Hello everyone,
We are all familiar with the two great frustrations of building LLM agents:
- They constantly break character to remind you they are “just an AI.”
- They burn expensive API tokens to answer static, factual questions they should already know.
Today, I’m open-sourcing a proof-of-concept that solves both issues using a biologically inspired architecture. I call it FlyBrain_MrSmith (CCEv5).
Instead of relying on brittle prompt engineering, this system routes user input through a PyTorch-based Spiking Neural Network (modeled on a fly’s Mushroom Body) *before* it ever reaches the LLM.
This achieves two remarkable things:
* The SNN dictates the Persona: It mathematically enforces the agent’s identity, completely overriding standard RLHF refusals. (Try asking it if it’s an AI—it will refuse, acting entirely as “Mr. Smith”).
* The Zero-Host Bypass: If the SNN detects high semantic resonance with an absolute biological memory, it bypasses the LLM entirely, answering directly and dropping your API token spend to exactly zero.
I’ve built a Gradio Glass-Box UI directly into the Colab so you can watch the neural raster plot fire in real-time, track the shifting system prompt, and see your token savings accrue as the agent learns via dopaminergic Q-learning.
Try the Sandbox here: Google Colab
(Note: You will need a free Groq API key for the fallback LLM generation, but you’ll be surprised how rarely the engine actually needs to use it).
I would love to hear your thoughts on combining Spiking Neural Networks with standard LLM pipelines. How can we push “Zero-Host” architectures further?