Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
April 24, 2025
·
Dublin
Single Embedding LLM Control
Live demonstration of altering a single embedding vector to direct a language model's output without retraining, showing how to produce arbitrary text.
Overview
What if I told you a single embedding could make a language model say anything you want? In this lightning-fast demo, I’ll show you how to hijack an LLM’s output with one sneaky vector—no retraining, no fuss. We’ll fire up the code, tweak an embedding live, and watch the model spit out whatever wild text we choose. An advanced trick with big implications: total control over LLMs!
Links
Project optimizes embeddings for specific LLM token generation using PyTorch.
acmc's Hugging Face Space demonstrates Simon Says prompt generation.
Tech stack