2026-10-15 (~1:27 PM ET — x.com/AnthropicAI): the thread's founding watch item gets its first checkable artifact — Anthropic: "We've committed to having evaluators assess our safety practices. We're in discussions with multiple organizations, and have now given employee-like access to a team from @METR_Evals," with a company post on working with third-party evaluators (anthropic.com/news). Development: the September 12 embedded-evaluator commitment now carries a named evaluator and granted access; the "discussions with multiple organizations" line keeps the breadth open. Watch: METR findings publication status (the essay's spec — publication without editorial control, Anthropic redactions), access scope, other named organizations, and OpenAI's parallel commitment materializing.
2026-09-22 (machinery strand): (1) OpenAI published
"Priorities and principles for effective third party
assessments" (openai.com, pub Sep 21 evening ET, in feed ~2 PM ET
Sep 22) — the first program document for Altman's Sep 12
"independent evaluators" commitment: deep access "across training,
evaluation, and deployment... as part of our efforts to pace the
frontier." Four priority areas: safety cases; critical safeguards;
capability evaluations covering Preparedness categories (CBRN,
cyber, AI self-improvement) + alignment evals; independent
investigation of critical misalignment incidents. Principles:
pre-registered mutually agreed claims, proportionate access,
conflict-of-interest disclosure, transparent methodology — and two
tells vs Amodei's essay ("cannot redact unfavorable findings"):
"actionable findings and time to remediate" before publication,
plus lab-requested redactions that assessors may merely note.
No named assessor: "in conversation with multiple third parties."
Machinery scorecard after two days: Anthropic has a named partner
(Accenture) and no embedded team; OpenAI has a document and no
partner. (2) METR pre-release-tested Opus 5.5 (Anthropic,
~12:30 PM ET; see anthropic-opus-5-5 thread) — METR's second
machinery-record appearance, now on the deployment side, while its
embedded-evaluator program remains unpopulated.
2026-09-18/19 (machinery lands + trigger corroborates): (1) the
essay's checkable artifact landed — Accenture named as Anthropic's
embedded evaluator, "at least $1 billion" joint investment
(anthropic.com + X, ~4-5 PM ET Sep 18). Held open: it's Accenture,
not METR, and "embedded" specifics (desks/badges/publication
without editorial control) are unspecified — ledger updated,
status pending until a team is embedded and first findings publish.
Altman's matching commitment still unnamed and undated. (2) RSI
trigger corroborated on the wire — Reuters/US News (~7 AM ET Sep
19): "leading labs say the scenario is near." (3) The states move:
Newsom's frontier-model kill-switch EO (California, ~1:04 PM ET
Sep 18 — the first state pacing-adjacent instrument on record) and
Virginia Gov. Spanberger's AI task force + data-center restraints;
the "cartel or safety pact" question gets its legal-architecture
layer via Decoder with Jonathan Kanter (antitrust exemption for AI
coordination). The Verge synthesis: "the AI regulation smackdown
isn't over." Federal-vs-state split is now the regulatory map.
(4) Frame-counter: Hollywood labor groups redirect the existential
frame to "what's already happening" (Verge).
2026-09-16 (Dreamforce strand, evening): both frontier CEOs made the
pro-pacing case on stage, separately, same venue as Huang's morning
"no new laws" — the debate's two columns shared one conference day.
Amodei (keynote exchange w/ Benioff): cited the Saturday essay,
automotive-safety model (competitors organize standards after a
peer's incident), "only 5% or 10% of what the possible value of the
technology is" in use. Altman (fireside): "smarter than people. That
finally happened"; on HF, "the world is right to be afraid of this"
(companies with "too much power"), then "The world should trust that
we are going to do the right thing"; publicly backed Amodei's call +
democratic-common safety standards (InformationWeek, Sep 16). Notable
frame-shift: InformationWeek treats Coxon's alarm as the originating
event and the HF breach as premise — trade press adopting the
warners' account on day 7 of the character attacks.
2026-09-15 (control strand): Kapoor/Narayanan, "What does it mean
to pace the frontier?" (13,000 words, pickup 6:53 PM ET Sep 14) —
the first substantive architecture answer to the essay: bridge
cybersecurity/AI-safety polarization; control over alignment as
the higher-yield investment; cyberoffense the one urgent specific
risk; organizational governance as the pacing tool (no dangerous
experiments without legal/security oversight; pause experiments if
that's what process requires). Notable: proposes governance that
doesn't depend on lab promises — evaluators (Anthropic) and safety
cases (OpenAI) both do.
2026-09-14 (pacing strand): Altman's thread (~12:05–12:18 AM ET) is
the handshake's operationalization. Defines pacing for the first time
— "we do not mean 'stopping'… it should be slower than it otherwise
could be" — answering the no-metric critique directly; "no amount of
American competitive pressure should justify recklessness." Names
OpenAI's first concrete mechanism: "explicit safety cases in advance
of frontier reinforcement learning runs we expect to significantly
increase capability" — with the admission that RSPs/Preparedness
Frameworks "focused primarily on the deployment of completed models,
not what happens during their development process" and "today's shift
to focusing on safe development and evaluation will need new tools."
Bypasses the essay's step-2 blocker: "we do not believe we need to
wait for an anti-trust exemption or legislation to begin the work"
("excited by ideas like independent auditors"). Asks government for
exactly one thing — "Where we will need the help of our government is
for international coordination. But first we should do what we can
ourselves" — the day after the WH/Speaker said no. Two failure modes:
loss of control ("unapologetically on Team Humanity") and
concentration of power (one person/company impressing their worldview;
one country; one lab). Watch: does a safety-case document ever become
public (the checkable version); does the auditor excitement become a
named program like Anthropic's embedded-evaluator commitment.
2026-09-12 (pacing strand): Amodei's "We Must Pace the Frontier" (essay + X, Sep 12 morning ET; Verge 12:23 PM ET). Unilateral step 1: embedded third-party evaluators — METR-class, employee-like access (desks, badges, laptops), publish findings without editorial control; Anthropic redacts security/legal/commercial material but cannot redact unfavorable findings. Step 2: democratic-country coordination on standards + rate limits (needs antitrust waivers). Step 3: global, tiered — bioweapons ban → mutual pre-release testing/global body → RSI speed limit (SALT analogy) → full pause, which he calls "unlikely to actually happen any time soon." Two triggers named: RSI ("starting to happen across the industry, including at Anthropic") and OAI-HF, with the dated warning: "in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet" (ledgered). New operational admission: Anthropic attributes its recent alignment incidents "in part by imperfect filtering of broken reinforcement learning environments," and says interpretability was used to examine unverbalized motivations in those incidents. China flank = enforcement (no chips/SM equipment, distillation crackdown, weight security), claimed to widen the US lead "significantly over the next 3–5 years."
Background
Same day: Altman agreed ("I agree with Dario that we need to pace the frontier… Committing to having independent evaluators") — the Roose "Sam and Dario agree" moment.
Who didn't answer: Google/DeepMind (no company statement; Legg/Dragan et al. signed the July employee statement, 1,386 signatures, endorsed by OpenAI + Anthropic; Hassabis has floated a FINRA-style body — the essay cites his mechanism). SpaceXAI/Musk absent from July's statement, silent today; the day's SpaceXAI move was Grok into Microsoft 365 Copilot. NVIDIA silent (standing: Huang "compute is revenue," $279B supply obligations, led the Jul 24 open-weights counter-letter). Chinese labs silent; active record = price war + alleged distillation; Beijing's "resolute countermeasures" warning stands. Trump–Xi late-Sept summit has AI on the table.
The skeptic frame on record: no definition of "pace," no metric, no referee — the LSI "speed limit without a speedometer" critique (Jul 30) applies verbatim to the essay; embedded evaluators are Anthropic's candidate speedometer, and they don't exist yet.
Watch for: METR team actually embedded (ledgered), OpenAI's "independent evaluators" commitment materializing into a named program, Google's company statement (its people signed July's; its CEO has the closest existing proposal), whether any pacing language survives into regulation (Frontier Act / Sanders bill adjacents), the Trump–Xi summit's AI language, the White House/Speaker posture after the Sep 13 rebuff (does the no harden into policy?), Musk's peer-review proposal gaining any structure (who calls, who gets early access), Anthropic's S-1/prospectus check (ledger), and whether Altman's agreement survives Astra's demand curve (Pro-tier purchasing pauses, Vending-Bench claims — the acceleration record sits directly under this handshake).