Neural Graffiti introduces a novel "Spray Layer" for real-time, neural-level modification of pre-trained Large Language Models (LLMs). This experimental technique allows users to inject memory and influence the model's internal state and output without retraining, targeting researchers and power users interested in exploring LLM behavior and neuroplasticity.

How It Works

Inspired by liquid neural networks and graffiti art, the Spray Layer integrates directly into the final stages of transformer inference. It injects a "memory trace" into vector embeddings, subtly altering the model's conceptual relationships and generative predictions over time based on past interactions. This approach influences the model's "thinking" process rather than forcing specific word outputs, allowing for gradual behavioral drift and the development of emergent "personalities."

Quick Start & Requirements

• Install: Requires Python and Hugging Face transformers.
• Prerequisites: A Hugging Face token is needed to download models.
• Demo: Available via Colab notebook.

Highlighted Details

• Enables real-time, in-inference modification of LLM behavior.
• Injects memory and influences conceptual relationships without retraining.
• Allows tracking of the Spray Layer's influence on each output.
• Aims to imbue LLMs with emergent personality and memory.

Maintenance & Community

The project is maintained by babycommando. Further community engagement details (e.g., Discord, Slack) are not specified in the README.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility is noted for any model exposing hidden states within the Hugging Face transformers ecosystem.

Limitations & Caveats

This is an experimental technique with results still being released. It does not guarantee specific word outputs and may not be suitable for business deployments due to the potential for creating highly specific, opinionated AI "entities."