Amodei's pacing plan caps the slowdown at the size of America's lead

Pacing, as Amodei writes it, still allows training and technical progress ^[7]. The essay asks for three things: permanent independent evaluators embedded inside frontier labs, and coordination among leading companies in democratic countries on shared safety standards and capability thresholds. The third is international agreements covering at least some dangerous capabilities ^[7][11]. How much pacing is permitted depends on one variable. Democracies can slow only by roughly as much as their lead over geopolitical rivals allows ^[8]. A larger lead, Amodei wrote, gives democracies the "breathing room we need in order to pace effectively" ^[9]. So the plan's first ask is a bigger lead: tougher restrictions on advanced chips and chipmaking equipment, on unauthorized model distillation, and on theft of model weights. Amodei believes those could widen the American advantage over the next three to five years ^[10].

The second layer is the one with a dependency outside the labs. Coordination on shared thresholds may need government help resolving antitrust barriers ^[11]. Asked whether AI companies should slow development or face more regulation, Trump pointed at China. "We're leading China in AI," he said. "Whoever wins AI wins" ^[12].

Beijing answered in two registers that do not carry the same weight. The Neuron notes that a Foreign Ministry statement is an official government position, while a Global Times commentary signals political thinking and binds nobody. That commentary called the proposal a "Cold War playbook" and argued its restrictions on China would entrench US dominance ^[5][6]. China Daily's objection was procedural. The three-company arrangement, it wrote, "sounds rather like a club whose membership rules have been drafted before the guest list is announced. A global AI-safety framework that excludes China is not quite global" ^[14]. Chinese regulators have said the underlying risk is real: the Standardization Administration of China, working with the Cyberspace Administration of China, says AI development must be monitored as it may go beyond human control ^[15].

The warning that started this came from Jacob Coxon, who quit Anthropic. Frontier labs, he said, are moving toward increasingly autonomous systems before anyone knows how to reliably control much more capable versions of them ^[16]. His concern centers partly on recursive self-improvement, and OpenAI says its own agents are already taking over larger pieces of the research process ^[16][18]. A roadmap from more than thirty Chinese researchers sorted self-improvement into five levels and surveyed nearly 500 papers, finding under 6% at the top tier. It flagged coding as the domain most likely to reach that tier first, since fixes there can be tested instantly ^[19]. Six percent of 500 is 30, so the top-level literature runs to a few dozen papers at most ^[20].

Products kept shipping through the argument. Apple began rolling out its rebuilt Siri, built in part on Google's models and absent from the EU and China at launch ^[24]. Salesforce shipped an in-house reasoning model trained entirely on synthetic data to keep customer information away from outside AI providers ^[26]. According to China Daily, the pressure on frontier spending is commercial: users are shifting demand toward cheaper tokens such as Kimi K3 (low) and DeepSeek V4 Pro over expensive frontier models ^[27].

The one text this week that would actually constrain a model's behavior is Microsoft's draft "Humanist AI" code of conduct, which would require its models to accept being shut off and forbid them from editing their own reasoning logs or reasoning in unreadable shorthand. It is open for public comment, and for now it binds only Microsoft's own systems ^[21].