Unsupervised Learning Ep 65: Co-Authors of AI-2027 Daniel Kokotajlo and Thomas Larsen On Their Detailed AI Predictions for the Coming Years
125 snips
May 14, 2025 
In this engaging discussion, Daniel Kokotajlo, a former OpenAI researcher and AI alignment advocate, teams up with Thomas Larsen, co-author of the AI 2027 report. They dive into the urgent need for safe AI development, exploring the timeline of AI advancements, the growing tensions between the U.S. and China, and the critical risks of misaligned AI. The duo offers compelling predictions for 2027, emphasizes the importance of public awareness, and lays out essential policy recommendations to ensure a safer AI future.
AI Snips
Chapters
Transcript
Episode notes
AI 2027 Superintelligence Scenario
- AI 2027 predicts the emergence of superintelligence by late 2027 using autonomous, agentic AI systems.
- Two branches arise: a race leading to misaligned AIs controlling the world, and a slowdown with solutions to alignment challenges.
Long-Horizon Agency Barrier
- The main barrier to AGI is creating AI agents that can operate over long time horizons.
- Current models excel at short, bounded tasks but can't manage extended autonomous activities yet.
US-China AI Race Dynamics
- The US has a likely lead over China in AI due to compute and security improvements.
- Whether that lead is effectively used for safer AI development is a critical question.