AI, Embeddings, and Why Regulatory Retrieval Fails
Description
In the 46th installment of The Atomic Exchange Podcast, co-hosts Dr. Goran Calic and Michael Tadrous break down their new research on how large language models struggle with nuclear regulation. They open with a recap of closed-door briefings at the Canadian Nuclear Association and the Canadian Nuclear Safety Commission, then explain embeddings, how vector spaces and cosine similarity drive retrieval, and why small differences across regulations blur together. The episode walks through accuracy and distraction tradeoffs, omission risks that cascade downstream, and live tests showing that even with many retrieval attempts models still miss or mix sections. They close with practical fixes, from human-in-the-loop workflows and structured prompts to rewriting regulations into machine-readable formats that reduce error at the source.
United States
