Get Dario Amodei on Dwarkesh Podcast

As we move toward recursive self-improvement, how do you think about the problem of value drift and corrigibility? Do you believe we can build models that are robustly aligned even as they become vastly more capable, or is there a fundamental boundary to what safety can achieve?

Dario, you've argued that scaling laws have held up far longer than skeptics predicted. Where do you see the next thresholds? Is there a ceiling to current approaches, and how do you think about compute, data, and algorithmic efficiency as complementary resources?

Scaling models requires enormous clusters and power. What do you see as the most likely constraints over the next decade—chip supply, energy grids, or something else? And what does that mean for the concentration of AI power in a few organizations or nations?