United States · China · AI safety
Washington would like Beijing to call if the AI causes an international incident
Sunday's U.S.–China talks produced a proposal. Treasury Secretary Scott Bessent said Washington suggested an AI-safety notification mechanism during his meeting with Chinese Vice Premier He Lifeng in New York. The idea would create a channel for incidents that rise to the level of national security and is expected to be considered by Presidents Donald Trump and Xi Jinping this week.
This is not an agreement. China's public response has not supplied the rules that would make the channel operational. The two sides would still need to define a reportable incident, the deadline for disclosure, the evidence shared, and how to distinguish state direction, outside compromise and an autonomous system exceeding its assignment.
Those ambiguities are exactly why the proposal matters. A serious AI-enabled cyber incident could be misread as deliberate state action while the governments are still working out what happened. Moving from opaque behavior toward notification would be a real diplomatic safeguard if both sides adopt and use it. For now, it is a concrete proposal with consequential details still missing.
Source: Reuters ↗TSMC · Advanced packaging · Taiwan
The next AI-chip factory needs a neighborhood for everything surrounding the chip
Taiwan broke ground on an advanced-packaging industrial park in Kaohsiung. TSMC plans a technology-validation laboratory and talent-training center there, both expected to begin operating in the fourth quarter of 2029. The Baipu park covers about 88.7 hectares, with roughly 53.6 hectares designated for industrial use.
The important phrase is “advanced packaging.” Modern AI accelerators increasingly combine multiple chips and stacks of fast memory inside one closely connected package. Linking those components without sacrificing bandwidth or wasting power has become a central engineering problem. Reuters notes that TSMC packaging technologies already support processors designed by Nvidia, AMD and Broadcom.
The park is meant to bring materials suppliers, equipment makers, validation and training closer together so new technology can reach manufacturing faster. Taiwan also says it has assessed electricity needs through 2035, a reminder that every cloud roadmap eventually becomes a physical one. Someone still has to manufacture the chip, package it, cool it and plug it into the wall.
Source: Reuters ↗China · Humanoid robots · Commercial reality
China's robot boom reached the dangerous question: does anybody buy them twice?
Chinese regulators are slowing some humanoid-robot IPO applications as they examine whether soaring valuations and revenue tied to state-supported projects reflect durable commercial demand. Reuters reports that the volatile debut of Unitree Robotics helped trigger the scrutiny: its shares rose more than fivefold after listing and then fell 55% from their peak.
The review does not mean Beijing has abandoned humanoids. China treats “embodied intelligence” as a strategic industry. The standard of evidence is shifting, though. Investors increasingly want factory deployments, order volumes and repeat customers rather than brief demonstrations or data-collection projects backed by local governments.
That is the distinction robotics needs. A machine can look capable for three minutes; commercialization asks whether it performs boring work safely and cheaply for years. It must pick boxes, tighten bolts, break infrequently and return tomorrow. A backflip is impressive. An invoice from a repeat customer is harder.
Source: Reuters ↗