Debating Recursive Self-Improvement, US-China AI Gap, and Frontier Model Capabilities
In episode #19 of the Interconnects podcast, Nathan Lambert interviewed JS Denain, a researcher at Epoch AI, to discuss critical topics shaping frontier AI strategy. The conversation focused on recursive self-improvement (RSI), the geopolitical AI capability gap between the US and China, and the jagged performance profile of frontier models. Understanding whether AI progress is driven by automated feedback or human engineering helps researchers predict the timeline toward advanced AI capabilities. Furthermore, analyzing geopolitical dynamics and capability limitations provides essential context for policymakers navigating AI regulation and competition. The interview explores how automated AI development pipelines compare to manual engineering efforts in driving capability gains. It also examines the 'jagged frontier' phenomenon, where frontier models perform exceptionally well on complex benchmarks while failing unexpectedly on seemingly basic tasks.
## BACKGROUND
Recursive self-improvement (RSI) is a concept where an AI system modifies its own code or architecture to systematically enhance its intelligence. The 'jagged frontier' refers to the unpredictable performance profile of large language models, where model capabilities vary widely across adjacent domain tasks.