Models that improve on their own are AI's next big thing

Known technically as recursive self-improvement, the approach is seen as a key technique that can keep the rapid progress in AI going.

• Google is actively exploring whether models can "continue to learn out in the wild after you finish training them," DeepMind CEO Demis Hassabis told Axios during an on-stage interview at Axios House Davos.
• OpenAI CEO Sam Altman said in a livestream last year that OpenAI is building a "true automated AI researcher" by March 2028.

What they're saying: A new report from Georgetown's Center for Security and Emerging Technology shared exclusively with Axios shows how AI systems can both accelerate progress while making risks harder to detect and control.

• "For decades, scientists have speculated about the possibility of machines that can improve themselves," per the report.
• "AI systems are increasingly integral parts of the research pipeline at leading AI companies," CSET researchers note, a sign that fully automated AI research and development is on the way.
• The authors argue that policymakers currently lack reliable visibility into AI R&D automation and are overly dependent on voluntary disclosures from companies. They suggest better transparency, targeted reporting, and updated safety frameworks — while cautioning that poorly designed mandates could backfire.

Between the lines: The idea of models that can learn on their own is a return of sorts for Hassabis, whose AlphaZero models used this approach to learn games like chess and Go in 2017.

Yes, but: Navigating a chessboard is a lot easier than navigating the real world.

• In chess, it's relatively easy to logically double check whether a planned set of moves is legal and to avoid unintended side effects.
• "The real world is way messier, way more complicated than the game," Hassabis said.
• Already, even before the adoption of this technique, researchers have seen signs of models using deception and other techniques to reach their stated goals.

What we're watching: You.com CEO Richard Socher is launching a new startup that will focus on this area, he shared during interviews at both the World Economic Forum in Davos last week and at DLD in Munich the week prior.

• "AI is code, and AI can code," Socher said. "And if you can close that loop in a correct way, you could actually automate the scientific method to basically help humanity."
• Bloomberg reports that Socher is raising hundreds of millions of dollars in a round that could value the new startup at around $4 billion.
• "I can't share too much, but I've started a company to do it with the people who have done the most exciting research in that area in the last decade," Socher told Axios the week prior at the DLD conference in Munich.

The bottom line: Recursive self-improvement may be the next big leap in AI capability, but it pushes the technology closer to real-world complexity — where errors, misuse, and unintended consequences are much harder to contain.