Why
The report frames pacing as a coordination problem, and coordination problems fail for two different reasons: no one can verify what anyone else is actually doing, and even willing parties can't legally agree to slow down together. Stage one solves the first problem by putting the verifier inside the system with real access, borrowing directly from bank supervision. Stages two and three try to solve the second problem by widening the circle of parties bound to the same standard. The segment's antitrust point matters because it identifies a failure mode that has nothing to do with AI capability at all: a plan that depends on companies privately agreeing to hold back releases is, on its face, coordination of commercial activity, and in most jurisdictions that is illegal regardless of the intent behind it. A safety plan that can't survive contact with competition law isn't a safety plan yet.
How it works
Stage one: embedded evaluators, borrowed from bank supervision
The report's analogy is bank supervision, where independent examiners sit inside a bank full-time, watching risk and practicing the institution's own processes daily rather than reviewing a quarterly filing (01:15–01:26). Applied to AI, the proposal is independent external evaluators with employee-level access resident in major labs, auditing training and safety processes in real time rather than after a model ships (00:35, 01:33–01:40). The distinguishing feature is access: not a questionnaire the lab fills out, but a person inside the room who can see what is actually being trained and how.
Stage two: democratic coordination
Once evaluators exist inside individual labs, the next step is getting Western democratic labs to agree on common evaluation standards and release-pace norms, so no single lab is disadvantaged for slowing down while a peer does not (00:19, 01:44–02:47). This is a trust problem among parties that broadly share incentives and legal systems, which is why the report treats it as the middle difficulty tier rather than the hardest one.
Stage three: global coordination, adversaries included
The hardest and, per the report, still necessary tier extends the same coordination to non-democracies and geopolitical rivals, China named explicitly (00:19, 02:06–02:13). Nothing in the segment describes a mechanism for this stage beyond naming it as the goal — it is presented as the destination the first two stages are meant to build toward, not a plan with steps of its own yet.
The China question, and why "it might fail" isn't an argument
Asked what happens if China pulls ahead while the West paces itself, Amodei's reported answer is that later-stage coordination is indeed less likely to succeed the wider the circle gets — but pre-judging that failure and therefore not starting is not an option (02:13–02:26). The logic is that stage one (an evaluator inside your own lab) and stage two (agreement among like-minded parties) are worth doing on their own merits regardless of whether stage three ever succeeds; declining to build the earlier stages because the final stage is uncertain throws away the parts that are achievable now.
Antitrust: coordination needs a convener with legal cover
The segment's sharpest point is procedural, not technical: pacing releases is, by definition, coordination of commercial activity among competitors, and companies agreeing among themselves to do that is the textbook definition of collusion under antitrust law (02:54–03:28). The report's conclusion is that the government's first job — before any large piece of AI legislation — is to convene the players and legally sanction the coordination as safety cooperation rather than collusion (03:15–03:43). Without that legal cover, the labs most willing to cooperate on safety are the ones most exposed to a competition-law suit for doing so.
Where it lands in Jayverse
- Auditor: embedded evaluator is the design target. A rule-checker resident in the system with access to what's actually happening, logging who saw what and by which rule — not a report generated after a transaction clears. Stage one of the KBS report is the closest real-world analogue to what the Auditor row is trying to build.
- Verex: consortium governance needs a convener with legal cover before it needs a standard. If Verex or a resolution-source consortium ever needs multiple competing platforms to agree on shared rules, the antitrust point applies directly — the missing piece is usually not the technical standard but who has the legal authority to convene the agreement without it looking like collusion, the same gap Tech #103's 21-bank stablecoin consortium and #65's KB Kookmin + Kinexys work had to solve institutionally.
- Knowledge Notes: sits next to #103, #65, #69, #102. Tech #69's "an invariant is a stop" and #102's MLflow tracing / LLM-as-judge both rest on the same premise as bank supervision here — evaluation has to sit inside the loop with access, not arrive as a postmortem.
- Eng: two ready interview questions. "How would you regulate frontier AI?" answered in three stages of rising difficulty, and "why can't companies just agree to slow down?" answered with the antitrust mechanism — both concrete enough to rehearse out loud.
Verified and unverified
Verified on 2026-09-19: Dario Amodei has publicly argued for pacing AI development relative to safety guarantees in essays including "Machines of Loving Grace" (October 2024) and "The Adolescence of Technology" (January 2026), and has long supported export controls and a "race to the top" on safety among labs; Anthropic maintains a public Responsible Scaling Policy describing internal capability thresholds and safeguards. Bank supervision does embed examiners on-site at large banks (Federal Reserve and OCC resident-examiner programs are real and documented). US antitrust law treats coordination among competitors as presumptively suspect, and safe-harbor mechanisms for legally sanctioned industry cooperation exist, such as the National Cooperative Research and Production Act (NCRPA), which is the kind of vehicle the report's "government convenes and sanctions the coordination" argument would need. Taken from the KBS report and not independently verified by me: the name "Pacing the Frontier" as a formal document title, the specific claim that Altman, Musk, and named US government and congressional figures endorsed the proposal, and the exact wording attributed to Amodei. I have not seen Anthropic's internal position beyond its public documents — this is my note on a news segment, not a claim about what Anthropic has decided internally. Korea angle: KBS covering this signals it's entered the domestic policy conversation; a Korean team building agentic finance should expect embedded-evaluator-style requirements under Korea's AI Basic Act (effective January 2026), noted here as context.
Sources: [YouTube — KBS News, "「[이런뉴스] "그러다 중국이 앞서가면요?"…'AI 개발 늦추자'는 아모데이 답변은」"](https://www.youtube.com/watch?v=ctWJw8sQghk) · Dario Amodei, "Machines of Loving Grace" (Oct 2024) and "The Adolescence of Technology" (Jan 2026) · Anthropic's Responsible Scaling Policy · related items: bremmer-strategic-thinking-update-worldview (Life), nazarov-cftc-tokenization-us-60-percent, microfactory-small-models-clutch-rollback, rometty-resilience-relationships-attitude, Tech #103 (21-bank stablecoin consortium), Tech #65 (KB Kookmin + Kinexys), Tech #69 (an invariant is a stop), Tech #102 (MLflow tracing, LLM-as-judge).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| pace / pacing (v.) | 속도를 의도적으로 조절하다(늦추다) · 이 항목 전체의 핵심 동사. "deliberately pace frontier-model capability gains and releases" |
| frontier model | 프론티어 모델(현재 최고 성능의 AI 모델) · AI 안전 논의의 표준 용어. "pace frontier-model capability gains" |
| embedded evaluator | 내장 평가자(조직 안에 상주하는 독립 평가자) · 1단계 제안의 핵심 개념, Auditor 설계의 참조점. "embedded evaluators inside individual labs" |
| resident examiner | 상주 감독관(은행 안에 계속 머무는 감독관) · 은행 감독 비유의 출처, Fed·OCC 프로그램의 실제 명칭. "Federal Reserve and OCC resident-examiner programs" |
| employee-level access | 직원급 접근 권한(외부인이 아니라 직원처럼 내부를 볼 수 있는 권한) · 평가자의 실효성을 가르는 핵심 조건. "independent external evaluators with employee-level access" |
| in real time | 실시간으로(사후가 아니라 즉시) · 사후 보고서와 대비되는 핵심 구분. "auditing training and safety processes in real time" |
| convene / convener | 소집하다 / 소집자(당사자들을 모으는 주체) · 반독점 문제의 해법으로 제시된 정부의 역할. "the government's first job... is to convene the players" |
| legal cover | 법적 보호막(어떤 행위를 합법으로 만들어 주는 근거) · 협력이 담합으로 몰리지 않게 하는 장치. "the labs most willing to cooperate... are the ones most exposed... without that legal cover" |
| coordination of commercial activity | 상업 활동의 조정(경쟁자들이 사업 행위를 맞추는 것) · 반독점법이 주시하는 행위의 정의. "pacing releases is, by definition, coordination of commercial activity" |
| collusion | 담합(경쟁자 간 불법적 사전 합의) · antitrust의 반대편 결과, 이 항목의 핵심 위험. "the textbook definition of collusion under antitrust law" |
| antitrust | 반독점(법)(경쟁을 보호하는 법 분야) · 이 항목에서 안전 계획의 최대 장애물로 등장. "how companies can coordinate release pace at all without triggering antitrust law" |
| safe harbor | 세이프하버(특정 조건을 지키면 법적 책임을 면제받는 조항) · NCRPA 같은 협력 면책 메커니즘을 가리키는 말. "safe-harbor mechanisms for legally sanctioned industry cooperation exist" |
| NCRPA | National Cooperative Research and Production Act(협력적 연구·생산법, 경쟁자 간 협력에 법적 면책을 주는 미국 법) · 정부가 조정을 승인할 실제 수단의 예시. "such as the National Cooperative Research and Production Act (NCRPA)" |
| OCC | Office of the Comptroller of the Currency(미국 통화감독청, 은행을 감독하는 연방 기관) · 상주 감독관 프로그램을 실제로 운영하는 기관. "Federal Reserve and OCC resident-examiner programs are real" |
| race to the top | 상향 경쟁(안전·품질 기준을 서로 높이려는 경쟁, race to the bottom의 반대) · 아모데이가 오래 지지해 온 입장. "a 'race to the top' on safety among labs" |
| export controls | 수출 통제(전략 기술·물자의 해외 반출을 제한하는 정책) · 아모데이의 또 다른 오랜 지지 정책. "long supported export controls" |
| on its face | 겉보기에, 액면 그대로 판단하면 · 법적 판단의 출발점을 표시하는 관용구. "on its face, coordination of commercial activity" |
| worth doing on their own merits | 그 자체만으로도 할 가치가 있는 · 나머지 단계 실패 여부와 무관하다는 논리를 표현. "worth doing on their own merits regardless of whether stage three ever succeeds" |
| pre-judge (v.) | 미리 단정하다, 예단하다 · 아모데이의 중국 질문 답변의 핵심 동사. "pre-judging that failure and therefore not starting is not an option" |