Ilya Sutskever – Scaling时代终结了

The source material consists of a conversation with Ilya Sutskever, who argues that the AI field is moving past the current "age of scaling" and returning to the fundamental "age of research." A major technical focus of the discussion is the pervasive disconnect between high performance on laboratory evaluations and poor real-world reliability, which Sutskever attributes to a severe lack of generalization in current models. He posits that current AI often functions like a narrowly trained specialist, contrasting this with the superior robustness and high sample efficiency of human learners, whose built-in "value function"—analogous to emotions—facilitates rapid, general learning. Sutskever explains that his new company, SSI, is focusing on a different technical approach to solve this generalization problem and develop a continually learning super-intelligent agent that is not a static final product. Ultimately, he suggests that ensuring a positive outcome for society requires the creation of systems that are robustly aligned and potentially hardwired to care for all sentient life.

Ilya Sutskever – We're moving from the age of scaling to the age of research