US-China AI Dialogue Preview: Anthropic Model Pushes for New Negotiations, Huawei Chips Provide the Strongest Card

According to Beating Monitoring, a new round of US-China AI safety dialogue is about to be launched. The immediate catalyst is the frontier model Mythos from American company Anthropic, which can autonomously infiltrate government databases, banks, and hospital networks, and is regarded by the US side as an “unprecedented cyber weapon.” Although Mythos is made in the United States, it proves that this destructive capability is technically feasible; once similar abilities are mastered by opponents or hacker organizations, both China and the US will be unable to withstand it. Meanwhile, DeepSeek has announced for the first time that its new model has been adapted to Huawei chips, indicating that Chinese AI is shedding its dependence on American chip giant Nvidia, adding leverage in negotiations. US officials confirmed that the White House hopes to open communication channels during the upcoming meeting between Chinese and American leaders.

The Biden administration has twice promoted US-China AI communication. During the first round of technical talks in Geneva in 2024, four participants revealed that China viewed the US concern about “AI losing control” as an academic hypothesis and instead pressured for export bans on chips, causing the talks to derail. In November of the same year, the two sides reached a limited agreement in Peru, promising not to let AI control nuclear missile launches. Then-US National Security Advisor Sullivan said the Peru agreement “broke the seal,” demonstrating that China and the US could achieve substantive results in AI negotiations, and recommended that the new team continue the effort during the transition, but this was not adopted. It was only after Mythos appeared that US Treasury Secretary Bostick, after hearing security reports from multiple banks, proposed federal review of future new models, leading to a complete shift in the White House’s stance.

But a more fundamental obstacle is that the two countries are running entirely different races. Silicon Valley giants are racing to develop “Artificial General Intelligence” (AGI, meaning AI with human-level comprehensive intelligence), believing that whoever first creates AGI will gain irreversible technological dominance. China is not pursuing the smartest AI but embedding existing AI into factories, hospitals, autonomous vehicles, and government systems to generate immediate economic value. CSIS senior advisor Kennedy said, “China does not see this as a single race but multiple parallel races.” After sitting down together, the US wants to discuss “what to do if superintelligence loses control,” while China wants to discuss “why can’t you let me buy chips.”

A recent Stanford report states that the performance gap between Chinese and American models has “virtually closed,” but the US still holds advantages: tech companies’ AI investments are ten times those of Chinese counterparts, and the number of data centers (the physical carriers of AI computing power) is also ten times greater. Industry experts suggest tackling specific risks one by one, similar to the Peru nuclear weapons agreement. However, Brookings Institution’s Kyle Chan warned that mutual distrust is fueling a “race to the bottom” in security, with both sides reducing security investments to run faster. Sullivan said that Chinese leadership generally agrees on the necessity of cooperation, but their sense of urgency is far lower than that of the US. “But their sense of urgency has now increased.”