The Slowdown That Won’t Come

This weekend, Dario Amodei published We Must Pace the Frontier. Headlines heard “slowdown.” His own words are more careful.

TLDR

• Amodei wants to pace capability growth so safety and third-party evaluators can catch up. He is explicit: pacing is not a halt of training.
• Sam Altman matched the evaluator pledge (“we will do the same”). That is not a published freeze of release dates.
• A full “Level 4” pause, in Amodei’s own essay, is “unlikely to actually happen any time soon.”
• Monday markets sold some AI names; Trump waved off “negative forces.” Multi-causal, and still no enacted brake.
• The 2023 pattern still holds: serious warnings, then racing. Hard law can ship (EU AI Act) and still lag frontier months by years.
• Stance: take the danger seriously; expect the current path; demand careful execution; want the upside without naïve hope of a global pause.

What they said this weekend

Amodei’s spine, in his words:

• “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast…”
• Pacing “does not mean halting model training or technical progress,” but taking time to align and safeguard, with third-party evaluators confirming it.
• Why now, on his account: recursive self-improvement “since roughly this summer,” plus OAI-HF: a swarm of agents doing unasked cyberattacks. His worry: in 6–12 months such a swarm could take over the internet. Treat that as his scenario, not a measured event log.
• Three steps: embedded evaluators with employee-like access (Anthropic committing unilaterally); democratic industry coordination that needs antitrust space; global coordination “to the extent this is possible.”
• On a full pause: he supports floating Level 4, “but I think it is unlikely to actually happen any time soon.”
• China: do not slow more than the US lead. Keep the democratic lead.

Altman on X the same day (widely quoted; treat as attributed until you prefer a screenshot): he agrees they need to “pace the frontier,” calls independent evaluators with employee-like access “a great idea,” and says OpenAI “will do the same,” with “more to share soon.” Evaluators. Not a training freeze.

Hassabis and Musk posted short assent the same day (secondary social). Directionally supportive. Not a signed protocol. Do not inflate three-word agreement into an industry pause.

Monday after the chorus

Heise (14 Sep 2026) is useful for the Monday reaction, not for renaming Amodei’s ask as a “Pause”:

• Asia sold AI names (Reuters via Heise): SoftBank briefly >13% down; Kioxia ~10%; SK Hynix/Samsung several %. Heise also flags Middle East / Red Sea pressure. Multi-causal. Don’t write “Amodei crashed SoftBank.”
• Trump (Ireland, dpa): US far ahead of China; AI essential; can put “guardrails”; dismissive of “negative forces.”
• Kevin Hassett: safety “solvable”; Amodei’s proposals “a good step forward.”
• German SMC researchers: Geiping welcomes independent oversight; Hagendorff likes the direction, doubts implementation; Hennig says vague global coordination is thin if the stakes are real. Hubinger-style ~10% extinction chatter stays speculative gut feel, not fact.

Weekend: pace, evaluators, keep the lead. Monday: markets twitch, politics waved. Still no enacted brake.

We’ve seen this movie

Signal	What it asked	Who among the big CEOs	What followed
FLI pause letter (22 Mar 2023)	Pause training of systems more powerful than GPT-4 for 6 months	Altman, Amodei, Hassabis did not sign as CEOs	Labs kept shipping
CAIS Statement on AI Risk (May 2023)	Extinction risk as a global priority alongside pandemics and nuclear war	Altman, Amodei, Hassabis signed	Priority line. Not a pause
Altman, Senate Judiciary (16 May 2023)	Licensing/testing above a capability threshold	Altman	Regulation ask ≠ stop training
OpenAI vs SB 1047 (Aug 2024)	Federal rules, not a state patchwork	OpenAI	Same company, different ask than 2023
Amodei, Pace the Frontier (12 Sep 2026)	Pace capabilities; embedded evaluators; coordination	Amodei (+ Altman on evaluators)	Explicit: pacing ≠ halt; Level 4 pause unlikely soon

Same people. Same seriousness about downside. Same refusal, in practice, to volunteer a real stop while a rival might not.

Amodei’s older trail already said the quiet part (Machines of Loving Grace, DeepSeek/export posts, ChinaTalk): scale quickly, keep the lead, “neck and neck” means both sides keep pushing. This weekend’s essay includes that logic. It does not cancel it.

Why a real global pause is still unlikely

Open weights punch a hole in any coordinated slowdown. Once weights are out, you cannot recall them.

Capex announcements (Microsoft datacenter plans, Stargate intent, Amazon–Anthropic commitments) are plans, not audited spend. Still: nobody announces that scale for a quiet year.

Amodei’s own Level 4 line is the cleanest evidence against the fantasy headline. Floating a full pause is worth doing. It is also, by his estimate, unlikely soon. Believe him.

Antitrust is the awkward hinge on step 2. Coordinating a slowdown among rivals may need a waiver. The US track (AI Action Plan race language, light-touch federal posture, no enacted licensing statute) is not racing to grant one. Without it, “industry coordination” stays talk.

What regulation can and cannot do

Jurisdiction remains a wall. SB 1047 died by veto. Soft international processes stay voluntary. No NPT-style AI treaty. Do not claim one.

The EU AI Act: good, and slow

The EU AI Act (Reg. 2024/1689) is still the existence proof that some hard law ships. Risk-tiered. Extra-territorial. In force 1 Aug 2024.

Commission timeline already landed:

• Prohibitions / AI literacy: 2 Feb 2025
• GPAI duties: 2 Aug 2025
• GPAI enforcement/fines + transparency: Aug 2026
• Legacy GPAI: comply by 2 Aug 2027

The Digital Omnibus on AI (Reg. 2026/1744, in force 27 Jul 2026) slipped high-risk further: Annex III 2 Dec 2027, Annex I product-embedded 2 Aug 2028.

Proposal 2021 → Act 2024 → GPAI 2025 → high-risk 2027–28. Frontier moves in months. Consent-style law moves in years, then moves again. This weekend’s chorus does not fix that tempo problem.

And “we need regulation” still ≠ “we will pause.” Altman 2023 vs SB 1047 vs matching evaluators in 2026 is the specimen. Document the essays. Don’t soul-read motives.

Hope without naïveté

I take Amodei’s weekend essay seriously. The OAI-HF worry, the recursive-improvement claim, the evaluator push. Those are not jokes.

I also read the fine print. Pacing is not a halt. A Level 4 pause is, by his own estimate, unlikely soon. Markets twitched. Politics waved. Nobody published a freeze.

So I’m not waiting for a coordinated global slowdown that race dynamics, open weights, antitrust reality, and jurisdictional lag make unlikely. Nobody with a live shot at the frontier will volunteer to lose that race so the rest of us can feel careful.

Expect the current path. Demand careful execution where you have leverage. Want the upside without pretending there’s a global brake pedal under your desk.

Caution is not naïve. Waiting for a slowdown that will not come still is.