Descriptions:
Nathan Labenz of the Cognitive Revolution podcast appears as a guest on the Intelligence Horizon podcast, hosted by Yale seniors Owen Jang and Will Sanuk Dufalo. The conversation covers the compression of AI timelines over the past five years, the evidence from interpretability research showing increasingly sophisticated world models inside AI systems, and why Labenz believes reinforcement learning scaling has moved AI beyond simple human imitation.
Labenz places his p(doom) estimate in the 10–90% range while expressing cautious optimism, arguing that a defense-in-depth strategy combining Goodfire’s intentional design work, Redwood Research’s AI control techniques, formal software verification, and pandemic preparedness could collectively keep society stable. The episode also addresses the US-China AI rivalry, concerns about the Department of War’s recent actions against Anthropic, and why Labenz would rather bet on human cooperation than on researchers’ ability to fully steer AI outcomes unilaterally.
On the technical side, the discussion engages directly with Yann LeCun’s critique that next-token prediction is the wrong objective for genuine AI understanding, with Labenz acknowledging the intuition while defending current scaling directions. The episode also references a Google DeepMind paper by researcher Rohan Shaw on upper bounds for reasoning in transformers versus state space models without externalized chain-of-thought.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published April 01, 2026
🏷️ Format: Podcast
