Help us evolve a new type of AI, trained directly by humans. To start, just click on the text you find most interesting.
Initializing...
Help progress our mission by uploading your current checkpoint to the FishNet checkpoint repository on GitHub
Interactive text evolution using neural networks
FishNet is an interactive application that uses Compositional Pattern-Producing Networks (CPPNs) to generate and evolve text sequences. The system employs a population-based evolutionary algorithm where neural networks compete and reproduce based on various evaluation criteria.
This project is an initial development aimed at evolving computer-generated text sequences, drawing inspiration from a platform called PicBreeder.* The project is ongoing, and any feedback or suggestions are highly appreciated.
My hypothesis is that, given PicBreeder's ability to learn abstract concepts such as symmetry, a similar neural network architecture might also be capable of learning more complex abstractions like language. Both PicBreeder and FishNet utilize a type of neural network known as a CPPN (Compositional Pattern Producing Network), which differs in architecture from traditional language models (LLMs). The full paper describing PicBreeder can be found here: https://wiki.santafe.edu/images/1/1e/Secretan_ecj11.pdf.
If you are interested in contributing, please visit the Evolve page and select the sequence that you find most intriguing. If you discover something noteworthy—such as a word, a recognizable pattern, or any other structure—please click the "I think I found something cool" button.
For context, the name FishNet originated during early testing when I trained it to output the phrase "fish fish fish fish fish fish it."
The source code is available here: https://github.com/Z-Coder672/FishNet/tree/main/code
*Please be cautious regarding this unofficial PicBreeder site. Its reliability is uncertain, and I recommend avoiding the save button. However, the link is: https://nbenko1.github.io/. The official PicBreeder website at picbreeder.org is currently unavailable.
The evolution follows a PicBreeder-inspired algorithm:
Uses the Gemma3 1B language model to evaluate text "interestingness". This method sends the generated sequences to a local Ollama instance and asks the AI to choose the most interesting one.
Statistical evaluation based on English language characteristics with automatic revert logic:
The Gemma3 1B evaluation method requires a local Ollama instance running on your machine. Without Ollama, the system will fall back to English scoring.
The mutation rate is set to the value of the learning rate slider.
Each individual in the population is a CPPN (Compositional Pattern-Producing Network) with:
ollama pull gemma3:1bollama serve