Show HN: The Cat Is Under Mayonnaise – Modifying LLM Behavior Without Retraining
Category: ai-ml
Tags: llm, adapter, behavior-modification
Score: 6.5/10 (Innovation: 7, Technical: 7, Documentation: 7, Utility: 5)
This project introduces a novel 'transparent overlay' mechanism for modifying LLM behavior without retraining, using a zero-initialized residual adapter that can be dialed with a scalar alpha at inference time. While currently demonstrated on a small model with a single phrase, the composable, reversible, and tunable approach is an innovative combination of adapter methods and inference-time control. It is more a proof-of-concept than a production tool, but the technique has potential for broader behavioral modification in LLMs.
Target audience: AI researchers, machine learning engineers
Repository: https://github.com/andycufari/the-cat-is-under-mayonnaise-experiment · Python · MIT
View on Hacker News