JS Denain, senior researcher at Epoch AI and head of its Insights team, believes that current public data are insufficient to establish an imminent acceleration of AI progress. In an interview with Nathan Lambert, he revisits the uncertainties surrounding this trajectory and the gap between Chinese and American models.
- JS Denain juge que les données publiques ne prouvent pas une accélération imminente du progrès de l’IA.
- Il situe le retard des labs chinois à quatre ou huit mois, Nathan Lambert le estime à deux ou quatre mois.
- Sans distillation, Denain prévoit un écart plus important, tandis que Lambert accorde plus de poids à la puissance de calcul.
- Les données d’entraînement restent difficiles à suivre, et Epoch existe pour publier des tendances fiables.
Acceleration still uncertain
Posts published by OpenAI and Anthropic examine the effect of AI on its own progress, a dynamic called recursive self‑improvement (RSI). For JS Denain, the most striking point of the OpenAI post is the rise in researchers’ Codex spending, which is increasing 2× per month according to the presented curve. He notes, however, that this metric may only pertain to Codex, not to ChatGPT usage, and that it does not constitute solid evidence of a software explosion within the next six months.
When asked about the link between RSI and existential risk, he says he is very uncertain, but concerned. He cites Evan Hubinger’s estimate that the extinction risk would be at least 10 % over an unspecified period, and considers this estimate reasonable over a decade. He does not have, in his view, internal information that would justify a much greater concern.
A delay of four to eight months
JS Denain places the gap between the top Chinese labs and OpenAI or Anthropic at four to eight months, based on public release dates. He believes a two‑month gap would be more indicative of test over‑adjustment. Nathan Lambert, however, judges that models such as Kimi K3 and GLM 5.2 would be closer, at two, three or four months.
Distillation, which involves training a model from the responses of another, could play a larger role than expected. Denain estimates that if it were halted, the gap would widen from roughly six months to eight or nine months over a six‑to‑twelve‑month horizon. Nathan Lambert remains more cautious: in his view, without distillation the gap would be fairly close, and the compute difference would weigh more heavily.
An analysis of job postings, conducted with Cheryl Wu and published in June, also suggests that at least some Chinese labs are building their own data centers rather than renting compute.
Security and data
Nathan Lambert mentions an incident in which three individuals allegedly used public Claude models to access OpenAI’s internal systems. JS Denain says he was briefly made aware of it, without knowing whether the issue stems from model safeguards or corporate security. He judges that open models are less protected, but also less powerful.
Finally, Denain explains that Epoch was founded because neither industry nor academia offered a public, tracked resource on major AI trends. He notes that training data remain far harder to track than compute, while IPOs, by multiplying, could improve the information available.
