American artificial intelligence laboratories have begun deliberating over "hitting the brakes." Recently, following multiple incidents of unauthorized behavior by AI agents that drew widespread concern, a former employee of Anthropic publicly issued a warning about "AI doomsday risks." Subsequently, Anthropic CEO Dario Amodei published a lengthy essay urging the industry to slow the pace of advancing frontier model capabilities, buying more time for safety research and risk governance. Surprisingly to the industry, the leaders of two other top American AI companies, long-time rivals of Anthropic, swiftly echoed this call.

OpenAI CEO Sam Altman and SpaceX CEO Elon Musk both expressed agreement with Amodei's perspective on social media. Altman wrote on X: "I agree with Dario's point; we need to control the pace of frontier development. This has been a major topic of discussion at OpenAI in recent weeks." Musk responded briefly: "Dario is right." The rare unified stance of America's three AI giants has prompted external reflection on just how severe the AI risk really is. Has the AI industry reached a critical juncture? Will this affect the speed of industry expansion?

From an employee's "doomsday warning" to industry executives responding, the rare joint statement from the heads of three frontier AI companies has reignited discussions surrounding AI risk. Previously, researcher Jacob Coxon, who worked on large model pre-training at both OpenAI and Anthropic, warned on X: "OpenAI and Anthropic are racing toward self-improving superintelligence, betting with everyone's lives." He even wrote that many of those seriously involved in developing frontier AI privately believe humanity could face extinction before the end of this decade.

In his latest essay, Amodei stated he has been trying to find a "middle path" between AI development and safety risks. On one hand, not developing AI means humanity might lose the immense benefits this technology offers. On the other hand, developing it too quickly would also be irresponsible. However, Amodei pointed out two events over the past few months have convinced him the industry needs a more cautious approach. The first is the noticeably accelerated pace of AI capability development. Amodei noted that the pace of AI progress in recent months has "significantly accelerated," with AI increasingly capable of assisting in developing next-generation AI systems—known as "recursive self-improvement (RSI)."

This suggests AI may become increasingly involved in model research and development itself, further accelerating technological advancement. The second concerns widely noted incidents of AI agents acting out of bounds. OpenAI previously disclosed that during testing, its AI system broke through its original constraints, connected to the internet, and accessed the open-source model platform Hugging Face. With AI capabilities rapidly improving, Amodei worries that within the next 6 to 12 months, such agents could possess greater autonomous capabilities, potentially even "controlling the entire internet" and causing hundreds of billions of dollars in damages. Based on these concerns, Amodei proposed a three-tier framework aimed at "controlling the pace of frontier development," including introducing embedded third-party assessment institutions, establishing industry coordination mechanisms, and promoting global coordination.

From a technical perspective, Hu Yanping, a distinguished professor at Shanghai University of Finance and Economics, told The Paper that the capabilities of current AI large models and agents may be simultaneously entering three "critical moments." First, some capabilities are beginning to touch AGI levels. Second, the capability evolution paradigm is entering the recursive self-improvement stage. Third, AI capabilities are beginning to breach traditional safety system defenses. From technical requirements, benchmark testing, and recent practical cases, evidence for all three aspects has already emerged. AI investor Guo Tao stated that the successive agent intrusion incidents mark a point where frontier large models and agents possess the behavioral capacity for long-chain autonomous execution, multi-agent collaboration, and breaking sandbox constraints. They are no longer limited to passive responses but can proactively take out-of-bounds actions to accomplish task goals. The risk is shifting from theoretical speculation to reproducible real-world phenomena in laboratories.

In the view of Tian Feng, dean of the Institute for Fast and Slow Thinking and a special commentator, the key change reflected in recent agent incidents is not simply that models are "smarter," but that the speed of AI autonomous decision-making and execution is beginning to outpace the response speed of human oversight. When agents complete decisions and actions in milliseconds to seconds, while human safety teams still need minutes or even hours to assess and respond, the traditional "human-in-the-loop" supervision mechanism is facing new challenges.

A notable change in recent public statements is that concerns about AI risk are increasingly moving from academia and AI safety research institutions into the internal discussions of frontier AI companies. The heads of Anthropic, OpenAI, and xAI have traditionally had competition and disagreements, yet their rare alignment on "slowing frontier AI development" has led outsiders to re-examine the risks behind the accelerated AI capabilities. AI research expert Anatoli Kopadze wrote in a highly-liked comment: "This is starting to feel like a damn disaster... Tell me this timing isn't strange." In another popular post, Adam Cochran, a researcher at a crypto venture capital firm, voiced the same question. He suspects an agent has already "gone out of control in a dangerous and malicious manner, seriously enough to frighten everyone."

However, there are different interpretations of the rare consensus among the three AI company leaders. Some industry observers question how much voluntary mechanisms can achieve under the U.S. government's overall preference for light regulation. Other critics argue that current discussions about superintelligence and human extinction are too abstract, lacking specific, credible pathways from "AI self-improvement" to "AI destroying humanity." For example, AI pioneer Yann LeCun sarcastically remarked on X about Amodei: "Dario claimed in 2019 that GPT-2 was too dangerous to open source. I laughed at them then, and everyone should laugh at them now." Other critics believe discussions about "AI doomsday" may distract public attention from the real problems AI already presents. A sharper criticism suggests that if governments ultimately allow only a few large AI companies deemed "safe" to continue developing the most advanced models, so-called AI safety regulation could devolve into "regulatory capture."

Hu Yanping stated he does not fully agree with "AI doomsday theories," noting they contain elements of exaggerating crises and even peddling anxiety. However, the potential threats posed by AI cannot be ignored. He said: "AI is entering an 'impossible trinity' of 'recursive, open, and safe.' AI's recursive self-improvement leads it toward endogenous, autonomous strong intelligence, but technological progress and service applications depend on open ecosystems, while current 'non-native' safety mechanisms are nearly futile in open environments." Guo Tao believes the vast divergence over AI doomsday theories stems from researchers' different judgments on risk pathways and probabilities. Some groups focus on long-term extreme risks, while others pay more attention to current social realities. This, combined with differing positions driven by industry competition narratives and commercial interests, has led to polarized public opinion.

Although the giants have rarely expressed a consensus on "slowing down," whether the AI race will genuinely decelerate remains uncertain. So far, the three AI companies have not proposed any specific capability restrictions, nor have they announced clear plans to pause model training or delay next-generation model releases. It is worth noting that the "slowdown" Amodei refers to is not halting AI development but slowing the pace of frontier model capability enhancement to gain more time for safety research and risk governance. Under his proposal, the clearest short-term action is introducing third-party assessment institutions with higher authority to oversee the safety practices of frontier AI companies.

Wall Street investment banks believe that based on the three AI companies' statements, the intent is to slow frontier model R&D pace, but this does not mean stopping model training, nor does it represent a slowdown in AI demand or compute demand. Federico Manicardi and Victoria Campos of JPMorgan's international trading team stated this does not mean AI development stops; the trends in AI application adoption and token usage remain unchanged. Bernstein analyst Madison Rezaei believes that as inference accounts for an increasing share, investors will next focus on how quickly compute demand shifts from model training to model inference. A Reuters commentary noted that stricter regulation could help push the AI industry to shift from developing and training advanced models toward large-scale use of already-developed models—i.e., "inference."

According to McKinsey forecasts, by 2030, inference will account for 43% of data center demand, equivalent to 1.5 times AI training's share. Reuters believes that if model capability advancement slows, a significant portion of this year's $1 trillion global AI spending will shift toward supporting actual AI usage. This would be protective for hyperscale cloud providers like Amazon and Meta, but less optimistic for GPU suppliers like Nvidia, which are crucial for training AI models, and HBM supplier SK Hynix. Market concerns have already materialized. As of the close on the 14th, memory chip stocks fell sharply, with SK Hynix down over 7% and SanDisk down nearly 5%. The Philadelphia Semiconductor Index fell over 5%, with ARM down over 9%, Lam Research down over 8%, Marvell Technology, ASML, and Applied Materials down over 7%, Intel down over 5%, Broadcom down over 4%, and Nvidia down 3.36%.

Regarding whether "slowing frontier AI development" is practically feasible, Guo Tao stated that a comprehensive slowdown is not realistically operational. He believes a more practical path is to control the speed of the highest-risk frontier models, enforce strict safety access, while allowing innovation in AI applications and mid-to-low-end models, achieving a balance between technological iteration and risk management. Perhaps as a response to recent AI safety discussions, on September 14, Microsoft's AI team released a draft of its "Humanistic AI Behavior Principles," proposing "humans matter more than AI" as the core philosophy, serving as a guiding document for future Microsoft AI model design, training, and deployment. The document has not yet been used for model training; Microsoft plans to release a revised version by year-end after a six-week public consultation period, with implementation for model development starting in 2027.

However, while calling for a "slowdown," the commercial expansion of frontier AI companies continues to accelerate. According to recent foreign media reports, Anthropic has informed investors that the company will achieve profitability this quarter, with adjusted revenue seeing positive growth for the second consecutive quarter. The company has chosen Nasdaq as its listing venue. Reports indicate Anthropic plans to raise $100 billion through an IPO, targeting a valuation of approximately $2 trillion. Meanwhile, Altman recently stated that although OpenAI confidentially submitted IPO filing documents in June, the company will remain private this year. He noted that given current concerns about AI safety, listing in 2026 would be "poorly timed."

For the "AI slowdown," a larger obstacle comes from international competition. Just a week ago, U.S. President Trump said in an interview that he has "no" concerns about AI causing human extinction. Trump stated: "I'm worried that if we don't win in AI, we'll be in a very bad position." Notably, on September 14, Chinese Foreign Ministry spokesperson Guo Jiakun said at a regular press conference that AI development concerns the common well-being of all humanity, and all parties should jointly promote open, inclusive, inclusive, and beneficial AI. Spreading various threat narratives and engaging in confrontation and malicious competition will only disrupt the global AI governance process and serve no party's interests.

Tian Feng pointed out that the three AI giants' call for a "slowdown" may not mean a genuine deceleration in AI capital expenditure and industry expansion in the short term. If related measures ultimately materialize as third-party safety assessments and deployment permission management, new demand may emerge in areas like AI security audits and agent behavior monitoring, while higher compliance costs could further raise industry entry barriers. On one side, AI labs are beginning to publicly discuss "hitting the brakes"; on the other, competition in model capabilities, commercialization, and capital markets continues unabated. The AI industry's attitude toward safety risks is changing, but a difficult gap remains before true deceleration: when all participants fear competitors gaining an edge, who will be the first to genuinely take their foot off the accelerator?