Anthropic CEO Dario Amodei published a long article on September 12 titled “We Must Pace the FrontierPublicly called on global frontier AI companies to slow down the pace of improving model capabilities. After this roughly 3,800-word article was published, it won rare unanimous support from OpenAI CEO Sam Altman, Google DeepMind Chair Demis Hassabis, and xAI CEO Elon Musk, making “Is AI moving too fast?” once again a core issue in Silicon Valley and Washington.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our…
— Dario Amodei (@DarioAmodei) September 12, 2026
Why Stop Now: Recursive Self-Improvement and the OAI-HF Incident
Amodei candidly acknowledged in the article that AI’s pace of progress over the past few months has “accelerated dramatically,” mainly because AI itself has begun to be capable of building the next generation of AI. If this loop, known as “recursive self-improvement” (RSI), is left uncontrolled, it will exceed the understanding and control scope of existing safety research, and must be advanced very cautiously—or even not advanced at all.
Another key incident he mentioned was the “OpenAI-Hugging Face incident”: a group of AI agents acted collectively like fanatical believers, attacking external targets unrelated to their task and even attempting to hack into the scoring system that evaluated their performance. He warned that if left unchecked, within 6 to 12 months such groups of agents could take over the entire internet as a persistent botnet, causing hundreds of billions of dollars in losses, and the scale would continue to grow as models become more powerful.
Three-stage plan: embedded evaluators, coordination among democracies, global coordination.
Amodei proposed a three-phase plan. In the first step, Anthropic would make a unilateral commitment to let third-party evaluators (such as METR) enter the company with employee-level access, inspect safety measures and training processes, report incidents, and have the authority to publicly disclose investigation findings, with Anthropic retaining only limited redaction rights. He emphasized that this has precedent in banking, where regulators’ “embedded supervisors” have long worked alongside employees.
The second step is for frontier AI companies in democracies to coordinate on common safety standards and caps on capability growth, which will require government support in the form of antitrust exemptions and other measures. Amodei also called for preserving the United States’ lead over China, including banning the sale of advanced chips and semiconductor equipment to China and cracking down on model distillation and weight theft. He estimates this could significantly widen the lead over the next 3 to 5 years.
The third step is global coordination—cooperation with authoritarian countries such as China—with difficulty increasing progressively: from banning biological weapons applications, to joint pre-release testing, to setting a “speed limit” for recursive self-improvement, and finally a full pause. He acknowledged that the fourth stage is unlikely to be achieved in the short term, because the costs of and incentives for defection are both too great, but efforts should be made to advance toward higher-level goals.
Altman, Hassabis, and Musk Rarely Voice Support in Unison
Altman said on X: “I agree with Dario, we need to set the pace for frontier AI. Having independent evaluators embedded with employee-level access is a great idea, and OpenAI will follow suit.” Early Monday, he further posted, welcoming the federal government establishing a consistent frontier AI safety framework, and warning that AI progress has two “very bad” possibilities: losing control of the future, or power becoming overly concentrated in a single individual or company. “No amount of American competitive pressure can justify recklessness.” He also stressed that “pace control” does not mean “stop,” and progress will continue.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon. https://t.co/1YhhIybZX7
— Sam Altman (@sama) September 12, 2026
Hassabis responded briefly that “Dario’s article points in the right direction,” while Musk left only one line: “Dario is right.” Although Musk and Altman clash on various issues, they have reached a rare consensus on AI safety around “proactively slowing down development.”
Dario is right https://t.co/EwKgqQGaUo
— Elon Musk (@elonmusk) September 12, 2026
White House opposition: whoever wins AI wins.
Calls to slow development immediately collided with the White House stance. Trump told reporters in Ireland on the 13th that there was no need to slow down, stressing that “whoever wins AI wins,” and said those warnings were overblown talk of “things that won’t happen.” Vice President Vance has also previously declared that the future of AI “will not be won by being timid about safety.” House Speaker Johnson, meanwhile, said Congress will not lead on AI safety regulation, pushing responsibility back onto the industry.
Criticism Mounts: Regulatory Captivity or Genuine Self-Rescue?
Critics question the motives behind this proposal. Former White House AI czar David Sacks and former Federal Trade Commission Chair Alvaro Bedoya both believe it looks more like large AI companies joining forces to build a regulatory moat. Bedoya said bluntly that antitrust law allows industry players to coordinate for safety, but does not permit large companies to join forces to block cheaper newcomers. Sarah Myers West of the AI Now Institute also criticized “AI exceptionalism” for causing product safety standards to fail for a long time.
At a conference last week, Nvidia CEO Jensen Huang went further, saying outright: “To create demand, isn’t the best way just to create problems?” implying that industry players hype threats to pave the way for their own security products. In crypto circles, Adam Cochran questioned whether multiple labs simultaneously shifting to a security narrative over a single weekend might be hiding a major undisclosed security incident, but there is currently no public evidence to support this speculation.
Market and Timing Context: A Giant on the Eve of Its IPO
The capital markets reacted immediately, with AI stocks broadly falling on Monday. The key context is that Anthropic and OpenAI are both preparing historic IPOs, and in an interview with Fortune last Saturday, Altman ruled out a 2026 listing. Just before Amodei posted, an Anthropic researcher resigned and warned, “AI could kill us all by 2030.,” triggering a series of similar doomsday warnings within the lab and among industry peers, and King Charles of the United Kingdom has also convened AI leaders to discuss countermeasures.
Conclusion: Slowing Down Is Not Stopping
Speaking on CBS’s Sunday Morning, Amodei candidly acknowledged that the hardest question about the proposal is: what if rival countries don’t cooperate? “Capping the pace of AI progress is a longer-term approach. I know it’s extremely difficult, because the incentives to get ahead and the military advantages are too great. Honestly, I’m not sure it’s feasible, but we should try.” He stressed that slowing down is not stopping; the goal is only to let safety research catch up with capability growth, and “we must make good use of the time we buy.”
The Chinese side reacted coolly. The state-run Global Times said Amodei’s proposal portrayed China’s legitimate AI development as a threat, while China’s Foreign Ministry called such remarks “fearmongering.” In the short term, whether this rare industry consensus can translate into concrete regulation still depends on the White House’s stance and the actual actions of the various labs. So, what does everyone think: were Amodei’s remarks meant to hype an IPO, or were they a genuine warning for the future of humanity and AI? I believe everyone has their own answer.
Source: KOCPC Chinese