Today in Tech Why AI still hallucinates — and how to stop it | Ep. 242
Aug 19, 2025
In a fascinating discussion, Byron Cook, a Distinguished Scientist and VP at AWS, dives into the world of AI hallucinations. He explains how these inaccuracies can be both misleading and surprisingly useful. The conversation covers the essential role of automated reasoning as a 'logic cop' to enhance AI reliability and the challenges in defining truth within business contexts. Cook emphasizes the need for safeguards around agentic AI and shares insights on when AI missteps can have serious consequences for users.
AI Snips
Chapters
Transcript
Episode notes
Hallucination Is Also Creativity
- Hallucination in LLMs is also the creativity users want for tasks like writing and art.
- We must combine creative models with tools that ensure contextual appropriateness.
Next-Token Prediction Lacks Thought
- Transformer models predict the next token without reasoning, so outputs lack intent or thought.
- That token-prediction design explains why models confidently produce wrong facts.
Truth Is Hard To Define In Practice
- Defining