Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI

Nathan Lambert interviewed Jean-Stanislas "JS" Denain, a senior researcher at Epoch AI who leads its Insights Team, in a podcast episode on the state and trajectory of AI. Topics listed for the episode include predictions for recursive self-improvement (RSI), robotics' role in AI acceleration, how far behind Chinese models are, whether distillation explains that gap, what Chinese job postings reveal about their labs, whether open or closed models are safer, how Epoch AI operates, and what a frontier post-training recipe looks like.

In the discussion, Denain said he does not think OpenAI and Anthropic blog posts on AI accelerating AI progress are strong evidence of imminent self-sustaining acceleration or full automation of the AI researcher job. He cited the OpenAI post's finding of a rise in Codex spending by researchers, describing it as a doubling per month, while noting it may be a measurement artifact and saying it is not strong evidence of a software intelligence explosion within six months. He said he does not currently think labs hold specific private information warranting substantially more alarm, and called the public evidence enough to track the dynamic.

On existential risk, Denain said he is extremely uncertain but worried, describing a figure of at least 10% x-risk over a decade-long timeframe as reasonable, referencing Evan Hubinger. He described himself as a "capabilities theory of everything" person and said the principal disagreement concerns how large AI capabilities get and how soon, with diffusion and other factors also mattering. A capabilities-maximalist path, he said, resembles the AI 2027 scenario: AI automating AI research, major robotics progress, and industrial expansion.

Lambert said he agreed with much of a distribution of capability levels discussed on the Dwarkesh podcast with Charlie, Beren and John but was surprised by timelines, including a 10X productivity target for AI researchers. Denain distinguished individual researcher productivity from overall company output, noting compute and other bottlenecks could prevent individual gains from translating into a 10X improvement for an organization.

Denain also said the OpenAI post noted surges in Codex spending among researchers, data teams and engineers that keep growing, while finding no large uplift in high-level strategic decision-making such as compute allocation or research direction choices. Lambert said he expects models to improve at scoped knowledge work, questioned whether harder environments can be produced on an order of magnitude, and called math an exceptional, jagged area; he said a major AI-driven architecture change or breakthrough in less falsifiable fields would update him significantly.