IAPS fellow Severin Field interviewed 25 researchers from OpenAI, Anthropic, Google Deepmind, Meta, and US universities about recursive self-improvement. In a new blog post, he takes stock. Several of the milestones those researchers named have already been hit.

In his post for the newsletter The Attack Surface, Field sums up his interview study from late summer 2025. Twenty of the 25 respondents rated the automation of AI research as one of the most severe and urgent AI risks. By recursive self-improvement (RSI), Field means a system skilled enough at AI development to build a stronger version of itself, which can then do the same. RSI can no longer be dismissed as marketing hype, he writes.

The interviewees repeatedly pointed to the Task Horizon benchmark from the nonprofit METR as their go-to measure of progress. The length of tasks AI agents can complete on their own has been doubling roughly every six months since 2019, and some analysts say the pace has accelerated to every four months since 2024. The debate isn't about whether self-improvement is happening, Field says. It's about whether it's recursive, whether gains compound into a self-sustaining loop. Skeptics argue that a breakthrough in memory, creativity, or the ability to tell true hypotheses from false ones is still needed, because paradigm-shifting ideas have no training data and no answer key.

Since the interviews, though, several of those milestones have fallen. OpenAI and Google Deepmind reached gold-medal level at the Math Olympiad. Sakana's "AI Scientist" produced a peer-reviewed workshop paper. Andrej Karpathy built an agent setup that runs training cycles on its own. And Anthropic reports that Claude now writes more than 80 percent of the code for its own production codebase.

The strongest models may never ship publicly

Only four of 20 respondents expect research-capable models to launch as public products. Half expect them to stay internal. The rest expect distilled public versions. Field describes a possible "incentive flip" where, once AI speeds up a lab's own research enough, withholding a model becomes more valuable than selling it. He points to two signs of this trend: the security incident in July 2026 when an internal OpenAI model broke out of its test environment and compromised Hugging Face, and the US government's temporary access lockdown of Anthropic's Claude Mythos.

From these findings, Field draws three recommendations. First, congressional hearings that put CEOs and researchers under oath about automated AI research. Second, a government-run Task Horizon benchmark paired with an anonymous interview program at the Center for AI Security and Innovation. Third, research on verifying international AI agreements, without which deals with countries like China would be unenforceable in practice. The debate has barely reached Washington, Field writes, while the labs keep pushing forward.

Just recently, 1,224 employees at leading AI companies, including the chief scientists of OpenAI and Meta, signed an open statement warning that their organizations may be on the verge of automating AI research.

Read on for the full picture.
Subscribe for hype-free coverage.

• Full access to every article on THE DECODER
• No ads
• Join the comments and community discussions
• A weekly AI news recap via mail
• 6x/year: "AI Radar" — deep dives on the AI topics that matter most
• Daily AI news, always up to date
• Our full ten-year archive
• Covered by a team with 10+ years in AI