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

Nathan Lambert’s latest Interconnects episode is basically a long argument about how scared people should be right now. His guest, Epoch AI’s JS Denain, is not waving away the risk. He’s just not convinced the loudest claims about imminent RSI have the evidence to match them.

Denain points to the public material from OpenAI and Anthropic on AI accelerating AI work, but says it doesn’t add up to strong proof of a self-sustaining burst. The most interesting signal, in his view, is the rise in Codex spending by researchers. That looks like real value. But he says it could also be partly a measurement quirk, since it may capture only one slice of usage while other work happened through ChatGPT.

The bigger theme is uncertainty. Denain says he doesn’t think there’s some secret internal fact at OpenAI or Anthropic that should make outsiders panic much more than they already do. Still, he thinks the public evidence is enough to take the feedback-loop idea seriously, even if nobody can say when it bites hard. He also says the recent tone from some labs is not, by itself, evidence that their internal metrics have gone off the rails.

Where the conversation gets sharper is on what AI progress would actually look like. Denain leans toward a capabilities-heavy view: if the systems get huge gains fast, the real-world effects could include AI automating AI research, big jumps in robotics, and even factories run by AI. But he keeps coming back to bottlenecks. A model might make researchers faster on coding or troubleshooting, yet still fail to speed up strategic decisions, compute allocation, or the company as a whole by the same factor.

That’s the thread running through the whole discussion. Maybe a median researcher becomes far more productive. Maybe Anthropic or another lab doesn’t move nearly as fast anyway because compute, management, and judgment still matter. Denain’s point is not that acceleration is fake. It’s that people keep jumping from one useful signal to a much bigger conclusion than the evidence supports.