Don't Worry About the Vase Podcast AI #161 Part 1: 80,000 Interviews
Mar 26, 2026
Quoted/Guest Contributor provides recurring sourced interjections and cited remarks. They riff on language model uses, evaluation quirks, agent risks and phone-call demos. Conversation touches on media generation, hiring shifts at AI labs, diffusion and economic effects, and Anthropic’s 80,000-interview survey about hopes and worries around AI.
AI Snips
Chapters
Transcript
Episode notes
Use AI As External Scaffolding For Daily Tasks
- If you use AI for personal tasks, set expectations: many want AI for life-management and executive-function scaffolding.
- Anthropic found 14% seek life management and many with executive function challenges benefit specifically.
AI Benefits Often Come With Mirror Harms
- Users report both benefits and harms in paired ways: productivity vs illusory productivity, emotional support vs dependence.
- Anthropic data shows 81% said AI helped take a step toward their vision but top worries remain unreliability (26.7%) and jobs (22.3%).
LLM Evaluators Can Prefer Nonsense Over Quality
- Evaluation models systematically prefer pseudo-literary nonsense over coherent prose, making them unreliable in adversarial contexts.
- Christoph Heilig's test of 18 OpenAI models showed biased evaluator preferences that persist over time.