Interrupting AI by Natural Language
Combining Stop Generating function with your prompt
The common LLM interaction
The cat user provides a prompt, and the LLM generates a response. This turn-taking is rigid. Unless AI stops itself or the system intervenes, it keeps generating until it hits the token limit.
Two key Questions
1. Why rely on system commands like "stop generating"? Are there alternative methods — perhaps using the most familiar interface to us, language itself?
2. Must we refine prompts manually? In human conversations, we adjust by adding words, not rephrasing entire statements. What if AI worked similarly?
A possible solution: Interrupting AI with natural language.
Our Approach
We fine-tuned LLaMA 3-8B on a custom dataset and built a web-based system.
When AI is generating, a user can interrupt naturally, e.g., “Any vegan options?”. This phrase acts as a prompt, eliminating manual stops or prompt revisions.
32 participants tested the system. Most found it efficient. Details in the research paper.
Tech Stack
Python, Llama, HTML, CSS, Javascript, Photoshop, Illustrator used.