TL;DR
Chinese AI researchers are not passively absorbing official narratives about US AI safety talk — they are radicalizing into their own version of it.
A researcher at DeepSeek, Liu Shengyu, wrote a blog post about his years spent optimizing kernels. Give it another six months to a year, he says, and AI-written kernels will likely match or beat his. So why keep going? Because if he slackens now, other companies will build their models anyway and end his career just the same. When everyone is this fixated on destroying themselves, he has no choice but to join the arms race.
What actually scares him
He sketches two futures: one where productive forces are massively liberated and living standards rise for everyone, and one straight out of Cyberpunk 2077, where a handful of tech companies hold most resources, a tiny minority gets the strongest AI, and everyone else makes do with a feeble version.
Which is why he most fears Anthropic holding the frontier forever. His words are blunt: the severity would be no less than letting Hitler get the atomic bomb before the Allies. That comparison is indefensible, but he plainly believes he is answering an accusation in kind.
The post was not taken down. Huxiu, 36Kr and state-owned Yicai covered it sympathetically, and the comments on Zhihu and Xiaohongshu are overwhelmingly supportive.
The conspiracy theory became the technologists' own
Many Americans sincerely believe liberal democracy should keep AI dominance, which makes chip export controls a natural conclusion. They do not feel they are acting.
That sincerity is hard to read from Beijing, which habitually looks for balance sheets and polls behind American motives, for a coordinated anti-China plot. In the past the story needed conjured partnerships; this time Anthropic's CEO, Dario Amodei, posted six essays all advocating export controls to China. The evidence arrived courtesy of Silicon Valley itself.
State media runs the old playbook, and it works. The more troubling part is the layer above: researchers at the very front of capability are drawing conclusions more radical than the Party's.
Liu graduated from Peking University in 2025 with offers from Carnegie Mellon and Berkeley and skipped them for financial reasons. A few years earlier and he might have spent a summer at OpenAI or Anthropic. That path is gone; American labs are now, for people like him, humanity's Goliath.
Why it matters
The story is not that Washington and Beijing are arguing again. It is that the people who understand the technology best no longer share any premises: Americans treat safety warnings as sincerity, Chinese researchers treat them as tactics. Once the two sides cannot agree on what good faith even looks like, coordination runs out of language.
Curated from high-quality sources, with concise summaries and key takeaways.
