OpenAI assesses agents revising model training
OpenAI evaluated autonomous coding agents tasked with writing, modifying, and monitoring training software for subsequent neural networks. While agents showed strong gains in task automation and monitoring runs, uneven capabilities across complex infrastructure tasks mean sustained recursive self-improvement remains unproven.
AI models modifying their own training pipelines is the practical start of recursive self-improvement, but the jagged frontier between routine monitoring and complex system engineering proves human engineers remain essential.
- –Autonomous completion rates on 2-to-4-hour research tasks reached 46%, yet success dropped to 25% on inference and serving-stack engineering.
- –The 79% success rate on monitoring training runs demonstrates agents excel at surveillance and routine remediation rather than novel architectural breakthroughs.
- –Researchers burning thousands of dollars daily in agent compute highlights high tooling utility without establishing compounding autonomous loops.
- –Without general breakthroughs in research taste and long-horizon planning, automated self-improvement hits rapid diminishing returns.
DISCOVERED
31m ago
2026-11-20
PUBLISHED
1h ago
2026-11-20
RELEVANCE
AUTHOR
Tugrul_Guner
