Study: AI Agents Cannot Yet Conduct Open-Ended AI Research

sebkrier · x · 2026-08-09

A new study suggests that AI agents are not yet close to recursive self-improvement (RSI), despite progress on narrow, verifiable tasks.

Researchers partnered with authors of two unpublished AI papers and tasked frontier AI agents with answering core research questions using thousands of dollars in API credits and six days of compute. The original authors ultimately rejected both agent-generated papers.

Log analysis revealed that agents lacked the judgment for open-ended research. While they occasionally proposed impressive directions, they typically rejected these ideas quickly based on low confidence.

Original post →

More from AGI Musings

AGI Musings channel →