Ilya Sutskever describes the recent period of AI development as an age of scaling that is giving way to a renewed age of research. Pretraining produced major gains, but simply multiplying compute no longer appears sufficient to make models learn or generalize like people.
Ilya Sutskever identifies sample efficiency as the fundamental difference. People can infer broad rules from a small number of examples, while current models often need vast datasets and still behave inconsistently outside familiar patterns. He says Safe Superintelligence is pursuing a different technical approach built around mechanisms that may narrow this gap.
The video connects that research direction to reports of a major NVIDIA investment and access to newer Vera Rubin computing systems. Those reports suggest the approach may now receive much more compute, but the underlying method remains undisclosed and the video is careful to infer rather than demonstrate what Safe Superintelligence has built.
The practical takeaway is that the next capability jump may depend less on another conventional pretraining expansion and more on finding a learning process that transfers knowledge reliably. Compute still matters, but only after research identifies an approach worth scaling.
Watch on YouTube
