inspired by how a team of humans might work:

The Forward Pass (Conceptualization): Instead of one agent, NoA builds a whole network of them in layers. The first layer tackles the problem from diverse angles. The next layer takes their outputs, synthesizes them, and builds a more specialized perspective. This creates a deep, multidimensional view of the problem space, all derived from the same base model.

The Reflection Pass (Refinement): This is the key to mining. The network's final, synthesized answer is analyzed by a critique agent. This critique acts as an error signal that travels backward through the agent network. Each agent sees the feedback, figures out its role in the final output's shortcomings, and rewrites its own instructions to be better in the next round. It’s a slow, iterative process of the network learning to think better as a collective. Through multiple cycles (epochs), the network refines its approach, digging deeper and connecting ideas that a single-shot prompt could never surface. It's not learning new facts; it's learning how to reason with the facts it already has. The solution is mined, not just retrieved. The project is still a research prototype, but it’s a tangible attempt at democratizing deep thinking. I genuinely believe the next breakthrough isn't just bigger models, but better processes for using them. I’d love to hear what you all think about this approach.

Thanks for reading.

physix• 11 months ago

Do you have any examples?

LLMs contain all knowledge – I built way to mine deep meaning from them

Related Stories