N N Taleb's Probability Questions (UNOFFICIAL) Stephen Wolfram visits RWRI 18 (Summer Workshop) (2023)
14 snips
Nov 11, 2025 
Stephen Wolfram, founder of Wolfram Research and creator of Mathematica, joins the discussion to delve into the intricacies of AI and computation. He describes ChatGPT as a 'stochastic parrot' and explores the flaws of statistical language models. Wolfram discusses the implications of automation on jobs, the need for constraints on AI outputs, and the potential for LLMs to serve as user interfaces. He also highlights the challenges in defining AI governance and the importance of integrating real-world knowledge to enhance decision-making.
AI Snips
Chapters
Transcript
Episode notes
Choose Tasks Where LLM Failures Are Tolerable
- Don't use LLMs when you need 100% precision; use them where 80–90% correctness is acceptable.
- Design applications so failures are tolerable and success yields meaningful gains.
Risk Of LLMs Learning From Their Own Output
- Growing web content produced by LLMs risks a feedback loop where models train on their own bland outputs.
- That may homogenize style and reduce informational richness over time.
Limit AI Actuation Not Internal Thought
- Constrain AI at the actuation layer rather than micromanaging internal thoughts.
- Put robust safeguards on what actions an AI can perform on systems and accounts.