The U.S. Still Leads China in the AI Race, the Author Argues

Summary

Noah Smith argues that the United States remains ahead of China in the AI race, citing stronger models, greater computing capacity, and much higher company revenue. He says the stakes extend beyond economics: a large and sustained lead in AI-enabled cybersecurity could shift the balance of power because offensive cyber capabilities may not be matched by an equivalent ability to retaliate or defend. At the same time, rapid progress by both countries could create shared catastrophic risks, giving Washington and Beijing an incentive to cooperate on safeguards and possibly on the pace of development. Smith argues that Chinese cooperation may be more likely when U.S. capabilities remain comfortably ahead. He points to the 2025 DeepSeek moment and later releases such as Moonshot’s Kimi K3 and Z.ai’s GLM-5.3 as evidence of Chinese progress. Z.ai reported that GLM-5.3 scored 84.5% on CyberGym, slightly above Anthropic’s Mythos 5 at 83.8%, but the results were not independently verified. Mythos remained substantially ahead on ExploitBench, where it scored 78.0% versus 54.4%, and completed more timed attack-development tasks. Smith also notes that Mythos was an earlier public release and that Anthropic had stronger internal models. He warns that U.S. policy could squander the lead, particularly by driving Chinese AI talent out of the country.