AI’s recursive self-improvement might not come so quickly after all
MIT Technology Review 1 week ago 28
Researchers at Princeton found that AI agents can handle the engineering aspects of AI research but lack the creativity and judgment needed for open-ended investigation, suggesting recursive self-improvement timelines may be longer than industry forecasts suggest. In a test where Claude Opus 4.8 had six days and $3,000 in API credits to conduct novel research on unpublished NeurIPS papers, both resulting papers were rejected by the original authors for lacking novelty and rigor despite competent experimental work. The finding challenges claims that AI will soon improve itself without human oversight, though it remains unclear whether open-ended creativity is essential or whether narrow task improvements alone could enable recursive self-improvement.