Risky Business with Nate Silver and Maria Konnikova What’s the Optimal Personality for AI?
19 snips
Aug 14, 2025 The hosts dive into the launch of GPT-5, highlighting its improved speed and user-selectable personalities. They discuss the challenges of AI reliability, emphasizing the need for human oversight in its outputs. Data privacy risks are examined, stressing the importance of safeguarding personal information. The conversation shifts to the geopolitical implications of U.S.-China tech trade, sparked by a new Nvidia deal. The episode wraps up with entertaining insights on the nature of AI accountability in a world of evolving technology.
AI Snips
Chapters
Books
Transcript
Episode notes
GPT-5 Is Improved But Still Unreliable
- GPT-5 shows mixed improvements: sometimes more detailed, sometimes flaky in actions like exporting to PDF.
- Users still need to monitor and correct outputs because the model can fail in practical tasks.
Itinerary Then Code Panic
- Nate asked GPT-5 to create a travel itinerary and then email it as a PDF.
- The model produced plausible itinerary text but failed when asked to generate a PDF and showed confusing code instead.
Hallucination Claims Overstated
- Maria tested GPT-5 on academic sourcing and found it still hallucinates nonexistent papers.
- Claims of 'no more hallucination' are overstated and require user expertise to detect errors.
