Summary
The episode works through four questions that students commonly find difficult about how large language models and AI agents produce their output. It uses real next-token probabilities from GPT-2 small to show what a probability describes, how lowering the temperature changes the ratio between two tokens, and how a rare continuation becomes likely across many samples. It then examines why an instruction to cite only real sources cannot supply knowledge, how a retrieval system can report a real source inaccurately, and how to judge the risks of an agent that would read public email for a fictional health authority team. The hosts close by debating whether a health authority should run a public chatbot, drawing on a 2024 British Columbia tribunal decision about an airline's chatbot, and they agree on a practical rule for answering only from approved sources that the user can see.