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

Summary

• Sutskever’s headline call: the age of scaling (roughly 2020–2025) is over. Pre-training data “is very clearly finite,” and he rejects the premise that raw compute still transforms outcomes — “Is the belief that if you just 100x the scale, everything would be transformed? I don’t think that’s true. So it’s back to the age of research again, just with big computers.” For capital allocators this reframes AI from a low-risk scaling recipe back into high-variance research bets.
• The eval–economy disconnect is real, and Ilya offers a possible structural explanation: models perform very well on hard evals yet lag in economic impact, potentially because labs inadvertently build RL environments “inspired by the evals” — Dwarkesh’s quip that “the real reward hacking is the human researchers” — and because models “generalize dramatically worse than people. It’s super obvious.” Ilya’s verdict on the current paradigm: it “will go some distance and then peter out.”
• Timeline: 5 to 20 years to a system that learns like a human and consequently becomes superhuman. Even in the stall-out scenario, incumbent labs “could make a stupendous revenue. Maybe not profits” — differentiation pressure eats margins, a direct caution on frontier-lab economics.
• SSI’s compute-poverty rebuttal: its $3B is more comparable to rivals than headline raises suggest, because their “big loans” are “earmarked for inference” and product staff — “when you look at what’s actually left for research, the difference becomes a lot smaller.” Context on the cofounder exit: SSI was fundraising at a $32B valuation, Meta offered to acquire, “I said no. But my former cofounder in some sense said yes.”
• Superintelligence is a learner, not a finished mind — “a superintelligent 15-year-old that’s very eager to go,” deployed like a new hire. He expects rapid economic growth from broad deployment but pushes back on winner-take-all: competition drives specialization into niches. “In theory, there is no difference between theory and practice. In practice, there is.”
• A genuine change of mind: Ilya now weights incremental deployment more, because “it’s very hard to feel the AGI” — you must be “showing the thing,” and this “may back-propagate into the plans of our company.” Predictions: labs “will become much more paranoid” about safety once AI feels powerful, and rival labs will converge on alignment strategies.
• His preferred alignment spec: an AI “robustly aligned to care about sentient life” — possibly easier than caring for humans alone since the AI itself will be sentient — plus a cap on the power of the most powerful superintelligence. His reluctant long-run equilibrium answer: humans become “part-AI with some kind of Neuralink++. I don’t like this solution, but it is a solution.”

Deep dive

Not yet available upstream; scheduled sync will retry.