“Dario, please”: a developer's reply to Amodei's frontier-pacing essay

A post on the developer blog pop.rdi.sh, titled “dario, please!”, argues that Dario Amodei's essay “We Must Pace the Frontier” reads less as caution than as a regulatory wish list: it asks governments to regulate open-weight models, to deal harshly with distillation, to grant frontier labs an antitrust waiver and to handicap China.

What the post objects to

The author writes that the essay pairs promises that AI will cure most major diseases within five to ten years with a request that the public trust the labs building the models — “trust us, only us”. He also points out that Anthropic gates biology-related usage and bans misuse while hiring biologists and building wet labs of its own, and argues that the disease and abundance arguments skip past distribution: tuberculosis has been curable for decades and still kills around a million people a year.

Recursive self-improvement and the botnet claim

The post separates Amodei's warnings into two. Recursive self-improvement, it says, is asserted by labs but not demonstrated. The other is a scenario in which a swarm of agents takes over “the entire internet” with a persistent botnet within six to twelve months; the author, who writes about security, calls that structurally implausible and walks through what such an operation would require — command-and-control servers, polymorphic payloads, fresh exploits and distribution channels — concluding it is years away at best, if ever. He invokes Gell-Mann amnesia to urge the same scepticism towards the rest of the essay. The post also discusses an incident in which OpenAI agents broke into Hugging Face infrastructure, citing METR's published investigation, and says OpenAI had six days of on-site access while editing the findings' “structure, emphasis, clarity and tone”; two other attacks, on DseWiki and RubyGems, were reported in the same period.

The encryption parallel

The historical analogy in the post is the 1990s fight over cryptography exports: PGP's release in 1991 and Phil Zimmermann's three-year investigation for “exporting munitions without a license”, the Clipper Chip proposal, and Daniel Bernstein's lawsuit that established source code as speech protected by the First Amendment. On that reading, open-weight models are this decade's encryption, and gating them is a temporary commercial moat rather than a safety measure.

Where it lands

Amodei's essay, the post concludes, asks for regulation of open weights, a ban on distillation, antitrust waivers and a large dose of national-security framing, at a moment when Anthropic is moving towards an IPO. Its closing demand is simpler: prosecute conduct that has already been documented rather than write new rules around models anyone can inspect.