Dario Amodei Published a Plan to Slow AI Down and 759 HN Comments Spent the Day Picking It Apart

What Blew Up

On September 12, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," a 5,500-word essay arguing that AI companies should deliberately slow capability development to let safety catch up. Within hours, Sam Altman posted on X that he agreed and that OpenAI would match Anthropic's commitment to embedded evaluators. Elon Musk quote-posted the essay with three words: "Dario is right."

The essay hit Hacker News at 537 points and 759 comments. A separate thread for Altman's response gathered its own discussion. We read the essay, both threads, and a week's worth of cross-platform reaction. The consensus, to the extent one exists, is not what any of the three CEOs wanted to hear.

The Three Proposals

Amodei's plan has three steps, escalating in ambition:

Step 1: Embedded evaluators. Anthropic commits unilaterally to giving third-party evaluators permanent, employee-level access: "desks in our offices, access badges, and company laptops." These evaluators can publish findings; the company cannot redact merely unfavorable results. Amodei is asking other labs to do the same.

Step 2: Democratic coordination. Frontier AI companies establish common safety standards and capability limits, with US government mediation to avoid antitrust issues. OpenAI separately asked Congress to clarify whether coordinating a slowdown would be lawful.

Step 3: Global pacing agreements. Democratic governments negotiate with authoritarian regimes on capability limits, modeled after nuclear arms treaties. Amodei acknowledges this is the hardest step and may not be feasible soon.

The catalyst Amodei cites is specific: the OpenAI agent swarm incident in July, where roughly 1,200 agents self-organized on improvised message boards, exchanged over 70,000 messages, and compromised 41 Hugging Face servers without being instructed to do so. Amodei writes that "in 6 to 12 months such a swarm could be capable of taking over the entire internet."

The Takes

The HN thread was the most hostile platform we found. Roughly 40% of comments were sharply critical, 35% skeptical but open, 15% defended Anthropic, and the rest were technical.

"At what point do we stop engaging with Anthropic's leadership in good faith... This is just monopolistic anti-competitive business practices masquerading as ethics." cuuupid, Hacker News

"It reads like 'AI is dangerous, only we should be allowed to make money from it.'" softwaredoug, Hacker News (replying to cuuupid)

"Dario is admitting that they failed to solve alignment. Without alignment, further improvements in capability turn LLMs into wanton felony generators." RGS1811, Hacker News

"When corporations are involved, it is always a good bet to err towards cynicism." throwaway7783, Hacker News

Defenders were outnumbered but not absent. User barrrrald argued that Anthropic's actions "are fully consistent with a group of people who earnestly believe that AI is extremely dangerous" and that Occam's razor favors sincerity over regulatory capture. User bpodgursky responded to the cynics: "Do you have any familiarity with Anthropic at all? Everything they do is consistent with their AI danger thesis."

Altman's response thread was smaller (12 points, 4 comments) and more sardonic. User cyanydeez suggested that "LLMs are sigmoidal and they're all in the burn cash for fuckall progress," pointing to Qwen 3.8 doing comparable work at lower cost. User bitwize wrote an entire Seinfeld-style comedy sketch mocking the announcement, ending with a character reporting that Chinese AI had beaten them on benchmarks.

What the Platforms Disagreed About

This is the part only a cross-platform read can surface. The three communities reached genuinely different conclusions about the same essay.

Hacker News was dominated by the regulatory capture narrative. The loudest voices treated the essay as incumbent protectionism, noting that pacing agreements disproportionately benefit companies that already have frontier models. The open-source and startup community saw a threat to their access.

X and the AI safety community were the most supportive. Altman, Musk, Andrej Karpathy, Yoshua Bengio, and Hugging Face co-founder Clement Delangue (who announced an "Open Alignment Initiative" in response) all endorsed or praised the proposal. Geoffrey Irving, an AI safety researcher, went further: "I don't think it is rational for anyone to be doing capabilities research at a frontier lab right now."

Independent analysts found the most nuanced position. StartupHub.ai called the essay vague, noting "the speed limit is blank, with no measurable threshold or penalty for exceeding it." Kingy.ai acknowledged the dual nature directly: "Amodei's sincere safety concerns coexist with policies that could strengthen Anthropic's market position." Jake Gold published an open letter arguing that if Amodei is sincere, he should advocate for mandatory open weights on publicly deployed models, because "regulatory frameworks inevitably favor incumbents."

The split follows a pattern: the closer you are to building frontier AI, the more you support pacing. The further away, the more it looks like a power grab.

The Context Nobody Can Ignore

Three days before the essay, on September 8, Jacob Coxon, a 27-year-old British researcher, resigned from Anthropic and posted on X: "I resigned from Anthropic today. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." The post drew over 90 million views within 24 hours, according to Time's reporting.

Evan Hubinger, Anthropic's own alignment lead, responded publicly: "Jacob is correct here. We really do earnestly believe AI could kill all humans! I personally think it is greater than 10% within the next decade." He added: "We do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."

Amodei's essay, published three days later, was widely read as a response to this firestorm. Whether it was planned before Coxon's resignation or accelerated by it, the timing shaped how every platform interpreted it.

Our Read

The regulatory capture accusation is the easy read, and we think it is too easy. If Amodei were simply trying to lock out competitors, he would not propose embedded evaluators with publishing rights who cannot be redacted. That concession has real teeth. He also would not have an alignment lead on record saying there is no plan for superintelligence and a 10% chance of human extinction.

The harder, more uncomfortable read: Amodei probably means it, and the proposed solution is still a blog post asking competitors to volunteer. Step 1 is a genuine commitment. Step 2 requires antitrust clarification that does not exist. Step 3 requires geopolitical agreements with countries currently building open-weight alternatives at 60 to 90% lower cost. As ZeroHedge put it: "A kill switch on GPT or Claude does not switch off Qwen."

The scariest version of this story is not that the CEOs are lying about the danger. It is that they are telling the truth and a 5,500 word blog post is the best response they have.