Epoch AI's JS Denain: no proof yet of imminent AI self-acceleration

The conversation is framed by a specific question: what are OpenAI and Anthropic actually seeing when they publish warnings about recursive self-improvement? Denain's answer separates public evidence from internal signals. He notes he does not currently think labs hold some hidden measurement warranting far more alarm, while still treating the dynamic of AI accelerating AI progress as worth tracking.

Much of the discussion turns on where AI uplift shows up inside a lab. Denain cites the OpenAI post's finding that Codex spending surged in the spring and then plateaued in the summer for much of the company, but kept growing or accelerating for researchers, the data team, and engineers. Yet he and Lambert both point out that high-level strategic decisions — compute allocation and research direction — showed little measured uplift, which is why Denain distinguishes a 10X productivity gain for the median researcher from a 10X gain for a lab as a whole.

The broader stakes hinge on bottlenecks that sit outside raw model capability. Denain says if robot capabilities were sufficient, financial incentives would drive fast build-out, while Lambert argues physical factories and a long tail of robotics reliability make a two-to-five-year industrial explosion implausible. For readers, the practical read is that the RSI debate may turn less on visible coding gains than on whether hard-to-verify tasks and physical build-out follow.