Epoch AI's Denain on RSI Timelines and the US-China Gap

In the latest Interconnects podcast, Nathan Lambert talks with JS Denain of Epoch AI about three contentious AI topics: recursive self-improvement (RSI), the US-China AI gap, and the 'jaggedness' of model capabilities. Denain argues that RSI timelines are inherently uncertain because progress depends on unpredictable feedback loops between capability gains and research automation.

On the US-China divide, Denain reportedly sees the gap as real but not static, shaped by export controls, compute supply, and talent flows rather than a single breakthrough. The conversation also explores 'jaggedness'—the observation that AI systems can be superhuman in some tasks while failing at simpler ones—which complicates both deployment decisions and safety forecasting.

Lambert and Denain agree that the field lacks reliable metrics for comparing frontier models across countries and that capability unevenness makes broad claims about 'superintelligence' premature. They differ on how quickly RSI could amplify existing gaps, with Denain more cautious about near-term timelines. The episode is a useful primer for anyone tracking AI policy debates without assuming linear progress.