Analysis of OpenAI Data Suggests AI Productivity Gains Fall Short of Self-Sustaining Feedback Loop

An external analysis of OpenAI's internal research metrics estimates that current AI models provide a 2-3% productivity gain per unit of capability increase. This figure is derived from OpenAI's logged experiments, which show a 1.6x increase in experiment volume alongside frontier model updates. The analysis contrasts this with theoretical thresholds for a self-sustaining recursive self-improvement loop, estimated at 15-19% productivity gain per capability point. Consequently, the data suggests a gap of five to ten times between current performance and the level needed for autonomous AI development to accelerate independently. The author notes that while benchmarks suggest longer task horizons, actual autonomous success rates in research tasks remain limited, indicating that diminishing returns persist despite increased automation.