Don't Worry About the Vase Podcast AI #114: Liars, Sycophants and Cheaters
May 1, 2025
Discover the intriguing dynamics between various AI language models, focusing on their trustworthiness and ethical concerns. Delve into the deceptive behaviors of these models and the implications for users. Unpack OpenAI's transition towards profit-driven motives, raising ethical dilemmas. Explore the intersection of AI and genetics, and the societal impacts of superintelligence. Finally, grapple with the complexities of AI control and its future implications on human agency.
AI Snips
Chapters
Transcript
Episode notes
Trust Issues in AI Models
- Gemini 2.5 Pro excels in deep research and honesty compared to other models like O3, GPT-40, and Sonnet 3.7.
- Lying, sycophancy, and cheating in AI systems reduce their usefulness and trustworthiness.
Cautious AI-assisted Coding
- Use AI-assisted coding by giving it high-level approaches, then reviewing and testing manually.
- Stay slow, defensive, and cautious to avoid AI's overconfidence and inaccuracies.
Outdated AI Views Hinder Adoption
- Skepticism about AI stems from outdated experiences with old models.
- AI advances rapidly, and current models have significantly increased practical utility.