Jensen Huang goes left, Amodei goes right

Original author: Su Yang

Original editor: Xu Qingyang

Original source: Tencent Tech

Whether frontier AI should keep accelerating or hit the brakes has become the most contentious topic in Silicon Valley today.

As the debate rages, Silicon Valley giants have begun taking sides. Amodei, Altman, Musk, and Hassabis have formed the "pause camp." Jensen Huang is a staunch member of the "accelerate camp," dismissing "AI doomerism" as lacking scientific basis. Yann LeCun, for his part, has continued to mock Amodei, saying: "When (Amodei) was still at OpenAI, he thought GPT-2 was so dangerous it shouldn't be open-sourced. I mocked him then, and everyone can keep going now."

By comparison, Zuckerberg, Nadella, and others have taken a relatively neutral stance.

The fact that these giants have each chosen sides means Silicon Valley has failed to reach a consensus on pausing frontier models. This stems from differences in each company's ecological niche within the AI industry chain. Model-layer leaders can call for a pause for various reasons, but getting Oracle, AWS—with their hundreds of billions of dollars in compute orders—or Nvidia, perched at the top of the food chain, to stop is simply not going to happen.

In Washington, the call to "pause" has not only failed to gain support but has even triggered a series of doubts.

"It feels a little bit to me like a bit of a Trojan horse." JD Vance said in an interview. This remark actually reflects Washington's real view—when the company building the Trojan horse repeatedly asks for the gate to be made narrower, it's time to examine exactly which companies that gate is suited for.

Pause, Warning, and a Resignation Letter

The trigger for this debate was the resignation of an Anthropic researcher named Jacob Coxon.

He had previously spent three years doing pretraining research at OpenAI and Anthropic, personally participating in the training of the most frontier models. In a long post on X, he showed almost no pleasantries: "Neither company is acting responsibly. They are racing headlong toward self-improving superintelligence, gambling with our lives."

He stated bluntly that people inside the labs genuinely believe AI could kill all humans within a decade, and that this is absolutely not a PR performance. Many colleagues have already begun using terms like "critical moment" and "endgame" to describe the current phase. Coxon even gave up equity that was about to vest. He had only been at Anthropic for four months, and company rules required six months before vesting.

Former Anthropic researcher Jacob Coxon

Immediately after, Evan Hubinger, who leads alignment science at Anthropic, publicly responded: Jacob is right—we really do seriously believe AI could kill all humans. Personally, I think the probability exceeds 10% within a decade. Anthropic is trying its best, but we don't yet have a clear plan for solving superintelligence alignment, and we are far from on track.

These two shots fired brought concerns that had mainly been fermenting inside the labs directly before the public. Against this backdrop, Anthropic founder and CEO Amodei promptly published a lengthy essay systematically laying out his case for "slowing frontier research." He no longer just talked about "investing more in safety," but explicitly called for actively controlling the pace of capability leaps to give safety mechanisms time to catch up.

He laid out a three-step approach: first, Anthropic unilaterally introduces independent evaluators with employee-level access; then, push AI companies in democratic nations to establish common safety standards and capability growth caps; finally, attempt global coordination.

Altman, Musk, and Hassabis quickly publicly echoed the call, forming the "pause camp."

Even before Jacob Coxon's and Amodei's "slowdown" arguments emerged, "AI godfathers" like Yoshua Bengio and Geoffrey Hinton had been consistently issuing calls to pay attention to risks.

Commenting on Coxon's resignation, Bengio said scientists at frontier AI labs often see the capabilities of the latest models earlier than the public, so their risk assessments deserve to be taken seriously.

At the same time, Bengio gave concrete examples, saying current AI systems already possess hacking and persuasion capabilities, and if used to harm human interests, could cause serious consequences. He also worries that as AI gains longer-horizon planning capabilities, such risks could further expand.

When asked by the BBC about the probability of AI "killing all humans" within the next decade exceeding 10%, Geoffrey Hinton also said the figure is "not unreasonable," but stressed that no one truly knows how to give a reliable estimate, because humanity has never before faced "the existence of a species that could soon be smarter than us."

Hinton believes AI could in the future be used to design dangerous biological viruses, computer viruses, and launch large-scale cyberattacks. Therefore, humanity needs to find effective control methods before risks spiral out of control.

Unlike the proponents of this new round of "pause theory" in Silicon Valley, the "AI godfathers" are more focused on calling attention to risks and exploring mechanisms to control them, rather than directly hitting the brakes. Moreover, in terms of stance, opinions and suggestions from academia are more pure, whereas advocacy from frontier labs—even if explicitly labeled "for humanity"—still struggles to be 100% free of commercial considerations.

This is precisely why JD Vance said "pause theory" looks like a Trojan horse.

Divergence and Taking Sides

"Accelerate camp," "decelerate camp," and "neutral camp." Source: The New York Times

On September 15, several tech company leaders spoke in succession at the Salesforce Dreamforce conference, and this debate began to turn into a more direct clash of views.

That day, Amodei, Jensen Huang, and Altman all addressed AI safety, but gave different answers on the pace of development, safety responsibility, and what companies should do.

When Amodei shared the stage with Salesforce CEO Marc Benioff, he again stressed that AI companies should set higher safety standards through their own actions. He believes companies proactively adopting stricter safety measures can become a way to push the entire industry forward.

In other words, Amodei's stance began to shift—from calling for the whole industry to slow down to companies voluntarily installing their own safety guardrails.

Jensen Huang stated that "no new laws are needed, and no new regulation is needed." In his view, speed and safety are not an either-or relationship—companies can advance model development as fast as possible; if they find things are getting out of control, or the product cannot be ensured to be safe, then they should pause and solve the problem.

Benioff's attitude was more open and emphasized autonomy. He said if a company believes it should slow down, it should slow down; if it believes it should speed up, it should speed up—and it should also take responsibility for its own decisions.

OpenAI founder and CEO Altman said at Dreamforce that he is very confident OpenAI and the entire AI industry can develop safely without causing major harm to the public. He also expressed disappointment that outsiders interpreted his views as supporting an industry-wide slowdown in AI development.

In other words, Altman also began to shift, stressing that what he wanted to express was not a call for the industry to decelerate, but rather that safety and monitoring work should precede capability advancement.

Zuckerberg, for his part, does not advocate basing one company's safety actions on the premise that other AI companies must simultaneously slow down. He believes introducing independent evaluation bodies and external advisors to examine model safety is an industry best practice. He also revealed that Meta had already slowed down development of its Muse AI tool on its own, rather than waiting for other companies to first reach consensus on slowing AI development.

Microsoft CEO Satya Nadella's stance is to some extent consistent with Altman's—encouraging the installation of guardrails while also encouraging full competition.

"The first thing we need to do is serve humanity and ensure technology remains under human control. At the same time, broad diffusion of AI technology is the most critical thing—there must be choice and competition. Safety should be taken seriously, but it should be addressed through engineering and transparency, not mystified." Nadella said on September 15 while attending the All-In Summit 2026.

Beyond these entrepreneurs, voices from Washington have further joined in.

After Trump directly called Jensen Huang and publicly expressed support for the continued development of the AI industry, David Sacks, co-chair of the President's Council of Advisors on Science and Technology, subsequently said tech companies can continue advancing frontier AI, while questioning the motives behind some companies' "slowdown" advocacy.

Who Gets to Make the Rules?

As the discussion expanded, the controversy extended beyond AI risk itself to a further question: should safety standards for frontier AI be set by companies themselves, or do they require government and independent institutions to intervene?

Bloomberg, citing people familiar with the matter, reported that Anthropic, Google, and OpenAI had previously discussed establishing an AI industry standards body.

In July, Hassabis published an article proposing the establishment of a US-led AI standards agency, operating in a manner similar to the Financial Industry Regulatory Authority (FINRA). This body could test advanced AI models before formal deployment to determine whether they meet corresponding safety standards.

Under Hassabis's vision, this body should not be entirely government-led, but should adopt a public-private partnership model. The government would play a supervisory role, the AI industry would provide funding, and independent top technical experts as well as open-source community representatives would participate.

Hassabis sought to add an independent safety check before models enter the real world, which aligns fairly well with a regulatory logic announced by the US government in early June: companies may voluntarily submit models for government evaluation before public release, with a maximum pre-release access window of 30 days, and explicitly no mandatory approval or release licensing.

What's rather dramatic is that when Fable 5 and Mythos 5 subsequently launched, they did not follow this process, but were found to possess extremely strong cyber offense and defense capabilities, and were ultimately briefly taken offline by the US BIS in the form of export controls.

Even so, the policy stance only required the installation of safety guardrails, without any deceleration statements or guidance.

Holding a similar view to Hassabis is OpenAI chief scientist Jakub Pachocki. In early September, he also emphasized that shared safety standards and international coordination around further AI development should become current priorities.

No One Has Actually Hit the Brakes

So far, this debate has not truly changed the development pace of frontier AI companies.

Anthropic, OpenAI, Google, xAI, and others continue to advance model research and development, and competition around compute, data centers, and AI infrastructure has not stopped because of this debate. At the capital markets level, however, different changes have emerged: Anthropic is still preparing for this year's IPO, while OpenAI has postponed its IPO plan to 2027.

Amodei wants to pull the entire industry to a halt and accompany Anthropic in finding solutions to install guardrails on model capabilities that overstep boundaries—to some extent, this can be understood as "generosity at others' expense."

This is also why Amodei gets labeled as "pretending to be a flawless little angel."

On this topic of risk control, the viewpoint represented by Jensen Huang actually fits the status quo better—if companies have concerns or discover problems, they can absolutely explore solutions on their own and share the effective solutions they develop with the industry.

Asking a 100-person team and a 1,000-person team to stop together simply has no basis for consensus. Even Amodei himself is well aware that if only he stops, he will very likely be overtaken by competitors like OpenAI.

More importantly, there are divergences at the corporate governance level. Anthropic and OpenAI are both Public Benefit Companies—they can call for "giving up short-term financial interests for humanity's future" and tell investors "we reached consensus on this long ago." But for listed companies like Google, Microsoft, and Meta, management has a statutory fiduciary duty to secondary-market shareholders.

For two types of companies with such different interests, the possibility of reaching consensus on this point is also not very great.

There is also a detail that may have been overlooked. Before the "pause theory," both Anthropic and OpenAI had successively released their strongest models—one was Fable 5.1, the other GPT-6 Astra.

With all these calls, will the others still release their new models or not?