Back to feed
News Story
APriority74
THE DECODER
1 sources

Top AI Lab Researchers Warned About Automated AI Research, and Several Predicted Milestones Have Already Been Hit

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, noting that several of the milestones those researchers predicted have already been achieved.

SynthePulse Insight · AI deep readingMembers

Automated AI Research: Milestones Reached, but Recursive Self-Improvement Remains an Open Question

Version 1 · 1 source

An interview study of 25 leading AI researchers reveals that automated AI research is seen as one of the most urgent risks, and several predicted milestones have been achieved. However, whether recursive self-improvement is truly occurring remains a point of contention.

  • IAPS researcher Severin Field interviewed 25 researchers from OpenAI, Anthropic, Google DeepMind, Meta, and US universities, 20 of whom ranked automated AI research among the most severe and urgent risks.
  • Researchers commonly use METR's Task Horizon benchmark to measure progress; task length has roughly doubled every six months since 2019, accelerating to about every four months after 2024.
  • Several predicted milestones have been reached: OpenAI and Google DeepMind achieved gold-medal level at the Mathematical Olympiad; Sakana's AI Scientist produced a peer-reviewed workshop paper; Karpathy built an agent that autonomously runs training loops; Anthropic reported that Claude writes over 80% of its production codebase.
Open section navigationInterview Background and Key Findings

Interview Background and Key Findings

In late summer 2025, IAPS researcher Severin Field conducted an interview study with 25 researchers from OpenAI, Anthropic, Google DeepMind, Meta, and US universities on the topic of recursive self-improvement (RSI). Field summarized the findings in his newsletter, The Attack Surface.

Twenty of the respondents ranked automated AI research among the most severe and urgent AI risks. Field defines RSI as a system whose AI development skills are sufficient to build a more powerful version of itself, and so on. He argues that RSI can no longer be dismissed as marketing hype.

Free for now

Read the full analysis

3 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article is based on a report from THE DECODER, which relayed Field's blog post and interview study. All specific figures and conclusions come from this secondary source and have not been verified with primary data.

Sources for this version

  1. Top AI lab researchers warned about automated AI research, and several of their predicted milestones have already fallenTHE DECODER

Primary report

THE DECODER

Primary source