AI’s recursive self-improvement might not come so quickly after all

A study led by Princeton researchers evaluated AI agents' ability to conduct open-ended AI research and found they can solve engineering problems but lack the creativity and judgment for original research at top conference levels. Using a new "shadow evaluation" method, they tested Anthropic's Claude Opus 4.8 on unpublished NeurIPS 2026 papers and observed limitations in autonomous AI research capabilities. This suggests that timelines predicting rapid AI self-improvement may be overly ambitious