Amodei's pacing plan buys outside scrutiny, not a slower frontier

*Buyers should price in independent eyes on the labs and ignore the brake until Washington signs the waiver.*

Opening read

Dario Amodei's plan to slow the most capable AI systems has three parts, and the useful way to read it is to sort them by who can actually deliver each one; that sort, not the headline, is what should change how you handle vendor safety claims this week.

Pacing the frontier: the badge is real, the brake is a request

Two events set this up. Last week researcher Jacob Coxon quit Anthropic and warned in a public post that the top labs are chasing AI that can improve itself, with the public's lives as the stake, and on Saturday his former chief executive, Dario Amodei, published a plan to slow the most capable systems so safety work can catch up. TechCrunch AI (https://techcrunch.com/2026/09/13/obama-urges-democrats-to-have-a-clear-plan-for-ai-safeguards/) Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing) Two days ago I argued that the wave of insider alarm was a political event more than a technical one. Amodei's plan confirms that and adds a twist: the politics produced a document faster than the technology produced a fix.

The consensus read is that the industry has finally agreed to slow down. A smart reader can hold that view because approval arrived fast: Sam Altman said OpenAI would follow Anthropic's commitment to embed third-party evaluators, Elon Musk posted his support, and Barack Obama told Democrats that AI safeguards need a clear plan at the center of their agenda. Amodei himself points to this summer's OpenAI Hugging Face hack, and Anthropic's own report that a Yemen-based engineering cell used Claude to help build guided weapons software gives the caution a second motive. TechCrunch AI (https://techcrunch.com/2026/09/13/obama-urges-democrats-to-have-a-clear-plan-for-ai-safeguards/) TechCrunch AI (https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/) Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing) The motive is real. The agreement is thinner than it looks: OpenAI and Anthropic already backed a petition in July asking Washington to support an international effort to pace the frontier, and two months later the industry is still asking for the mechanism. TechCrunch AI (https://techcrunch.com/2026/07/28/sam-altman-is-ready-to-decelerate/)

My read is that the plan delivers one thing and asks for two. Amodei laid out three strategies and committed Anthropic on its own to only one: outside evaluators who work inside the company on the same footing as its own risk reviewers, with credentials, a workspace, and hardware, and with carve-outs only where law or contracts intervene. The other two, common safety standards and limits on how fast unchecked progress runs, depend on governments granting a narrow antitrust waiver, a legal exception that lets competitors agree on shared rules without breaking competition law. No government has granted it, and Altman's matching pledge is a principle with details still pending. TechCrunch AI (https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/) Cohere Blog (https://cohere.com/blog/who-gets-to-define-the-rules-for-ai) Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing) So the deliverable is scrutiny, and the slowdown is a request.

Cohere's objection shows why that split matters. Its complaint is that a rule set written inside a few powerful labs in one country would become binding on every other developer once governments sign on, while Amodei's own line, offered in August, is that the rules Anthropic proposes are built to disadvantage frontier companies like his and favor smaller rivals. Cohere Blog (https://cohere.com/blog/who-gets-to-define-the-rules-for-ai) TechCrunch AI (https://techcrunch.com/2026/08/16/anthropic-ceo-says-ai-backlash-is-fundamentally-a-crisis-of-trust/) Cohere competes with those labs, so discount its motive, but the mechanism it describes is accurate: a waiver that lets the biggest players coordinate is also a waiver that lets them set the bar smaller rivals must clear. Amodei's claim cuts the other way. I do not know which one holds, whether this protects the public or the incumbents, and the evidence cannot show it while the waiver does not exist.

The security review is where this lands for a business leader choosing which model vendor carries the least surprise risk: the evaluator commitment is the useful part, and it changes what you demand. Most safety claims you are handed may rest on the vendor's own testing, and you can find out in an hour: pull your last AI vendor assessment, count the safety claims backed by a named outside party versus the vendor itself, and if the count is zero, ask Anthropic or OpenAI which evaluator gets the badge and whether its findings will be published. The answer tells you whether an embedded evaluator means an audit or a decoration. Engineers planning around release cadence should not stretch a roadmap on the promise of pacing. Nothing here changes shipping dates; it changes who gets to look.

Navier-Stokes: the credit fight got uglier than the math

OpenAI used an unreleased internal model to resolve the Navier-Stokes problem, one of the Millennium Prize Problems (famous unsolved math questions with a prize attached), and NYU professor Tristan Buckmaster accused the company of foul play over credit. The weekend added detail on both sides. OpenAI launched the run after hearing a false rumor that two Millennium Problems had fallen, and Buckmaster says Sebastien Bubeck offered either a joint posting the next day or a paper crediting OpenAI's model with Levent Alpoge dropped as an author. Bubeck and Noam Brown deny acting unethically, Bubeck has offered screenshots, and Altman says his team worked in good faith. Simon Willison's Weblog (https://simonwillison.net/2026/Sep/8/on-navier-stokes/) Don't Worry About the Vase (https://thezvi.substack.com/p/brand-new-ai-solves-a-millennium) Even the timeline has a wrinkle. The 88-hour solve time circulates as fact and shows up in Simon Willison's account, but one write-up traced the number to a satirical post rather than OpenAI's own statement. Simon Willison's Weblog (https://simonwillison.net/2026/Sep/8/on-navier-stokes/) Latent Space (https://www.latent.space/p/ainews-openai-reports-navier-stokes) The sharpest fight in the record is over process and credit rather than the mathematics, though at least one account still treats the result itself as disputed. AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/AI-Model-Month-Is-Off-to-a-Blistering-Start-e3ok1na) Either way, a research leader deciding whether to share early results with a lab should settle credit terms in writing before the first call.

Hard constraints: MIT checks the rule on the output, not the steps

MIT researchers built HardFlow, a method that makes a generative model (the kind of AI that produces new outputs such as images or robot motions) obey hard rules, meaning limits that must never be broken, such as a robot staying out of a marked zone, without retraining. HardFlow nudges the model toward what you asked for as it generates, then tests only the finished result against the rule rather than policing every step along the way. MIT News - Artificial Intelligence (https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914) The design choice is the news here, not the perfect compliance the team reports on its own robotics, maze, and image-editing tests. MIT's Navid Azizan pitches it as keeping the model's creative range while holding the nonnegotiable requirements fixed. MIT News - Artificial Intelligence (https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914) Checking the final answer instead of every intermediate step is cheaper and easier to wrap around a model you already run. Engineers weighing a compliance fine-tune against a bolted-on checker get an argument for the bolt-on here, with the caveat that three lab test beds are not a certification.

Augur calls

• Call: OpenAI will publicly name at least one outside evaluation organization with embedded, employee-like access to its AI systems by the end of 2026.
• Criteria: An OpenAI blog post, press release, or official statement naming the organization and describing on-site or employee-equivalent access proves hit; no such public naming by the due date proves miss.
• Due: 2026-12-31
• Odds: 55%
• Call: The US government will not grant frontier AI labs an antitrust waiver permitting coordinated pacing of AI development by mid-2027.
• Criteria: A federal statute, executive order, or DOJ or FTC business review letter authorizing such coordination among AI developers proves miss; absence of any such instrument by the due date proves hit.
• Due: 2027-06-30
• Odds: 80%

What to watch

• OpenAI still owes the details behind Altman's evaluator pledge, and the test is whether its access matches Anthropic's badge-and-desk standard. Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing)
• Cohere's objection is the one on the record from a rival developer; if other developers join, the waiver becomes a fight rather than a formality. Cohere Blog (https://cohere.com/blog/who-gets-to-define-the-rules-for-ai)
• Democrats now have Obama pushing AI safeguards onto their agenda, which is the venue where any waiver would be decided. TechCrunch AI (https://techcrunch.com/2026/09/13/obama-urges-democrats-to-have-a-clear-plan-for-ai-safeguards/)
• Sebastien Bubeck's promised screenshots would settle whose account of the credit talks holds up. Don't Worry About the Vase (https://thezvi.substack.com/p/brand-new-ai-solves-a-millennium)

Verification notes

Altman's evaluator commitment is a stated principle with details pending, and the antitrust waiver Amodei's plan depends on has been requested, not granted. Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing) Neither gap changes today's read, but both bound how far it can be pushed. The Navier-Stokes credit dispute rests on competing accounts, with screenshots offered but not shown in the evidence, one account treats the result itself as still disputed, and one write-up traces the 88-hour figure to a satirical post rather than OpenAI's own statement. Latent Space (https://www.latent.space/p/ainews-openai-reports-navier-stokes)

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-14?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-14)