80,000 Hours Podcast What the hell happened with AGI timelines in 2025?
289 snips
Feb 10, 2026 A deep dive into why predictions about transformative AGI swung wildly in 2025. The conversation covers how new reasoning models briefly tightened timelines and why that optimism faded. It explores technical limits like inference-time gains, RL inefficiencies, and scaling costs. Non-technical factors and measured forecast updates around a 2032 pivot point are also discussed.
AI Snips
Chapters
Transcript
Episode notes
Demo Impressiveness Outpaces Real-World Usefulness
- Impressive demos haven't translated into equivalent workplace productivity gains.
- Models improve visually faster than they improve usefulness in messy, real-world contexts, widening the demo-to-impact gap.
Auto-Coding Won't Instantly Automate AI R&D
- Automating software engineering alone won't guarantee fully automated AI R&D because many research tasks aren't pure coding.
- As capabilities improve, bottlenecks shift and more tooling may be needed just to sustain prior progress rates.
Forecasts Shifted A Few Years Longer
- Metaculus' median strong-AGI forecast moved from July 2031 to November 2033 in one year.
- That ~2.5-year shift reflects industry sentiment moving moderately longer, not a collapse of progress.