Beren Millidge, John Schulman, and Charlie O'Neill debate how close we are to recursive self-improvement

Beren Millidge、John Schulman、Charlie O'Neill 对谈递归自我改进离我们还有多远

John Schulman, Beren Millidge, and Charlie O'Neill discuss why 2036 might not bring superintelligence. Schulman points to a repeating cycle: each new model feels like AGI at launch, then feels dumb after a month, because models still have weak judgment and self-checking. Millidge flags the sim-to-real gap—models ace benchmarks but stumble in the real world—and says unsolved meta-learning and continual learning could keep it that way. O'Neill frames it as a question of whether the Transformer-plus-RL recipe needs another Moore's-law-style discontinuity to keep climbing, or whether we're simply far from the optimal learner a chip can run. No one gives a firm timeline, but all agree we're nowhere near the ceiling.

Why it matters: A podcast conversation among three frontline researchers debating the real distance to recursive self-improvement, with concrete observations and clashing views. Hits all three HKR axes, but as a discussion piece rather than a product launch or paper, the information density i...