AI Progress Accelerates Beyond Forecasts, Raising Recursive Self-Improvement Concerns

The article highlights a significant acceleration in AI capabilities, surpassing even recent expert forecasts. Ajeya Cotra, a prominent AI thinker, has revised her 2026 predictions, stating that her earlier forecasts for software engineering capabilities now feel "much too conservative." This revision is driven by recent METR results, which show AI systems like Opus 4.6 achieving a "time horizon" of 12 hours, significantly outperforming Cotra's January prediction of approximately 24 hours for the end of 2026. She now anticipates AI agents will achieve a time horizon exceeding 100 hours for software tasks by the end of the current year, suggesting a potential breakdown of the "time horizon" concept as AI approaches multiple full-time-equivalent weeks of work.

This rapid progress signals a potential "software explosion," where AI systems become extremely proficient very quickly, leading to rapid colonization and growth across the economy. The implications extend to the concept of AI R&D Automation (AIRDA), where AI begins to build itself—a phenomenon often termed recursive self-improvement. This is viewed by many as an "event horizon," beyond which the future becomes increasingly difficult to predict.

In response to these accelerating timelines and the profound implications of AIRDA, researchers from GovAI and the University of Oxford have proposed 14 distinct metrics to measure the extent to which AI companies are succeeding in building and overseeing AIRDA. These metrics are crucial for understanding progress towards recursive self-improvement, which is considered a necessary prerequisite for AI building AI. The proposed metrics include measuring AI performance on AI R&D tasks, comparing it to human and human-AI teams, assessing "oversight red teaming" (human supervision effectiveness), quantifying misalignment in AIRDA, computing efficiency improvements, surveying staff on AI use and productivity, examining AI use in high-stakes decisions, and analyzing how AI researchers spend their time.

The urgency of these metrics is underscored by the dual nature of AIRDA: while it could accelerate AI's benefits, it also hastens the arrival of "destructive capabilities," such as those related to weapons of mass destruction or significant societal disruptions like unemployment. The article emphasizes the critical need for robust measurement and oversight as AI progresses towards increasingly autonomous R&D.

_Context: This intelligence report was compiled by the DailyAIWire Strategy Engine. Verified for Art. 50 Compliance._