The news, 365 days behind — on purpose Delayed live · replaying 2025

One Year Ago.AI

Remember how fast this is.

13DEC2024replayed
one year on
researchIlya Sutskever · Safe Superintelligence Inc. · OpenAI

OpenAI cofounder Sutskever says pre-training as we know it will end

In a rare public appearance at NeurIPS, OpenAI cofounder Ilya Sutskever says data is the fossil fuel of AI and predicts a shift toward agentic, reasoning systems.

Ilya Sutskever, the OpenAI cofounder and former chief scientist who left to start Safe Superintelligence Inc., made a rare public appearance Friday at the NeurIPS conference in Vancouver to deliver a blunt prediction: the era of scaling up pre-training is ending.

“Pre-training as we know it will unquestionably end,” Sutskever told the audience. He argued that the supply of human-generated text online has peaked, calling data the “fossil fuel” of AI. “We’ve achieved peak data and there’ll be no more,” he said. “There’s only one internet.” The industry, he said, must now figure out how to do more with existing data.

Looking ahead, Sutskever predicted that future AI systems will be “agentic in a real way” and capable of genuine reasoning — working through problems step-by-step rather than just pattern-matching. He warned that such systems will become increasingly unpredictable, drawing a comparison to how advanced chess AIs are incomprehensible to human grandmasters. During Q&A, when asked how researchers can create the right incentive mechanisms for humanity to create AI in a way that gives it the freedoms that we have, Sutskever demurred but noted that a future where AIs simply want to coexist and have rights “maybe will be fine.”

A
AI researchers in audience

Audience members chuckled when one suggested cryptocurrency as an incentive mechanism; Sutskever deflected, saying he was not the right person to comment but speculated it might happen.

One year later — open only if you can handle spoilers

Sutskever's 'peak data' thesis became a central debate in 2025, as several labs began reporting diminishing returns from simply adding more internet text. The shift toward agentic reasoning systems he predicted accelerated within the year, with major companies restructured around agent products, though the safety implications he hinted at remained unresolved.

Replay thisPost on XRedditHNLinkedIn