The era of building bigger AI models and watching magic happen might be over. That's the message from Ilya Sutskever, the OpenAI co-founder who now runs Safe Superintelligence, and it's a pretty significant shift in how Silicon Valley thinks about artificial intelligence.

Three Eras of AI Development

In a conversation with YouTuber Dwarkesh Patel released Tuesday, Sutskever laid out a clean timeline of how AI has evolved. From 2012 to 2020, he said, we were in "the age of research." Breakthroughs came from clever algorithms and new model architectures—basically, smart people figuring out better ways to do things.

Then from 2020 to 2025, we entered "the age of scaling." The formula was simple and seductive: bigger models, more computing power, better results. Just keep cranking up the dial and watch performance improve. It worked remarkably well, which is why everyone from startups to tech giants poured billions into massive computing clusters.

Why Scaling Hits Diminishing Returns

But now? Sutskever says we've hit a wall. The scale has gotten so enormous that simply making things bigger doesn't deliver the same transformative results anymore. He put it bluntly: the belief that 100 times more computing would radically change everything "just isn't true." Sure, it would make a difference, but not the revolutionary leap people expect.

This matters because the entire AI investment thesis of the past few years has been built on scaling. If that's no longer the path forward, the industry needs a new playbook.

Research Returns, But With Better Tools

Sutskever's solution? Go back to the research-focused approach of the early 2010s, except now researchers have access to supercomputers. "It's back to the age of research again, just with big computers," he said. In other words, innovation will come from smarter ideas, not just bigger models—but those ideas can be tested with unprecedented computing resources.

From OpenAI Drama to a $32 Billion Startup

Sutskever's perspective carries weight partly because of where he's been. He was at the center of OpenAI's messy 2023 leadership crisis that briefly resulted in CEO Sam Altman's removal before his quick return. Shortly after leaving OpenAI, Sutskever co-founded Safe Superintelligence in June 2024.

The stealth-mode startup attracted serious attention fast. By September 2024, SSI had raised $1 billion from heavyweight venture firms including Andreessen Horowitz and Sequoia Capital. Reports earlier this year indicated that Meta Platforms (META) tried to acquire the company, which is now valued at $32 billion, but couldn't close the deal.

If Sutskever's right about AI's new direction, his timing for launching a research-focused company looks pretty smart.