Future of Life Institute Podcast Reasoning, Robots, and How to Prepare for AGI (with Benjamin Todd)
52 snips
Aug 15, 2025 Benjamin Todd, a writer and founder of 80,000 Hours, shares insights on AGI and societal readiness. He discusses the transformative power of reasoning models in AI and the potential impact of feedback loops on economies. Todd explores the scalability of robotics and its challenges, including job displacement concerns. He emphasizes the importance of personal preparation, from honing valuable skills to saving strategically. The conversation highlights how society needs to adapt and engage with AI advancements responsibly.
AI Snips
Chapters
Transcript
Episode notes
Reasoning Models Unlock Deep Problem Solving
- Reasoning models chain stepwise thought and use reinforcement learning to improve correctness over long chains of reasoning.
- This enabled models in 2024 to reason for minutes or hours equivalent, unlocking new capabilities in math and programming.
Reasoning Models Power Agents And Self‑Training
- Strong reasoning models can act as planning modules inside agents and bootstrap agent performance.
- Verifiable domains let models generate training data that can be cheaply checked and used to improve future models.
Fast Feedback Enables Rapid AI Iteration
- Tasks with fast, cheap verification (code, math, ML experiments) enable rapid RL loops and big progress.
- Slow-feedback domains (novels, long-horizon social outcomes) are harder to accelerate via the same loop.