This op-ed is worth your time not for model benchmarks but for how clearly it maps the definitional gap on AI safety. On the US side, Anthropic's Dario Amodei worries about the "adolescence of technology" — models getting powerful enough to harm humanity — so he bakes Universal Declaration of Human Rights values into Constitutional AI. On the Chinese side, Z.ai chief scientist Tang Jie's July staff letter and August op-ed call for embedding national laws and security concerns directly into model foundations, and he wants legislation to enforce ideological alignment. National Security Minister Chen Yixin's recent article doesn't mention runaway AI at all; the threat is hostile forces using AI against "political, institutional, and ideological security."
On the product level, Z.ai's GLM-5.3 is open-weight and close to Anthropic's Mythos 5 in cyber defense, but discovered vulnerabilities get reported to a database run by Chen's ministry. The author flags a real concern: deeply embedded behavioral rules in open models are hard for downstream users to strip out, so those values travel with the weights.
I'd read this as background for setting expectations around Thursday's summit, not as technical analysis. The author's own conclusion: the two sides don't share a definition of "safety," so room for substantive agreement is extremely narrow.