FishNet evolves language via human-guided CPPNs

FishNet represents a rejection of the transformer paradigm, demonstrating that complex linguistic abstractions can emerge through evolutionary search rather than gradient descent. Its CPPN architecture treats text as a coordinate-based pattern, allowing for unique structural generation and the ability to zoom into linguistic patterns. By incorporating human-in-the-loop selection, the project identifies meaningful linguistic fragments that raw statistics might miss, democratizing model training through shared checkpoints and automated evolution options using local LLMs. This highlights a path toward discovering language capabilities in compact architectures without massive datasets.