RSI is not happening [R]

A new paper suggests that Recursive Self-Improvement (RSI) is not imminent, as current AI agents like Codex/GPT-5.6 Sol and OpenClaw/Opus 4.8 cannot perform open-ended machine learning research. The study involved having these agents attempt to replicate accepted, but unpublished, NeurIPS papers. The original authors graded the agents' work, finding they were unable to advance the research, which challenges the idea that AI agents can accelerate AI research by taking over entire projects.

Unlike many theoretical discussions, this paper directly tests AI agents like Codex/GPT-5.6 Sol and OpenClaw/Opus 4.8 against human-level research tasks, providing empirical evidence against imminent RSI.