Jensen Huang turns left, AMD turns right

Whether cutting-edge AI continues to accelerate or hits the brakes has become the most controversial topic in Silicon Valley right now.

Around this controversy, Silicon Valley giants have begun to take sides. Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis form the "Pause Camp". Jensen Huang is a firm "Accelerationist", criticizing that the "AI doomsday theory" lacks scientific basis; Yann LeCun has continued to mock Amodei, saying: "When (Amodei) was still at OpenAI, he thought GPT-2 was too dangerous to be open-sourced, and I laughed at him back then, and now everyone can continue to do so."

In contrast, Mark Zuckerberg, Satya Nadella and others have shown a relatively neutral stance.

The fact that the giants have taken their respective positions means that Silicon Valley has not reached a consensus on pausing cutting-edge model development, which stems from the differences in their ecological niches in the AI industry chain. Leaders at the model level can call for a pause for different reasons, but it is impossible to get a positive response if you want companies that hold hundreds of billions of dollars in computing power orders such as Oracle, AWS, and Nvidia, which stands at the top of the food chain, to stop.

In Washington, the calls for a "pause" have not only failed to gain support, but have even sparked a series of doubts.

"It feels a little bit to me like a bit of a Trojan horse," J.D. Vance said in an interview. This sentence actually reflects the real view in Washington: when companies that build Trojan horses keep asking to narrow the door, it is time to examine which companies this door is really suitable for.

01 Pause, Warning and a Resignation Letter

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

He previously worked on pre-training research at OpenAI and Anthropic for three years, and was personally involved in the training of the most cutting-edge models. In his long post on X, he was barely polite: "Neither company has acted responsibly. They are rushing straight towards self-improving superintelligence, betting our lives on it."

He said bluntly that people inside the labs genuinely believe that AI could kill all humans within ten years, and this is by no means a PR stunt. Many colleagues have begun to describe the current stage as "the critical moment" and "the endgame". Coxon even gave up the equity that was about to vest. He only stayed at Anthropic for four months, and according to the company's regulations, equity can only be exercised after six months.

Former Anthropic researcher Jacob Coxon

Shortly afterwards, Evan Hubinger, Anthropic's head of alignment science, publicly responded: Jacob is right, we truly believe that AI could kill all humans. I personally think the probability of that happening within ten years is over 10%. Anthropic is doing its best, but we do not yet have a clear plan to solve superintelligence alignment, and we are far from being on the right track.

These two statements directly brought the concerns that had previously been brewing mainly inside the labs to the public. Against this background, Dario Amodei, founder and CEO of Anthropic, soon published a long article systematically putting forward the proposition of "slowing down cutting-edge research". He no longer only talked about "investing more in security", but explicitly called for actively controlling the speed of capability leap to leave time for security mechanisms to catch up.

He listed three steps: first, Anthropic unilaterally introduces independent evaluators with employee-level permissions; then promote AI companies in democratic countries to establish common security standards and upper limits for capability growth; and finally try global coordination.

Sam Altman, Elon Musk and Demis Hassabis soon publicly echoed the proposition, forming the "Pause Camp".

Before the "deceleration" theory put forward by Jacob Coxon and Dario Amodei, "AI godfathers" such as Yoshua Bengio and Geoffrey Hinton had been issuing initiatives to pay attention to risks.

Talking about Coxon's resignation, Yoshua Bengio said that scientists in cutting-edge AI labs can often see the capabilities of the latest models earlier than the public, so their judgments on risks deserve to be taken seriously.

At the same time, Yoshua Bengio also gave specific examples, saying that current AI systems already have hacking and persuasion capabilities, and if used to harm human interests, they may cause serious consequences. He also worried that as AI acquires the ability to plan over longer time horizons, such risks may further expand.

When asked by the BBC whether the probability of AI "killing all humans" in the next ten years exceeds 10%, Geoffrey Hinton also said that this figure is "not unreasonable", but at the same time emphasized that no one really knows how to give a reliable estimate, because humanity has never before faced "the existence of a species that may soon be smarter than us".

Geoffrey Hinton believes that AI may be used in the future to design dangerous biological viruses, computer viruses and launch large-scale cyber attacks. Therefore, humans need to find effective control methods before risks get out of control.

Different from the new round of "pause theory" advocates in Silicon Valley, the "AI godfathers" are more calling for attention to risks and exploring the establishment of mechanisms to control risks, rather than directly hitting the brakes. Moreover, in terms of stance, the opinions and suggestions in the academic field are more pure, while the initiatives of cutting-edge labs, even if ostensibly labeled as "for humanity", are still difficult to completely separate from commercial considerations.

This is precisely why J.D. Vance said the "pause theory" is like a Trojan horse.

02 Divergence and Taking Sides

The "Acceleration Camp", "Deceleration Camp" and "Neutral Camp". Source: The New York Times

On September 15, a number of leaders of technology companies spoke one after another at the Salesforce Dreamforce conference, and this debate began to turn into a more direct confrontation of views.

That day, Dario Amodei, Jensen Huang and Sam Altman all talked about AI security, but gave different answers on the development speed, security responsibilities and what actions companies should take.

When Dario Amodei shared the stage with Marc Benioff, CEO of Salesforce, he emphasized again that AI companies should set higher security standards through their own actions. He believes that enterprises taking more stringent security measures on their own initiative can become a way to promote the development of the entire industry.

In other words, Dario Amodei's attitude began to shift, from calling for the whole industry to slow down, to enterprises voluntarily adding safety guardrails.

Jensen Huang said "no new laws are needed, nor new regulation". In his view, speed and security are not an either-or relationship. Enterprises can advance model development as quickly as possible; if they find that they are out of control, or the product cannot ensure safety, they should pause to solve the problem.

Marc Benioff's attitude is more open and emphasizes autonomy. He said that if companies think they should slow down, they should slow down; if they think they should speed up, they should speed up, and at the same time they should take responsibility for their own decisions.

Sam Altman, founder and CEO of OpenAI, said at the Dreamforce site that he is very confident that OpenAI and even the entire AI industry can develop safely without causing significant harm to the public. He also said he was disappointed that the outside world interpreted his views as supporting the entire industry to slow down the development of AI.

In other words, Sam Altman also began to shift, emphasizing that what he wanted to express was not to call for the industry to slow down, but that security and monitoring work should precede capability improvement.

Mark Zuckerberg does not advocate that a company's security actions be based on the premise that other AI companies must slow down synchronously. He believes that introducing independent evaluation agencies and external consultants to test model security is the industry's best practice. He also revealed that Meta had previously voluntarily slowed down the development of the Muse AI tool, rather than waiting for other companies to reach a consensus on slowing down AI development first.

Satya Nadella, CEO of Microsoft, is to some extent aligned with Sam Altman, encouraging the addition of guardrails and full competition.

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

In addition to these entrepreneurs, voices from Washington have also joined in.

After Donald Trump directly called Jensen Huang and publicly expressed his support for the continued development of the AI industry, David Sacks, co-chair of the President's Council of Advisors on Science and Technology, later said that technology companies can continue to advance cutting-edge AI, while questioning the motivation behind the "deceleration" proposition put forward by some enterprises.

03 Who Will Make the Rules?

As the discussion expands, the controversy has extended beyond AI risks themselves to whether the security standards for cutting-edge AI should be set by enterprises themselves, or require the intervention of governments and independent agencies?

Citing people familiar with the matter, Bloomberg reported that Anthropic, Google and OpenAI have previously discussed establishing an artificial intelligence industry standard organization.

In July, Demis Hassabis published an article proposing the establishment of a US-led AI standard organization, operating in a model similar to the Financial Industry Regulatory Authority (FINRA) of the United States. This organization can test advanced AI models before their official deployment to judge whether they meet the corresponding safety standards.

According to Demis Hassabis's vision, this organization should not be completely led by the government, but adopt a public-private partnership model. The government plays a supervisory role, the AI industry provides funding, and independent top technical experts as well as representatives of the open source community are also involved.

Demis Hassabis tried to add an independent safety check before the model enters the real world, which is more in line with a set of regulatory logic announced by the US government in early June: enterprises can voluntarily submit models to the government for evaluation before public release, provide up to 30 days of pre-release access, and explicitly do not engage in mandatory approval or set release licenses.

What is more dramatic is that the subsequent launches of Fable 5 and Mythos 5 did not follow this process, but were found to have extremely strong cyber attack and defense capabilities, and were eventually taken offline briefly by the US BIS in the form of export control.

However, even so, the policy attitude is only to require the addition of safety guardrails, and there is no statement or guidance to slow down.

Sharing a similar view with Demis Hassabis is Jakub Pachocki, Chief Scientist of OpenAI. He also emphasized in early September that sharing security standards and carrying out international coordination for the further development of AI should be a current priority.

04 No One Is Actually Hitting the Brakes

So far, this debate has not really changed the development pace of cutting-edge AI companies.

Anthropic, OpenAI, Google, xAI and other companies are still continuing to advance model R&D, and the competition around computing power, data centers and AI infrastructure has not stopped because of this debate. Different changes have taken place at the capital market level: Anthropic is still preparing for its IPO this year, while OpenAI has postponed its IPO plan to 2027.

Dario Amodei wants to pull the whole industry to stop and accompany Anthropic to find solutions to add guardrails when model capabilities cross boundaries, which can to some extent be understood as "being generous at others' expense".

This is why Dario Amodei has been labeled as "pretending to be a perfect little angel".

On the topic of risk control, the viewpoint represented by Jensen Huang is more in line with the current situation. If enterprises have concerns or find problems, they can fully explore solutions on their own, and share the proven effective solutions with the industry.

It is not possible to reach a consensus to make a team of 100 people and a team of 1000 people stop together. Even Dario Amodei himself is very clear that if he is the only one who stops, he is likely to be overtaken by competitors such as OpenAI.

More importantly, there are divergences at the corporate governance level. Both Anthropic and OpenAI are Public Benefit Companies, which can shout "abandon short-term financial interests for the future of humanity" and tell investors "we reached a consensus on this long time ago", but listed companies such as Google, Microsoft and Meta, their management has a fiduciary duty to secondary market shareholders.

Companies with two different types of interests are unlikely to reach a consensus on this point.

There is another detail that may be overlooked. Before the "pause theory", both Anthropic and OpenAI released their most powerful models, namely Fable 5.1 and GPT-6 Astra.

Under a series of calls, will other companies still release their new models?

This article is from the WeChat official account "Tencent Tech", author: Su Yang, editor: Xu Qingyang, published with authorization from 36Kr.