Besent (贝森特) says U.S.–China AI gap under six months; Stanford research says difference nearly erased

Claim and context

It has been reported that Besent (贝森特), a Chinese AI firm, believes the performance gap between leading U.S. AI models and Chinese equivalents is now less than six months. Stanford University researchers, reportedly in a recent analysis, conclude the difference has been nearly erased. Those are striking assertions: if true, they imply that techniques and engineering workarounds — not just raw compute — are driving parity.

How parity is being achieved

The key technical story is distillation and model compression. Distillation — popularised in a 2015 paper by Geoffrey Hinton and colleagues, and prefigured by model‑compression work dating to 2006 — teaches a smaller model to mimic the outputs of a much larger one. In China’s developer community this idea has spread beyond pure engineering: people are "distilling" expert judgment into reusable "skills" (for example, community projects that attempt to capture a public figure’s decision patterns). Reportedly, these practices make it possible to deploy much of a big model’s behaviour on far cheaper hardware, accelerating local progress.

Geopolitics, supply constraints and social questions

Why does this matter geopolitically? U.S. export controls and chip sanctions have limited Chinese access to the highest‑end accelerators, so architectural tricks and data‑efficient methods have become strategic. Distillation reduces reliance on top‑tier silicon and can compress capability into models that run on more widely available hardware — a pathway to rapid catch‑up. But there are social and ethical stakes too: as Chinese commentary has noted, “skill” distillation raises questions about labour, consent and what it means to turn human expertise into replaceable software. Will policy and corporate governance keep pace with the technical pull to close the gap? That is the question now confronting regulators and industry on both sides of the Pacific.