When should “learn as we go” stop being the default?

Zach DavidsonOct 3, 2026

Miles Brundage says he regrets advocating for iterative deployment in the GPT-3 era, arguing that the risk picture has changed as AI capabilities and real-world harms have grown. That raises a hard question: how should deployment decisions change as the stakes rise? What evidence should make a lab slow down, and who should get to decide?