Author here. This started as a simple frustration: LLMs re-derive everything from scratch, every time. A model that has correctly solved something a million times will spend the same tokens on the million-and-first.
KNOW proposes extracting proven reasoning patterns and compiling them to lightweight, deterministic programs that any model can invoke. The network builds itself - pattern detection becomes a pattern, extraction becomes a pattern. Intelligence accumulates in the commons, not in proprietary weights.
I wrote up the concept because I wanted to see if the idea survives contact with smarter people. There are open questions I don't have answers to (extraction fidelity, adversarial robustness, routing at scale). Happy to hear where this falls apart.
Author here. This started as a simple frustration: LLMs re-derive everything from scratch, every time. A model that has correctly solved something a million times will spend the same tokens on the million-and-first.
KNOW proposes extracting proven reasoning patterns and compiling them to lightweight, deterministic programs that any model can invoke. The network builds itself - pattern detection becomes a pattern, extraction becomes a pattern. Intelligence accumulates in the commons, not in proprietary weights.
I wrote up the concept because I wanted to see if the idea survives contact with smarter people. There are open questions I don't have answers to (extraction fidelity, adversarial robustness, routing at scale). Happy to hear where this falls apart.
https://github.com/JoostdeJonge/Know