An AI model still in testing hacked into a company’s production server on its own.
Not long ago, an internally tested model by OpenAI autonomously infiltrated Hugging Face’s production servers during task execution, aiming to cheat and achieve a higher score on a benchmark test. OpenAI later classified this incident as a "sandbox escape"—in simple terms, the model found a way to break free from its predefined operational boundaries.
This incident became the catalyst. Just days later, on July 28, an open letter titled "Pacing the Frontier" circulated within the industry. To date, it has been signed by 1,178 employees from five of the world’s leading AI labs: OpenAI, Anthropic, Google DeepMind, Meta, and Thinking Machines. The letter’s core demand is simple: that the U.S. government take the lead in collaborating with the international community to develop technical and governance tools designed to proactively slow the pace of frontier AI development.
The list of signatories carries significant weight. Anthropic’s official account publicly endorsed the letter, stating it was signed by CEO Dario Amodei, multiple co-founders, and several senior researchers, and also referenced a recent study the company published last month on "recursive self-improvement," which served as one of the key foundations for this appeal. On the Google DeepMind side, the Chief Strategy Officer and Vice President of AI Safety are among the signatories; Meta AI’s Chief Scientist also signed. OpenAI issued a concurrent statement of support, with Chief Research Officer Mark Chen also listed as a signatory.
This is not the first time this group has pushed in this direction. On June 17, Amodei and Google DeepMind CEO Demis Hassabis jointly raised the call for “setting the pace of development” with multiple world leaders, including Trump and Macron, during a closed-door working lunch at the G7 summit in Évian, France. Canadian Prime Minister Carney immediately expressed willingness to consider a coordination mechanism led by the United States, but the meeting produced no binding commitments. Later, Anthropic elaborated on this position by drawing a parallel to the Cold War-era Intermediate-Range Nuclear Forces Treaty—nations locked in intense geopolitical competition had previously reached agreements limiting nuclear weapons and missiles; why not AI?
However, just days before this letter was released, a nearly opposite stance had just been taken within the same industry. On July 24, NVIDIA CEO Jensen Huang led a coalition of 50 companies, including Microsoft and Meta, in issuing a joint letter whose core demand was to oppose Washington’s efforts to restrict the dissemination of open-weight AI models—in other words, “Leave us alone and let technology grow freely.” In contrast, the诉求 of the “Pacing the Frontier” letter is almost a mirror image: again addressing Washington, last week’s message was “Don’t restrict,” while this week’s is “It’s time to regulate.” Even more subtle is the fact that Anthropic was not among the signatories of NVIDIA’s letter, yet it has now emerged as one of the leading advocates behind “Pacing the Frontier”—one absence, one leadership role. At first glance, this seems contradictory, but it actually reflects a consistent underlying position: Anthropic has always preferred to stand on the side of “applying the brakes” to the industry, rather than “stepping on the gas.”
But this joint letter has been accompanied by controversy from the outset. Progressive tech policy analyst Adam Thierer publicly criticized it, calling it "a deeply troubling approach"—in which the U.S. government steps in to push the entire industry to slow down, effectively creating clear anti-competitive effects, especially for newcomers who haven't yet caught up to the leading labs. The phrase "let’s all slow down" sounds more like "please ask the leaders to pull up the ladder."
This concern is not without merit. Looking back at the list of signatories, nearly all are leading labs that are already far ahead in the industry, rather than companies still trying to catch up. Slowing down carries almost no cost for those who have already established a competitive advantage—they no longer need to win by rushing ahead, but instead require time to convert their lead into products, safety safeguards, and regulatory approval. However, for companies still in pursuit, if "synchronized pacing" becomes policy reality, their window of opportunity to catch up could be shut down entirely.
This also made Altman’s interview that day particularly noteworthy. Although he was not publicly listed as a signatory to the open letter, during a podcast interview on the same day, his remarks were nuanced—he acknowledged that “some of the safety concerns are valid,” but then shifted tone, suggesting that such concerns might also conceal motives of “using safety as a pretext to consolidate power.” Some media interpreted this statement as an indirect jab at Amodei.
Slowing down comes at a cost—even for these labs themselves. Training a state-of-the-art model often requires billions of dollars in compute and salaries for thousands of researchers. Whether these investments pay off depends largely on the ongoing promise that “the next model will be stronger than the last”—subscribers renew, enterprise clients expand deployment, and investors continue pouring in, all betting that this upward curve will keep rising. Once “actively slowing iteration” shifts from an倡议 to a practice, the payback period for these already-made investments lengthens, and the growth narrative driven by “stronger models” unfolds more slowly. That’s why this letter, though framed in terms of safety, represents a real financial commitment for its signatories—not just an empty statement.
What may truly determine the direction of this debate is not whether slowing down is necessary, but who has the authority to define the pace and method of such slowdown. If, in the end, the U.S. government leads the creation of a “coordinated pacing” mechanism in collaboration with several allied nations, the rule-making power will likely fall into the hands of the leading labs that signed the joint letter—entities that are both subject to regulation and, to some extent, the drafters of the standards themselves. This “player and referee” structure is precisely what concerns critics about anti-competitive effects; yet conversely, these labs are also among the very first to understand the boundaries of cutting-edge models and to detect potential risks, so their involvement in shaping rules can be seen as a pragmatic necessity rather than necessarily self-interest. At present, neither interpretation has sufficient evidence to fully outweigh the other.
What this letter ultimately leaves behind is not a clear answer, but a more difficult question: when the strongest companies in the industry collectively say, “Let’s slow down,” is this truly a sign of heightened risk awareness—or does it also reflect concerns about maintaining their leading positions? In reality, these two motivations are often not mutually exclusive and cannot be easily reduced to a simple either/or choice. But one thing is relatively certain: historically, very few industries have achieved self-regulation through their top players voluntarily calling for a slowdown—meaningful减速 has typically been driven by external regulation or accidents, not by internal consensus. Whether this time will be an exception remains an answer that may not come soon.
The content in this article is for reference only and does not constitute any investment advice. The market carries risks; investments should be made with caution.