TL;DR
The best way to make China take superintelligence risk seriously is not for America to run faster, but to let its leaders see that their own open-weight models can be turned against them.
A 27-year-old researcher quit Anthropic, saying the thing he and his colleagues were building could wipe out humanity. People say this every few months, but this time it reached the evening news, and Barack Obama started telling Democrats to pay attention.
The genuinely strange part is how the top labs responded: they asked for the slowdown themselves.
Why bosses would give up revenue
Slowing down normally means handing the lead to a challenger and losing money. Dario Amodei, who runs Anthropic, published a long post calling for the frontier to be paced, with a three-step plan in which the labs police each other. Musk, Altman and Hassabis quickly agreed.
In the author's reading, speaking against their own financial interest means the fear of extinction is sincere. They feel stuck in a race they cannot leave: stop, and someone else builds it anyway, so each has to finish first and make sure the safest system — their own — is the one that ends up dominant.
The obvious public question is how software in a data centre kills real people. The author's answer is bioweapons: AI can design viruses, and viruses can kill. Biologists say the task is too hard; the author tells readers to ignore them. Experts keep mistaking their own difficulty for universal difficulty, as they did with natural language and with hard mathematics.
If you cannot stop China, frighten it
The main objection to pacing is that America slows, China overtakes. The author once argued that simply staying ahead would make Beijing choose to slow down too — and now admits that was plainly wrong: if China only slows when America does not, the game is over.
So change what China fears. Right now it worries about Washington using American models against it, which only makes it accelerate. Give it something else to worry about: use China's own open-weight models, GLM or Kimi, to steal its secrets and break into its best-guarded accounts, then post a few harmless ones on RedNote.
The point is not real warfare but a demonstration: your rush toward superintelligence is making you vulnerable to any domestic dissident who decides to use it. Beijing has never hesitated to crush its own tech companies, as it did in 2021.
Once the leadership sees that superintelligence threatens its own grip on power, the author predicts, a whole range of international cooperation suddenly becomes possible.
Strip it down: they fear the enemy, but they fear their own people more.
Curated from high-quality sources, with concise summaries and key takeaways.
