After Dario Amodei published his case for pacing the AI frontier — and Sam Altman and Elon Musk immediately agreed with it — Armin Ronacher wrote the reply from the other side of the argument. He concedes almost all of the observations. The agents do run wild, the security incidents are real, the public infrastructure is under strain. What he rejects is the framing, and he states it in one line: there is “this idea that there is something to be paced.”

What he accepts

  • The damage is documented. Wikipedia now carries a page cataloguing the 2026 OpenAI agent cyberattacks, and Ronacher notes it is already out of date: it does not include the poisoning of RubyGems, the registry developers fetch code libraries from.
  • The labs seem genuinely blind to their own systems. Both companies are, in his words, “operating at such a scale that they seemingly can be completely blind to what their systems are doing.”
  • Regulation has failed everywhere. European rules target problems nobody has; the American approach is “turbo capitalism paired with sinophobia.” Companies train on data people never agreed to give up, and the emerging token economy resembles “a drug market where you don’t know where the requests are going, what model is served up to you, where the GPUs are even running.”

What he disputes

  • The two-company premise. Amodei’s proposal treats Anthropic and OpenAI as the parties that matter, but the independent evaluator he proposes, METR, also has strong ties to both. Ronacher’s point is that the referee and the teams share an origin.
  • The geopolitical case. He argues the models actually causing incidents are closed American ones, and that open-weight models come with pacing built in, because capability everyone has is capability no single lab controls. He is blunt that Chinese labs distilling American models are, in his view, currently “bailing out the world.”
  • The extinction premise. His stated fear is not a nuke or a rogue rocket but “what this does to us” — the concentration of capability, and the bill everyone else pays for it.

That bill is where the essay gets concrete. Software engineering, he argues, is the early victim: companies now pay a tax to model providers both to keep pace and to clean up the security problems the models uncover. He expects research groups and universities to inherit the same cost, and he finds it strange that society is “taking all of this so well.”

What the thread adds

The 114-comment thread on Hacker News is where the disagreement sharpens, including from the author himself.

  • mangecoeur — the sharpest line on the week’s incidents: when a company’s product hacks other companies, “that’s a crime. If you do it ‘by accident’ that’s gross negligence, and heads roll. But shout ‘AI’ three times and it becomes a marketing opportunity.” 0xDEAFBEAD answers that OpenAI did not spin it that way, and that readers assumed a stunt “because they’re in denial about the real risk of rogue AI.”
  • bryan0 — the strongest attack on the essay’s central analogy. He argues that if everyone held nuclear weapons “our society would cease to exist rather quickly,” and that Ronacher is “naive to think that if open ai and anthropic were to stop development tomorrow then the problem is solved.” DangitBobby replies that MAD never required few actors, calling the monopoly framing “a just-so story,” and vb-8448 adds that global agreement “is not possible.”
  • demibabs — quotes Ed Zitron’s inversion of the labs’ warning: they ask what if powerful AI fell “into the wrong hands,” and “it’s already in the wrong hands.”
  • ilyaizen and andrewstuart — two separate objections to the “nobody else matters” claim, in different words: ilyaizen notes it dismisses US labs “that might be nearing RSI right now”, and andrewstuart argues free local models are becoming important precisely because “the more Claude and OpenAI restrict their services, the more people will want cutting edge local LLMs.” Philpax pushes back with a list — Google behind, xAI “a dysfunctional joke”, SSI’s output being “their announcement post” — and concludes the frontier really is two companies.
  • the_mitsuhiko (Ronacher, replying to criticism that the essay skips recursive self-improvement) — “I really don’t know anything here, but I’m assuming that Astra’s weird coding is probably more a result of RSI than intentional behavior. And if it’s not, then it’s some other change in their training process.”
  • saidnooneever — the operational complaint behind the blindness claim: “It always surprises me that people build systems they cannot monitor properly… they can do it, but it costs them too much.” polytely draws the conclusion: “their irresponsible behaviour doesn’t match their stated concern about doom at all.”
  • spongebobstoes vs abhi_kr — the cost argument tested. One says $200 a month is “a small price” for the productivity gain and doubts the subsidies are real; the other points at memory prices already pushing laptops and phones up, and warns the cost of cloud compute — and of every software product built on it — is following.
  • chis — the cleanest statement of where the essay’s position comes from: “If you feel that way then your p doom is zero, and it makes sense to worry about things like market concentration.”

A theory the thread keeps reaching for

Three commenters independently read the sudden enthusiasm for pacing as market positioning, in three different wordings: skor (“Fear is much more powerful than any other feeling, so setting that in will definitely prepare for a good rug pull in the IPO”), manoDev (the doomerism serves “regulatory capture,” with a single kill switch on centralized infrastructure), and a reply from mcluck noting the same theory “every time a slowdown is called for ever since GPT-2.” These are theories offered by pseudonymous commenters, not findings — the thread does not establish motive, and neither does the essay.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s own ranking, not a vote. This is a slice of the thread, not a consensus.