An "AI Doomsday" drama by five men
The debate that seemingly concerns the fate of humanity and technological ethics actually has nothing to do with either.
If something could truly end humanity within a few decades, it cannot be solved through debate. And if it can be resolved through debate, it was never an apocalypse in the first place.
Interest Considerations Behind the Anxieties
On September 14 local time, at the All-In Summit in Los Angeles, Jensen Huang was on stage participating in a live panel discussion when his phone rang. The caller was the President of the United States.
He switched the call to hands-free mode so that the entire audience could hear it.
Trump said on the other end of the line:
"Robots will not take over the world, and artificial intelligence will not control the entire planet. This kind of thing will never happen." He believes that recent claims about artificial intelligence getting out of control are "all a hoax".
He also added that data centers can create wealth for local communities, and stated that if the United States slows down its pace out of concerns over risks, the party that will be most delighted will be China.
Jensen Huang replied: "You are right, we will not let that thing (AI slowdown) happen." Applause immediately broke out at the scene.
The reason why Trump made such remarks is that the recent "AI doomsday theory" has been making a huge stir.
This round of debate was ignited by Dario Amodei, CEO of Anthropic. On September 12, he published a long article titled *We Must Slow Down the Frontier*, warning that "at the current rate, within 6 to 12 months, a group of autonomous AI agents may have the ability to take over the entire Internet through botnets, causing losses amounting to hundreds of billions of dollars."
He thus called for slowing down the upgrade speed of the most advanced models to allow safety research to catch up, and proposed a three-step plan: introduce embedded third-party evaluation, establish common industry safety standards, and promote coordination at the global level.
That same night, Sam Altman of OpenAI responded:
"I agree with Dario, we need to control the pace of frontier development."
Shortly after, Elon Musk of xAI reposted the content and added a note:
"Dario is right."
What is the relationship between these three people? Amodei and Altman are the most direct competitors; Musk and Altman have publicly fallen out for many years. Musk has also sued OpenAI, and it is commonplace for the two to confront each other on social media.
However at this moment, enemies and competitors have all stood on the same side, jointly calling for a slowdown in AI development.
They may indeed have concerns about the threats posed by AI, but judging from past events, such concerns are not many — as the "twin stars" in the current AI industry, Anthropic and OpenAI have always been emphasizing safety on one hand, while tearing down safety guardrails to accelerate their development on the other.
At the time of its establishment, OpenAI clearly set its mission statement as building general artificial intelligence that "benefits humanity safely and is not constrained by the need for financial returns". However, in the company's documents at the end of 2025, the key expressions of "safety" and "not constrained by profit-seeking needs" have been deleted.
In fact, even before that, OpenAI had already put profits above safety:
In January 2024, media reports stated that OpenAI quietly abolished the "ban on military and warfare use" of ChatGPT. As the US military's demand for AI technology grows increasingly strong, OpenAI has also quietly weakened its previous position of opposing military applications;
At the same time, OpenAI also abolished the long-held transparency principle, leaving external observers without the factual basis to judge OpenAI's governance structure, conflict of interests and safety decision-making mechanism;
In May 2024, 5 employees responsible for safety-related work resigned. Among them, Jan Leike warned the public that the company attaches too much importance to "dazzling products" and neglects AI safety.
Similarly, Anthropic also placed safety at its core position at the time of its founding.
Dario Amodei used to be the Vice President of Research at OpenAI. Out of doubts about OpenAI's values and concerns about safety, he left collectively with six other core OpenAI employees in 2021 to found Anthropic (which means "human"), and launched the star product Claude.
From the very beginning, Anthropic has taken the label of "the frontier company that takes 'AI safety' most seriously in the world", and released "Constitutional AI" to train AI — by creating a set of clear value principles, letting AI use this set of principles to judge and correct its own outputs.
In addition, Anthropic also launched the governance framework *Responsible Scaling Policy* in September 2023, to assess the potential catastrophic risks of models. If certain standards are met, the company will voluntarily suspend training until safety requirements are met.
However, in the 3.0 version of the *Responsible Scaling Policy* released by Anthropic in February this year, the previous promise of "not training more powerful models before risk mitigation measures are guaranteed to be in place in advance" was cancelled. It is replaced by a more moderate transparency mechanism centered on roadmaps, risk reports and external reviews.
It has to be admitted that the development of AI will bring a series of safety issues, but at present, AI has only taken the first step in a long march. Are the risks really large enough to reach the "doomsday" level that will destroy humanity?
This is a very abstract and quite doubtful statement. All such claims are "unfalsifiable and unverifiable". No technology company can provide credible specific scenarios to prove that such doomsday crises are imminent. They just keep emphasizing the crisis, never put forward decent solutions, and instead keep tearing down safety guardrails to run at full speed.
This makes people can't help but think that hidden behind these anxieties about the survival of humanity are considerations of interests.
A Play by Five Men
In June this year, Anthropic warned in its report *When Artificial Intelligence Starts Self-Constructing* that AI systems may achieve recursive self-improvement within two years, that is, autonomously design, train and improve successor models, and human involvement in model training will gradually decrease. Therefore, the report called on major global AI labs to slow down, or even suspend the R&D of frontier models.
However, just a few days before this report was released, Anthropic had secretly submitted its IPO application, with a valuation close to one trillion US dollars.
This time Dario Amodei called for a slowdown in AI development again, which also coincided with the critical period when Anthropic is about to go public.
As is known to all, the IPO window has relatively high requirements for the corporate narrative capability. Since its establishment, Anthropic has been focusing on "safety", which is also its core selling point that distinguishes it from competitors. By continuously strengthening the "safety narrative", it can further raise its valuation premium.
Sam Altman once sharply attacked this self-marketing method of Anthropic: "We made a bomb, are about to drop the bomb on you, and then sell you a 100 million dollar bomb shelter, which is only sold to people we select — this is obviously an incredible marketing stunt."
But despite the criticism, Altman also stood on the same side as Amodei this time, not out of concern for the fate of humanity, but because he also needs the "safety narrative" at the moment.
OpenAI had a net loss of 38.5 billion US dollars in 2025, while Anthropic's annual net loss was about 4.2 billion US dollars. In the first half of 2026, Anthropic's revenue growth rate far exceeded that of OpenAI, and for the first time surpassed the latter in single-quarter revenue scale in the second quarter, while turning to profitability. OpenAI not only continued to lose money, but its growth rate also slowed down.
Therefore, for OpenAI, now is not a good time to go public. The decision to postpone the IPO can be narrated in two ways. One statement is "we are delaying the listing for the safety of humanity", and the other is "we are waiting for a better valuation window". Altman obviously chose the former as his public statement, quickly switching from an "accelerationist" to a "decelerationist".
Therefore, when *Fortune* interviewed Altman recently, in response to the question "whether you have to run fast due to IPO pressure", his answer was crisp and clear:
"We will not rush to go public. Considering everything that is happening in the safety field, it would be an unwise moment to go public now, and we have no pressure on this."
Compared with Altman, Musk has always been emphasizing the dangers of AI. As early as 2014, he believed that "AI may be more dangerous than nuclear weapons", and also predicted that the probability of AI leading to the destruction of humanity is about 20%.
Later, he still kept warning about the dangers of AI, but at the same time founded xAI. His core logic is: Since AI is inevitable, rather than worrying outside the field, it is better to step in personally to ensure that AI develops in the direction he considers safe.
However, xAI is in the position of a follower in the frontier model competition. Calling for a "slowdown" is beneficial to the followers. And at the same time as he stated his approval of the slowdown, the pre-training of xAI's Grok 4.8 has just been completed, and reinforcement learning training has started immediately.
Meanwhile, at the All-In Summit on September 15, Musk talked about a plan to regulate AI: leading AI companies (including xAI) will test each other's models before the models are officially released, and competitors will be responsible for discovering and putting forward potential safety hazards.
This also reflects Musk's considerations: he wants to slow down the leaders through mutual supervision to benefit the followers, and wants to formulate industry rules as a participant. The industry self-regulation approach can also avoid government intervention in AI regulation.
The three people in the "deceleration theory" camp hold the same view, but each has their own calculations.
As the "shovel seller" in the AI industry, Jensen Huang is a complete "accelerationist". NVIDIA's performance guidance relies on the "exponential growth of AI infrastructure", while the deceleration narrative precisely threatens this narrative, so NVIDIA will have other answers on the deceleration issue.
Therefore, he strongly argues that the predictions of those with negative attitudes towards AI have not been fulfilled according to the schedule: radiologists will be replaced by AI in a few years, 90% of codes will be generated by AI in the short term, and a large number of junior positions will disappear rapidly.
Another heavyweight figure in the "acceleration theory" camp is Trump. In all his relevant statements, he repeatedly mentioned a judgment: the United States is currently "about one year ahead of China" in the AI field, but this advantage may be lost at any time.
He explicitly said: "I am not worried at all (about AI leading to human extinction)" "What I am worried about is that if we cannot win the artificial intelligence race, we will fall into a very disadvantageous position."
This kind of anxiety directly drives his actions. In May 2026, he was originally going to sign an executive order requiring frontier AI models to undergo a 90-day government safety review, but David Sacks, the former White House AI affairs director, made an emergency call on the day of signing, warning that the review mechanism would "seriously slow down the innovation speed of the US AI industry, and even cause the United States to lose the AI competition to China". Subsequently, technology executives including Musk and Zuckerberg also called to exert pressure, and Trump finally canceled the signing at the last minute.
In addition, Trump's strong opposition to the "AI deceleration theory" is also for the upcoming US midterm elections. The Trump administration has basically adopted a non-intervention attitude towards AI, but the American public is increasingly dissatisfied with AI construction, especially data centers, which are accused of driving up electricity prices.
A poll by NBC News in March this year showed that the approval rate of AI is even lower than that of the US Immigration and Customs Enforcement (ICE). A Gallup survey in May also found that seven out of ten Americans oppose the construction of AI data centers in their local areas.
For this reason, while Trump is emphasizing that the voices advocating that AI has fatal risks and calling for government intervention in regulation are all "absurd conspiracy theories", he also mentioned that "data centers are the oil of the next 20 to 25 years, their value far exceeds the Internet, and they are much more important than it".
He wants to characterize a livelihood issue that is unfavorable to his party as a "hoax" before the midterm elections.
All in all, the debate that seemingly concerns the fate of humanity and technological ethics actually has nothing to do with either.
If something could truly end humanity within a few decades, it cannot be solved through debate. And if it can be resolved through debate, it was never an apocalypse in the first place.
This article is from the WeChat official account "Shangyin She" (ID: shangyinshecj), the author is Haoran, and it is published by 36Kr with authorization.