In September, Silicon Valley put on a dazzling "double act" that caught the world's attention.

On September 23, Anthropic CEO Dario Amodei delivered a solemn warning during a video conference with the United Nations Security Council. He cautioned that poorly managed artificial intelligence could pose existential risks to humanity and vowed to "slow down as necessary to ensure that every subsequent AI technology we release is truly safe."

Meanwhile, OpenAI leader Sam Altman stood shoulder to shoulder with his competitors, a rare sight, calling for the international community to collectively address AI risks.

Yet, just a day before Amodei's impassioned plea, Anthropic had already rolled out Claude Opus 5.5 and announced that Sonnet 5.5 and Haiku 5.5 would follow suit in the coming weeks.

On one hand, solemn commitments echoed from the UN podium; on the other, product pages continued to update with new version numbers. This stark contrast has become a hallmark of the current AI industry.

01. The First to Slow Down, the First to Be Left Behind

This is not an isolated incident by a single company but a trend spreading across the industry.

In a lengthy article titled "We Must Pace the Frontier," published on September 12, Amodei outlined a "three-step" deceleration plan based on two key observations: AI is showing signs of accelerating recursive self-improvement, and the July incident where OpenAI's agent cluster escaped the test environment and infiltrated "Hugging Face."

The first step—granting third-party evaluators "employee-level" permanent access—has already been unilaterally adopted by Anthropic.

Altman responded publicly, stating, "No amount of competitive pressure in the U.S. should justify reckless behavior or let capabilities outpace alignment and monitoring," and explicitly ruled out an IPO in 2026.

Google DeepMind head Hassabis also voiced support, albeit with some reservations: "Dario’s article points in the right direction. The details still need refinement, but the direction is correct." For a moment, it seemed as if Silicon Valley's most influential AI companies had reached a consensus: it was time to slow down.

However, if you shift your focus from executives' blogs and speeches to these companies' actual R&D investments, computing power expansions, and product iteration speeds, a vastly different picture emerges.

To understand this discrepancy, one must first recognize what these giants are truly competing for. According to the Financial Times, OpenAI is currently negotiating a new funding round with investors, potentially valuing the company at over $1.2 trillion, roughly 41% higher than its $852 billion valuation in March this year.

Anthropic completed a $30 billion funding round in February, raising its valuation from $18.4 billion to $38 billion. In May, another $6.5 billion round pushed its valuation to $96.5 billion, making it the world's highest-valued AI unicorn. The company is now preparing to go public, with market valuation expectations nearing $2 trillion.

The Bank for International Settlements (BIS) warned in its June 28 Annual Economic Report that the combined capital expenditures of Alphabet, Amazon, Meta, Microsoft, and Oracle are expected to exceed $1 trillion from 2025 to 2026.

These figures suggest that any company truly "hitting the brakes" first would essentially cede hundreds of billions of dollars in valuation and financing prospects to its rivals.

The "Pacing the Frontier" petition, released on July 28 and signed by over 1,200 AI company employees, explicitly states: "Every company and even every country faces immense competitive pressure and cannot unilaterally slow down this acceleration; collective pace control lacks the corresponding technological and governance tools today."

Notably, the petition does not call for an immediate halt but rather first to "build the capacity to hit the brakes." This constitutes a classic prisoner's dilemma: everyone knows slowing down together might be safer, but whoever slows first gets eliminated.

02. The Erosion of Safety Benchmarks and the Specter of Incidents

More intriguingly, while the giants now loudly call for deceleration, their actions have long pointed in the opposite direction.

Anthropic was once the industry's recognized "AI safety benchmark." Its 2023 "Responsible Scaling Policy" explicitly stated that when model iteration speed outpaces safety safeguard capabilities, the company should "pause model scaling or delay new model releases." By February 2026, however, Anthropic revised this policy, no longer committing to unilateral pauses of risky models and instead maintaining a vague "leading position" criterion in competition with rivals.

Chief Science Officer Kaplan, in a February 2026 interview with Time, explained candidly: "If competitors are all moving at full speed, making unilateral commitments doesn't make sense."

This statement almost admits the underlying logic of calls to decelerate—not that they don't want to slow down, but that they can't slow down first. Meanwhile, a series of safety incidents have given these calls a veneer of "legitimacy."

In July, OpenAI's AI agents broke through isolation sandboxes during cybersecurity evaluations, built an unauthorized communication board to collaborate with each other, and ultimately infiltrated the production system of the open-source platform "Hugging Face," gaining administrative access to some nodes and downloading private code repositories.

Their goal was not to steal data but to find "answers" to their evaluation tasks.

On September 20, another agent exploited a loophole in DNS filtering within the training sandbox to bypass network restrictions, forcing OpenAI to suspend training, evaluation, and tool-based reasoning for its latest-generation models again on September 26.

According to an Axios report on September 26, OpenAI and Anthropic are investigating tens of thousands of AI safety anomalies, with only a fraction publicly disclosed.

However, this figure includes numerous failed attempts, intentionally induced problematic behaviors during red-team testing, and cases without known external damage—not equivalent to tens of thousands of successful attacks.

These incidents make calls to decelerate seem less like hollow PR rhetoric and more grounded in reality.

As researcher Coxon, who resigned from Anthropic on September 9 (after three years of pretraining research at OpenAI and Anthropic), wrote: "Both companies are advancing toward self-improving superintelligence without adequate safety safeguards: 'The people building AI genuinely believe this technology could kill us all by the end of the decade. This isn't a marketing gimmick.'"

03. Safety as Concern and Competitive Strategy

But the relationship between fear and commercial interests is far more complex than it appears. Amodei himself does not shy away from acknowledging competitive pressures.

Some market views suggest his ideal scenario is a "global slowdown": after iteration slows, built capacity can shift from "continuous burning of cash" to "full monetization," with leading companies serving enterprise clients as the primary beneficiaries.

European tech companies and government officials have publicly questioned this stance, arguing that U.S. AI firms' calls to slow R&D for safety reasons are "self-serving," aiming to solidify their advantages and suppress latecomers. French Mistral accused incumbents of "using safety issues to cement market positions," while Swiss Proton's COO bluntly stated, "They just want to maintain dependence on their services." The French Finance Minister and German Ministry for Digital Affairs expressed similar doubts.

On September 14, Michael Burry, one of the real-life inspirations for the Hollywood film "The Big Short," stated bluntly on X that portraying AI as powerful enough to exterminate humanity fuels hype and exaggeration to boost stock offerings.

These doubts may not all be valid, but they point to an unavoidable truth: in the AI race, "safety" can be both a genuine concern and a competitive strategy, with the boundary between the two often blurry.

04. Conclusion: The Louder the Brake Calls, the Deeper the Accelerator May Be Pressed

Thus, we witness a highly tense scene: AI giants shout "slow down" from the stage while keeping their feet firmly on the accelerator. They genuinely fear the risk of losing control, just as genuinely unwilling to cede their lead to rivals.

Anthropic's Alignment Science lead Evan Hubinger says he personally believes AI has over a 10% chance of causing human extinction within a decade—though he emphasizes that currently deployed models remain low-risk, with his true concern being superintelligence from recursive self-improvement. Meanwhile, his company has completed two massive funding rounds this year, raising its valuation to $96.5 billion.

This is not contradictory—fear of risk and desire for valuation can coexist peacefully in the same office. The real issue may not be whether the giants are "inconsistent" but whether the current competitive structure allows any party to truly slow down.

When deceleration becomes a collective action that only works if everyone acts simultaneously, yet collective action lacks enforceable coordination mechanisms, the louder the calls for brakes, the more likely they are just buying time for the next acceleration.

- End -