AI recursive self-improvement might not come so quickly after all (August 2026) (www.technologyreview.com)

A recent study led by researchers from Princeton University indicates that the ambitious goal of achieving recursive self-improvement in AI may be farther away than previously thought. While AI models, including Anthropic's Claude Opus 4.8, can effectively solve narrow engineering problems related to AI research, they falter significantly in open-ended research tasks. The study employed a new evaluation method called "shadow evaluation," where AI agents were given the task of answering complex research questions from unpublished papers submitted to the NeurIPS 2026 conference. Despite their capabilities in managing experimental setups and data analysis, the agents' submissions were ultimately rejected due to their lack of creativity, judgment, and effective use of feedback—critical components needed for innovative research. This finding has significant implications for the AI/ML community, as it challenges the optimistic timelines for fully automated AI research and highlights a crucial gap in AI capabilities. Researchers suggest that while AI can excel in narrow tasks that have straightforward success metrics, the absence of intuitive creativity may hinder progress in more abstract, open-ended areas of inquiry. The study underscores the necessity to rethink training methods for AI systems to foster the kind of innovative thinking required for groundbreaking advancements, raising important questions about the future trajectory of AI development and its potential for self-improvement.