Overview
-
The Generalization Gap: Current AI models perform impressively on benchmarks but generalize dramatically worse than humans, making basic errors despite superhuman eval scores. This disconnect between eval performance and real-world reliability remains poorly understood.
-
Return to Research Era: After years of scaling pre-training, compute is now so large that simply adding more may not transform capabilities. We're entering a new research era requiring novel ideas, not just bigger clusters.
-
Alignment Through Sentience: Sutskever suggests building AI that cares about sentient life may be easier than human-only alignment, since AI itself will likely be sentient and empathy emerges from self-modeling circuits.
Takeaways
Ilya Sutskever shares his evolving views on AI progress and safety in this wide-ranging conversation. His key insight: the fundamental bottleneck is that models generalize far worse than humans, and solving this—not just scaling—will determine the path to superintelligence.
"The whole problem of AI and AGI? The whole problem is the power. When the power is really big, what's going to happen?"