Doom Debates DEBATE: Is AGI Really Decades Away? | Ex-MIRI Researcher Tsvi Benson-Tilsen vs. Liron Shapira
27 snips
Nov 7, 2025 In a thought-provoking debate, Tsvi Benson-Tilsen, an ex-MIRI researcher and founder of the Berkeley Genomics Project, argues that AGI is much further away than commonly believed. He emphasizes the limitations of current AI, pointing out tasks it struggles with, like generating novel scientific ideas. The conversation also explores the need for clear benchmarks in predicting AI progress and debates whether advances could trigger an AI winter. Tsvi proposes germline engineering as a solution for enhancing human intelligence to tackle future challenges.
AI Snips
Chapters
Transcript
Episode notes
Early Timeline Probabilities Stated
- At the start Tsvi says he assigns a 1–3% chance of AGI in the next five years.
- Liron reacts that this is much lower than mainstream consensus and pushes for goalposts.
Benchmarks Can Miss Core Intelligence
- Tsvi argues LLM benchmarks miss the deep capabilities that indicate true general intelligence.
- He notes LLMs fail at novel, edge-case reasoning and produce nonsense on significant novelty.
Creativity And Sample Efficiency As Barriers
- Tsvi predicts AI won't lead most research by producing novel concepts that interest human experts.
- He cites Cantor's theorem and sample complexity as concrete examples of current limitations.