A DeepSeek engineer says he especially fears Anthropic reaching AGI

The pacing debate sounds different from inside a Chinese lab. An engineer at DeepSeek published an essay naming the outcome that worries him most, and it was not an American rival beating China. It was Anthropic winning.

The essay

Shengyu Liu is described by the South China Morning Post as an engineer behind the company’s V4.1 models, and says himself that he wrote the main attention kernel powering V4.1 Flash, the low-level code that makes the model fast. His post went up on WeChat on Monday, September 14, 2026, under a title that OfficeChai translates as I Have No Choice but to Bury My Talent in Yesterday.

Liu wrote that he does not trust Anthropic or OpenAI to make cutting-edge AI open and affordable, and that he especially does not want Anthropic to master AGI. The South China Morning Post, reporting on the essay, notes that he invoked Nazi Germany to make the point about concentrated control.

He frames the future as a fork. Either the productivity gains get shared broadly, or a handful of companies hoard the strongest models in what he calls a Cyberpunk 2077 outcome. His stated reason for staying at DeepSeek is to keep frontier models open rather than to move to a better-funded lab.

A different argument for open weights

The Western version of this debate is mostly about safety, and it treats open weights as a risk to be managed. Liu’s argument runs on a separate axis. His case for open models is not that they are safer but that closed ones decide who ends up owning the capability.

That distinction matters for anyone reading the pacing letters from US labs as a global consensus. The engineers building the fastest open models are not weighing the same tradeoff, and a slowdown argued on safety grounds reads, from that side, as a decision about market position.