Leadership1 publisher3 min readPublished
Amodei's pacing plan would put outside evaluators inside Anthropic with the right to publish
Three rival lab CEOs backed the pacing argument within two days of the essay. The checkable part of it is access for embedded evaluators, plus an antitrust waiver from Washington that has not been granted.
The Board Room · Leadership desk
What happened
- Dario Amodei used a weekend essay to warn that within 6 to 12 months an agent swarm could take over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage.
- Four of the largest labs landed on the same side of the argument within two days, with OpenAI saying it would match Anthropic's independent-evaluator commitment.
- Memory took it worst: SK Hynix and Kioxia each closed down 6.4% and the KOSPI fell roughly 3.2%, while ASML was off 4.4% at the European open and SoftBank dropped as much as 13%.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- exposure An evaluator with a badge, a company laptop and no editorial control over its own findings becomes a second publisher of a model's weaknesses. Enterprise buyers would get information they currently have to ask the vendor for.
- decision Joint capability limits sit behind a government decision nobody has made yet. Until the antitrust waiver exists, the only move a lab can execute is granting access on its own terms. The terms differ lab by lab.
- contradiction Bernstein and the newsletter that gathered the endorsements read the same document opposite ways, one as a modest deceleration and one as face-saving. Both readings hold, because the text does not ask for a compute cap.
- cost Suppliers carried the sentiment risk, not the labs: Korean memory and European equipment holders paid for the interpretation in a single session while Nvidia and Micron were roughly flat.
The only step in the plan Anthropic can take without anyone's permission is the first one [10]. Amodei is specific about what embedded third-party evaluators would get: in his words, "Desks in our offices, access badges, and company laptops," permissions "mostly comparable to what internal risk assessment teams have," and the right to publish findings "without editorial control by Anthropic" [4][5][6]. An outside party can check that. It also changes who sees a model before customers do.
Step two moves the decision to Washington. Common standards and capability limits across democratic-world labs would need the government to "issue a narrow waiver for certain kinds of safety conversations" because of antitrust, Amodei wrote [7]. Step three, coordination with authoritarian governments, he concedes is the hard one [8]. On China he proposes nothing new for US labs: keep chip export controls, crack down on distillation, harden weight security [9].
The endorsements arrived faster than the detail. Sam Altman said he agrees "we need to pace the frontier," called it "a primary topic of discussions we've had at OpenAI in recent weeks," and said OpenAI would match the independent-evaluator commitment [11]. Satya Nadella backed pacing on Sunday and said Microsoft would open its first-party MAI model behavior rules to public consultation [12]. Elon Musk replied "Dario is right" [13]. Three of the four labs attached a named procedural commitment to the endorsement [15].
A skeptic would call this theatre, and the newsletter that assembled the responses did: "Personally, I think it's a bullshit face-saving maneuver and none of these companies intend on slowing development," its author wrote [16]. Bernstein's Stacy Rasgon reads the text more narrowly, arguing the market reaction overshoots it, and his team put it this way: Amodei "is not calling for a halt to training, but rather a shift from 'extremely fast' to 'only somewhat fast'" [17]. Both readings survive the document. The essay asks for evaluators and export controls, and it does not ask for a compute cap on US labs [19].
I'd test intent against the capital calendar. The Financial Times reports Anthropic has told shareholders it expects a second straight quarter of positive adjusted operating income on gross margins above 80%, against $11.5B in Q2 revenue [22]. Business Insider reports the company has picked Nasdaq for an October listing, and Reuters separately reports talks for Nvidia to anchor with up to $10B [23][24]. Every one of those Anthropic figures is single-outlet reporting from unnamed sources with no company confirmation [26]. Bloomberg reported on Monday that SoftBank closed an upsized $11.87B two-year loan from roughly 20 banks to fund its OpenAI investment [25]. Altman told Fortune on Friday that safety makes right now "an ill-advised moment to go public," and confirmed: "I would say not 2026, yeah" [27].
Under the urgency is a single capability claim. Amodei wrote that "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI" [2]. In the same week, a shop called Specific published Real-SWE, which replaces public GitHub issues with licensed private codebases and sets tasks touching roughly 11 files each, against about 6 on comparable public benchmarks, or about 1.8 times the file count [28][30]. The best score is Fable 5.1 at 38.8%, with GPT-6 Astra at 33.8% and Gemini 3.8 Flash at 31.2% [29].
What to watch
- Whether Washington grants the narrow antitrust waiver Amodei asks for, and how narrowly it is drawn.
- Whether OpenAI publishes evaluator terms as specific as badges, laptops and publication without editorial control.
- Whether the October Nasdaq listing and the 80%-plus gross margin figure appear in a filing rather than unnamed sourcing.