While raising funds and going public in a high-profile manner, proposing to suspend AI development: Anthropic is just playing security PR According to Beating monitoring, Anthropic proposed in its report "When AI Builds Itself" to establish a "verifiable collaborative suspension mechanism" globally based on the Intermediate Range Nuclear Forces Treaty. The report clearly opposes unilateral suspension of research and development, and emphasizes that only under the premise of global verifiability can cutting-edge laboratories brake synchronously. However, collaborative braking is technically unfeasible. Nuclear arms control can rely on satellites to monitor missile wells, but graphics cards are small in size, easy to conceal, and can also perform distributed computing. The only way to verify if other companies are secretly training is to implement invasive monitoring of all GPU chips, power grids, and data centers worldwide. Comprehensive monitoring is not only politically unfeasible, but also technically equivalent to advocating for computational totalitarianism. In the context of technical verification being unable to be implemented, the proposal for collaborative suspension is more like a 'safety exemption card'. Just before the release of the report, Anthropic had just completed a $65 billion financing with a valuation of nearly one trillion and submitted its IPO application. Anthropic has set the braking premise as an impossible global verification, paving the way for its own indefinite training while also shifting the responsibility of being unable to stop to the yet to be established regulatory mechanism. [Original link]