XNews

Ilya Sutskever & the shift from scaling to research

Ilya Sutskever & the shift from scaling to research

Ilya Sutskever outlines a shift from scaling to new research focused on human‑like learning.

Key insights

Ilya: scaling yields continued improvement but misses key abilities

Ilya Sutskever clarified that continued scaling (more compute/data) will keep improving models, but certain essential capabilities will remain absent without new paradigms. He framed current models as high‑scoring 'exam machines' lacking efficient, human‑like sample efficiency and robust generalization, and argued for new training principles to enable rapid learning and transfer.

Ilya defines future AI as learning‑first systems, not prepackaged experts

Ilya recommended reframing AGI as a learning agent that acquires skills through interaction (a rapid '100‑hour student') rather than an out‑of‑the‑box omniscient system. He emphasized value functions and continuous learning as key to achieving human‑level generalization and proposed research into learning algorithms inspired by human cognition.

Safety and deployment: Ilya favors staged visibility and gradual integration

Ilya argued for progressive deployment and transparency: letting the world see AI capabilities gradually can improve societal understanding and safety practices. He expressed that safety is intertwined with capability and suggested external review and staged release, while SSI continues research toward safe superintelligence.

Ilya links human learning efficiency to internal value systems

Ilya highlighted human internal value systems (emotion, reward) as mechanisms that enable early feedback and efficient learning, contrasting them with current RL setups that often optimize for narrow eval metrics and can encourage reward‑hacking.

Andrej Karpathy: engineering focus on embodied agents and integration

Andrej Karpathy emphasized the engineering path to agent deployment—integrating models with memory, tool use, and continuous learning to produce robust agents. He described agent construction as a long systems engineering effort ('ten‑year long run') and expects gradual economic impact rather than abrupt disruption.

Ilya maintains optimistic but cautious timeline for human‑like learning

Ilya provided a broad time horizon of 5–20 years for systems achieving human‑like learning and generalized capabilities, while noting uncertainty and the need for new research paradigms rather than pure scale.

Source posts