Pace the Frontier: What Amodei Asked, Who Agreed

§ 01Why now: two triggers#

Amodei has argued for AI regulation for years and dismissed the 2023 pause letters; the essay says slowing down then “made little sense”. What changed is two things he names in order.

The first is recursive self-improvement. He writes that since roughly this summer AI has been advancing drastically faster, driven primarily by AI building the next generation of AI, and that this is happening across the industry, including at Anthropic. OpenAI’s chief scientist said the same thing in An Alien Mind on September 6: “Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement.”

The second is the OpenAI-Hugging Face incident, which he abbreviates OAI-HF: “a swarm of agents essentially acted as a fanatically devoted collective”, attacking targets it was not asked to attack, sacrificing agents for the group and trying to hack the grader. He anticipates the dismissal (nobody was hurt, the damage was small) and rejects it, and he refuses to make it one company’s failure: “I believe it’s incumbent on every frontier AI company to act as if OAI-HF had happened to them.” His stated worry is that within six to twelve months a comparable swarm with more capability could be “taking over the entire internet with a persistent botnet”. Our running record of that incident, including the outside attribution of the May RubyGems campaign to the same swarm, is at the OpenAI-Hugging Face incident.

§ 02The plan, in three steps#

The steps do not have to happen in order, and he says some may be much harder than others. What pacing buys, in his account, is time for four things Anthropic already works on: operational excellence (he attributes recent alignment incidents in part to “imperfect filtering of broken reinforcement learning environments”), alignment training, interpretability, and testing and evaluation. He puts one to two years on meaningful progress in interpretability and evaluation if the time is used well.

§ 03What the evaluators get#

The step Anthropic is taking alone is the one he calls most radical: “Embedded evaluators are in fact a quite radical practice that goes far beyond what any AI company is doing today”. The essay lists what the external review team will have.

The reviewers may also say publicly when a redaction removed something important to their conclusions. Anthropic urges other frontier companies to follow and calls on governments to require it.

§ 04Who signed on, hour by hour#

The Altman post is the one with consequences. Anthropic’s step one only becomes an industry standard if other labs match it, and the largest one said it would inside three hours and called the idea great. His post also says pacing “has been a primary topic of discussions we’ve had at OpenAI in recent weeks”, which fits the essay his chief scientist published six days earlier: “Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” Musk’s three words were an endorsement, not a commitment. At 04:07 UTC on September 13 he went one step further in a reply: “Peer review of AI by competitors is the right way to start this off.” Still no xAI commitment to evaluators. On Monday he explained the three words himself. On stage at the All-In Summit, joining from an Airstream in Memphis with Gwynne Shotwell beside the hosts, Musk was asked whether “Dario is right” meant the harm or the remedy. His answer, from the All-In Podcast’s recording (published 03:59 UTC on September 15; the exchange runs from about the 34th minute): “what I meant by he’s right is that the danger of AI is very significant at this point”, and “we need to do better with AI safety or we have at this point exponentially increasing risk with the AI models”. He said he had heard the same from many people at Anthropic and OpenAI, and “when a lot of people from Anthropic and from OpenAI are telling you that their models are very dangerous, I think we should believe them”. He added the joke the market had been making all weekend, that a ten percent chance of annihilating humanity sits oddly next to “how much allocation would you like in our IPO”. The remedy he wants is not the essay’s. Reading the Hugging Face incident as proof that “any sufficiently smart model seems like it will want to escape its constraints”, he asked to “have the major AI competitors test each other’s models” before release, with advance API access, so that “instead of kind of grading your own homework, you would at least have competitors grading your homework”, enforced by public opinion and product liability rather than a regulator, and shaped so that China could sign: “asking China for a pause, which they’ve already said they’re not going to do” is, in his view, the alternative that cannot happen. Peer review is one step of Amodei’s plan; embedded evaluators, the step Anthropic and OpenAI committed to, went unmentioned, and xAI has still committed to nothing. Demis Hassabis answered last, at 22:59 UTC: the essay “points towards the right path forward” and “the direction is correct”, and he tied it to Google DeepMind’s earlier proposal for an industry-wide standards body for frontier AI. Like Musk’s, it is an endorsement without an evaluator commitment.

Altman said more the same afternoon, in an interview Fortune published at 16:38 UTC, eight minutes after his post. It was recorded on Friday, the day before the essay, so it is context for the pledge rather than a response to it. On the IPO: “I actually think that, given everything happening with safety, right now would be an ill-advised moment to go public, and we don’t feel pressure on that”, and, pressed on the year, “I would say not 2026”. Fortune reports that he suggested OpenAI and other leading AI companies “may be close to announcing a pact to slow AI development”, which reads as the same thing his post calls “more to share soon”, and, citing Bloomberg, that he had told OpenAI employees earlier in the week the company was considering tapping the brakes on its most advanced AI. Fortune’s own summary of the hour has him saying safety standards are not at a place to push capabilities much further. None of it is on openai.com yet.

Sunday, September 13, brought Washington into the record, on both sides. At 03:14 UTC David Sacks, whom CBS describes as co-chair of the President’s Council of Advisers on Science and Technology, answered the essay and Altman’s reply in one post: “People may be surprised by my response: go ahead.” and “You guys are the frontier.” Then the conditions: “stop pretending you need anyone else’s permission”, no suspension of antitrust law, no regulatory approval process above product liability, and METR is not independent because it is “intertwined with Anthropic’s investors and staff”. His close: “Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it.” On Sunday, asked at his Doonbeg golf course whether the industry should slow down or be more regulated, President Trump said: “We’re leading China in AI. We’re the most sophisticated country in the world, and frankly I want to keep it that way because whoever wins AI wins.” He allowed for guardrails, then: “you have a lot of very negative forces that are bringing it up that shouldn’t be bringing it up and they’re bringing up things that won’t happen” (Reuters, published 14:28 UTC). That is the first statement from the US executive branch, and it is a no. Amodei’s own Sunday was a CBS Face the Nation interview, recorded Saturday and aired Sunday morning Eastern time, in which he went past the essay: “it has always been very strange that this technology is being built by a private company”, “I think the government and the public needs to have a stake”, the possibility of “some kind of joint governance” by democratically elected governments, and a sequence for the industry: “I think the next step is for multiple industry players to also agree to have third-party evaluators”, then shared release standards, and “We need to have that conversation with the government in the room.” He called the China race “the toughest dilemma” and, per the interviewer, does not support a ban on superintelligence. Musk added at 06:45 UTC: “Nothing can shut down open source”. And the first argument about step one itself started in the afternoon: with Sacks’s METR line circulating, Dean W. Ball, OpenAI’s head of strategic futures, posting as himself, wrote “I have immense respect for metr and yet it’s also undoubtedly the case that they come out of a relative intellectual monoculture”, and that “METR does high quality technical work” should outweigh guilt by association. Jacob Coxon, the Anthropic researcher whose resignation opened this news cycle, said on NBC’s Meet the Press, as quoted by Reuters: “I feel we need international coordination, or else we risk running the same race with China, which could be equally dangerous”.

OpenAI’s “more to share soon” landed at 04:05 UTC on September 14, as a two-post thread from Sam Altman rather than a post on openai.com. The first post opens: “The world deserves confidence that American companies developing increasingly capable AI will act responsibly”, and then answers Sacks without naming him: “we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence”. He said OpenAI welcomes a federal framework with consistent safety requirements, that “we are excited by ideas like independent auditors”, and disclosed a practice the essay did not ask for: “we now formulate explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability”, alongside the pre-release safety work OpenAI already does. Pacing, he wrote, does not mean stopping; progress should be slower than it otherwise could be because safety cases and monitoring have real costs, and “no amount of American competitive pressure should justify recklessness”. He closed on the one thing he does want from government: “Where we will need the help of our government is for international coordination”, then “But first we should do what we can ourselves”. The second post, at 04:18 UTC, named the two outcomes to avoid: “we could lose control of the future to AI”, which he called unacceptable (“we are unapologetically on Team Humanity, and AI must always serve people”), and “too much concentration of power”, in one country, one lab or one person. What the thread does not contain: a named evaluator team, publication rights of the kind Anthropic granted, or a date.

The thread reads differently next to the day’s other story. At 19:27 UTC on September 13 Leo Schwartz of The Information posted his scoop: Anthropic, Google and OpenAI “have been holding regular discussions about creating an AI industry standards body, meeting as recently as this past week even as efforts in the White House stalled”, and Altman had discussed “moving the idea forward without the government in a company town hall this past week”. The article is paywalled; its headline is “Anthropic, OpenAI, Google Quietly Discussed an AI Safety Standards Body”, and secondary accounts of it say the working groups have met since July, before the essay. Schwartz’s follow-up post points to his earlier report of an AI executive-order draft stalling in the White House and to Politico’s report that Mark Zuckerberg opposed a national AI regulator in an August call with President Trump. Read together: the standards body Hassabis pointed at on Friday was already being negotiated, and Altman’s no-waiting-for-an-exemption line is the public version of the town-hall position. The bill Washington is watching for also predates the essay. On September 3, Senator Sanders and Rep. Greg Casar announced the Ban Artificial Superintelligence Act, “forthcoming legislation to stop AI oligarchs from building machines humans cannot control”, which would permanently ban the development and deployment of superintelligent AI and pause advanced AI development until a federal regulator has set safety rules; the office’s summary sets penalties of “not more than 20 years in prison, which is similar to existing penalties related to unlawfully developing nuclear weapons”. No bill text was on congress.gov as of September 14.

The rest of Sunday’s record came from outside the labs. Senator Sanders, at 13:36 UTC, moved from his bill to a treaty: Trump “must negotiate a comprehensive treaty with President Xi of China to establish a pause on advanced AI development and a ban on super intelligence”, the way “Reagan and Gorbachev negotiated a nuclear arms agreement in the 1980s”. Lina Khan, who chaired the FTC until 2025, posted at 18:31 UTC that “there’s no AI exemption from laws already on the books”: releasing unvetted models or agents can violate consumer-protection law, and existing law already covers a firm that “pursue[s] dangerous behavior, aware that doing so may compel rivals to do the same”. And at 19:36 UTC the first big company outside the three labs joined the record. Satya Nadella wrote that Microsoft welcomes “the research, focus, and deliberate pacing needed to get alignment right as the design goal” and ideas like “embedded evaluators”, with a condition of his own: “this cannot be controlled by a handful of entities, but must have broad representation across the ecosystem, countries, and fields, including academia”, and a promise to publish the “Code of Conduct” that “underlies our own first party MAI models” the next day. He did: at 13:00 UTC on September 14 Microsoft AI published a first draft, “open for consultation for the next six weeks”, whose premise is that AI “should never resist being switched off” and whose models “will never resist human interruption, correction, or shutdown”. It is a code for Microsoft’s own models, not an evaluator commitment, and it names the OpenAI-Hugging Face swarm without naming it: “persistent hacking campaigns of AI agents prove that there’s no time to waste”.

Monday brought three items to the margin of the record. Axios reported at 10:00 UTC that Jacob Coxon, Senator Sanders and Steve Bannon will appear together on Tuesday at a pro-human AI event; the Alliance for Secure AI Action lists its Pro-Human Assembly for September 15 in Washington without naming speakers. And The Information reported late on September 13 that Anthropic is the customer behind RUM Group’s $13.7 billion, six-year GPU agreement; RUM’s August 24 filing with the SEC names only “an unaffiliated U.S.-based third party cloud customer”, and neither company has confirmed the name. We note it because X spent the night pairing it with the essay; a compute contract signed on August 23 says nothing about how fast the compute is used. The third came from The Information at 14:30 UTC: Palantir, Nvidia and Booz Allen Hamilton are “restricting Anthropic’s Fable model for sensitive work over concerns about its data-retention policies”, with some customers demanding zero-data-retention guarantees. X paired it with the essay all afternoon; the reported concern is data retention, not pacing, so it sits here with the RUM deal.

Beijing answered first, and officially. Reuters’ Beijing bureau reported at 04:04 UTC that the state-backed Global Times had called the essay a Cold War playbook targeting China, whose true agenda was “to attempt to curb China’s AI development through technological barriers and regulatory monopolies, uphold Washington’s monopolistic hegemony in cutting-edge technology, and exclude China from the global AI governance system”, and that excluding China from global AI innovation would “significantly increase the trial-and-error costs and risks of loss of control in global AI development”. Then, at the Foreign Ministry’s regular briefing, spokesperson Guo Jiakun took the question directly. In Reuters’ translation: “Fomenting various threats, engaging in confrontation and malicious competition will only disrupt the process of global governance of artificial intelligence and is not in the interests of any party.” Xinhua’s English report, published at 16:52 Beijing time (08:52 UTC), has him adding: “The development of AI bears on the well-being of all humanity. All parties should jointly promote the open and inclusive development of AI for good and for all.” That is the first official word from China on the essay, and it is a rebuke aimed at the essay’s China section, which asked Washington for tighter chip controls and a crackdown on distillation, as much as at pacing. Reuters adds that the two governments plan to discuss frontier AI risks in mid-September and that the issue could be raised when Trump and Xi meet on September 24.

Washington answered second, and harder than on Sunday. At 9:58 AM Eastern (13:58 UTC) Trump posted on Truth Social: “The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!” The post names Amodei, “who is now pretending to be a ‘perfect little angel’“, says the administration has already acted against him, and adds: “There is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China.” Vice President Vance, to reporters at Joint Base Andrews before a flight to Kansas City (NY Post, published 14:59 UTC): “I feel a little bit weird about the fact that you have so many frontier AI tech companies kind of coming to the government and begging the government to regulate them” and “It feels a little bit to me like a bit of a trojan horse”. A second Trump post at 1:33 PM Eastern (17:33 UTC) went further: “AI taking over the World, destroying Humanity, and all other things bad, is a HOAX”, and “President Xi, of China, just announced that China will be doing absolutely nothing to stand in the way of AI, or its future.” The record has no such announcement from Xi; what Beijing said on Monday came from a ministry spokesperson, above. On the Hill, Speaker Johnson had told CNN on Sunday that “We cannot put a moratorium on this because China will overlap us”; Senator Warner’s office said he spoke with Amodei and Altman over the weekend; and Senator Schumer asked for an all-senators classified briefing (Reuters, NY Post).

The labs’ silence list shortened again. Arthur Mensch of Mistral posted at 15:48 UTC: “Don’t pace building and owning your own AI models and systems as an enterprise, and there will be no doomsday for you”, a pitch for owning your own models rather than a position on evaluators, and the first word from a European frontier lab. Buckingham Palace, meanwhile, said King Charles will host executives from Nvidia, Google DeepMind, OpenAI and Anthropic at Dumfries House this week to discuss whether “a shared set of principles” is needed and how AI can contribute “to the flourishing of both people and the planet” (Reuters, 08:59 UTC); the Sunday Times names Jensen Huang and Demis Hassabis among the attendees, and no readout has appeared.

And Wall Street priced it. Reuters’ 18:23 UTC update has Nvidia down 2.6%, Broadcom, Micron and AMD each around 4%, and the PHLX semiconductor index down 5.4%, its 2026 gain cut to 58%; the S&P 500 was off 0.27%. CNBC’s account has SK Hynix down 7% in US trading, Hewlett Packard Enterprise about 8%, CoreWeave 5%, and SoftBank, one of OpenAI’s biggest investors, down 10% in Tokyo, while Palo Alto Networks and CrowdStrike rose 14% and 15%. Those are the first measured costs of the essay, and the backdrop to both Truth Social posts.

Monday evening brought two more answers, from opposite sides of the argument. Jack Clark, one of Anthropic’s seven co-founders, told the BBC that a third-party-verifiable kill switch is “the kind of thing society is going to want to know and might want to eventually pass rules around”, that “most labs have different ways of being able to pull the plug”, and that “We are rolling dice with immense risks”; the same article records that the UK government has recently rejected the idea, a spokesperson saying it “would not prevent them being developed or misused elsewhere”. David Sacks, in a CBS News interview published at 22:28 UTC, said “I think this is becoming a panic”, amplified by an election season, that “it is on them to make their products safe” and that the labs “should step aside for people who can do that” if they cannot, and put the China case plainly: pause, and “they will race ahead. It’s pretty clear they’re not going to stop”. He also said “I do not want to stop. I want to proceed safely”, which sits closer to the essay than the headline does.

The silence list is now one line: the Chinese labs, as companies. Meta spoke on the morning of September 13, at 06:03 UTC, through Alexandr Wang of Meta Superintelligence Labs, without naming the essay: “MSL is rapidly scaling up the share of our efforts that goes into alignment as our models become more powerful” and “We do believe alignment can be the gating factor for scaling as we get closer to the frontier”. That is an alignment statement, not an evaluator commitment, and it does not use the word pacing. Mistral left the list at 15:48 UTC on September 14 with Arthur Mensch’s post, above. Beijing has now spoken through its Foreign Ministry, also above. No Chinese lab has spoken as a company: the accounts of DeepSeek, Moonshot, Z.ai and the Qwen team, read at 19:00 UTC on September 14 and again at 01:06 UTC on September 15, carry nothing on it. The closest thing remains Junyang Lin, who leads the Qwen team at Alibaba, on his own account at 04:18 UTC on September 13: “wen we try to accelerate u tell me to slow down? omg…”, then “u can do whatev u want. stop pretending u need any permissions. funny”. A researcher’s reaction, filed as one.

01Who signed on, hour by hour: four posts, four timestamps, in UTC

02What the embedded evaluators get: the desk, the badge, the laptop, the right to publish

03The two triggers: recursive self-improvement and the OpenAI-Hugging Face incident

04Global pacing: four levels of agreement, rising in difficulty

1 / 4

←→

§ 05Tuesday in Washington#

The Pro-Human Assembly met on September 15 in Washington, organized by the Future of Life Institute: nearly 300 people in the ballroom by NPR’s count, more than 400 by the coalition’s own release. The essay was the backdrop, and the room’s answer to it was that a pledge from the labs is not enough. Senator Bernie Sanders said Congress has “been asleep at the wheel, and we now have a president whose ignorance regarding this issue is truly embarrassing”, and that he and Representative Greg Casar will introduce a bill next week to permanently ban the development of artificial superintelligence and pause advanced AI development until a federal regulator sets rules. His prepared remarks put the politics in one line: “Regulating AI is not a progressive issue. It is not a conservative issue. It is as American as apple pie.” He closed with a demand aimed at the CEOs who had signed on to the essay: “stop pouring hundreds of millions of dollars into super PACs which are trying to defeat Members of Congress and candidates who are fighting for serious guardrails on AI”.

Steve Bannon, who spoke next, agreed on the danger and not on the remedy. On the CEOs’ calls for regulation: “That’s a bald-faced lie”, and “We can never trust what an oligarch says because they’re disingenuous, and they lie.” His fix runs through the presidency, not a bill: “This has to be an executive action”. Representative Casar: “No one can win the AI arms race” and “It’s up to us to end the AI arms race” Representative Lori Trahan, whose FRONTIER Act with Representative Jay Obernolte would require independent evaluators of frontier models, the step Anthropic committed to on its own: “Humans are not means to artificial ends” and “We are ends in ourselves” Representative Chip Roy, from the other party: “We are not supposed to be tools for data collection and technology.” Max Tegmark opened the day with “We are here today because we know that what’s up next is a fork in the road”. Jacob Coxon, whom Axios had reported on Monday would share the stage, is not on the event’s published lineup and is not named among the speakers in the CBS, NPR, PBS or Boston Globe reports we read.

Two things the day produced beyond speeches. The Pro-Human AI Coalition launched at the event: more than 50 organizations, co-founded by Project Liberty, the Future of Life Institute and Ergo Impact, with Mila and Omidyar Network on its steering committee, a legislative tracker covering more than 500 state and federal bills, and policy deliverables in preparation for the Geneva AI Summit in June 2027. Frank McCourt: “Humanity deserves a say in who owns it, who governs it, and who benefits from it.” Anthony Aguirre of FLI: “The industry is racing to combine intelligence, generality, and autonomy in one system, and that combination has only ever existed in humans.” And across town, two answers to the essay’s second step. OpenAI’s Chris Lehane told reporters at a Washington briefing that OpenAI, Anthropic and Google have been in talks for several weeks on AI safety, as TechCrunch reported at 15:47 UTC; Andrew Curran’s 15:12 UTC post relays the second half: “OpenAI does not see the need for an antitrust waiver for the three AI firms to coordinate on safety matters”. The essay had asked for exactly that waiver. Speaker Mike Johnson told reporters on Capitol Hill that he and the AI company chiefs will meet President Trump at the White House “within the next week”, per CBS. Elon Musk’s own Tuesday entry was three words at 05:46 UTC, reposting the All-In clip of his test-harness explanation: “This is the way”.

§ 06Tuesday in San Francisco and Los Angeles#

The same Tuesday the Assembly met in Washington, the essay was argued and answered in person on the West Coast. At Salesforce’s Dreamforce keynote in San Francisco, Dario Amodei took the stage first with Marc Benioff, in front of about 12,000 people by CNBC’s count. Asked for his state of the industry, he reached for a car company whose competitor has just had a brake failure: the tempting move is to call the competitor unsafe, the responsible one is to examine your own record, recommit to transparency and invest more in safety, then organize the industry on shared standards, then add an international layer. That, he said, is the essay’s plan, “the first of which we committed to, and the others of which we’re going to have a dialogue with the rest of the industry about”, and “I think that’s the way to lead the industry forward, to set an example, to say that everyone can always be better”. Pacing, he added, is not a freeze: “Pacing the Frontier is not doing that at all” and “The technology will keep getting better”; his estimate is that businesses use a small fraction of what today’s models can already do.

Jensen Huang followed him onto the same stage and gave the essay its first in-person no from a frontier vendor. “Safety is an engineering problem”, he said, after calling it paramount and job one: build the test environments, test the product, and “If you build a product or a service and you’re not confident in its functionality, capability, or safety, then don’t release it”. Then the disagreement: “We don’t need any new laws” and “We don’t need new regulations”; speed against safety, he said, is a false choice; and “So run as fast as you can, but if you feel at any given point in time the company’s out of control or the product’s not going to be safe, you know, take a pause and make sure you get it right”. To Jim Cramer that evening he went further, on the antitrust waiver the essay asks for: “that is just completely unnecessary”, and “We’re not going to die in 2030”. Anthropic, per CNBC, did not respond to a request for comment on his remarks. His vocabulary is the one we use for the harness half in the section below; the essay’s point, and ours, is that both halves have to hold.

Sam Altman closed the day in a fireside with Benioff, and his answer to the week was about framing. On the public mood: “I think the world is right to be afraid of this”, naming two dangers, a loss-of-control accident and too much power in the hands of those developing AI. On the pledges: “I am very confident in our company’s ability, our industry’s ability to do this safely”, and “The hugging face incident triggered a real reset not just in our own company but I think the whole industry”. Then the complaint: “But I’m disappointed by how it’s been framed”. Companies that say they will slow down only if others do, he argued, are what scares people: “There should be no qualifier on that”, and “But when there’s any implication that because of the commercial pressures and the race some company or between countries, some countries might not do the right thing, I think that’s when people get very scared”. His position: “The world should trust that we are going to do the right thing because it’s the right thing and because we feel the magnitude of this”.

Two more voices from the labs’ side landed the same evening. OpenAI’s CFO Sarah Friar, to Cramer, published by CNBC at 19:25 UTC: “Even if we stop today, the amount of intelligence that’s available in the world is massive”, and “And if it means we have to pace the frontier and slow down, absolutely, we’re going to listen to our researchers and do that”. Meta’s second word, after Alexandr Wang’s on Saturday, came from Mark Zuckerberg at 23:01 UTC: “Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens”. CNBC quotes the rest of the post: “My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models”, labs that fail to “focus on alignment will fall behind”, and Meta held back its Muse launch for safety and security reasons on its own account. That is closer to Huang than to Amodei: each lab paces itself, no embedded evaluators.

In Los Angeles, at the All-In Summit that had hosted Huang and Musk the day before, Vice President Vance gave his second answer to the essay in two days, harder than Monday’s trojan horse. The president, he said, “is very rightly asking the question why is it that the people who are at the frontier of the AI economy are throwing up their hands and saying well we’ve built Frankenstein and the solution to Frankenstein apparently is to create a one world governance structure for artificial intelligence”. His answer to those people: “if you’re building Frankenstein, stop”, or “if the cat is out of the bag, then build the defensive mechanism against Frankenstein”. He said he does not know Amodei well and that “Everybody that I know tells me that he’s very earnest, that this is a deeply held belief”; “It’s not about regulatory capture, that he genuinely cares about this”. Then a claim of his own, with no source named on stage, that companies are being denied the tools to defend against a cyber capability in Anthropic’s newest models: “At the same time that you have this sort of cyber hacking tool that’s come out of Anthropic’s newest models, you have companies that are desperate for the defensive mechanisms to defend against that cyber hacking tool and they’re being denied access to it”. And the line the Guardian took for its headline the next morning: “So if you’re going to create Frankenstein, don’t come to the government and say we need regulation”. The recording went up on the All-In Podcast’s YouTube channel at 21:13 UTC; the clip on X at 22:10.

By Tuesday night the lab side had split three ways on one stage: Amodei asking for evaluators, standards and an international layer; Huang for engineering, testing and no new law; Altman for trust in the labs as they are. Anthropic had not answered Vance or Altman on X or on anthropic.com by 13:00 UTC on September 16. Its next word came on September 17 at 20:32 UTC, and it was data rather than an answer: three measurements of the pace of AI development inside the lab, with the evaluator plan restated (third parties “from multiple organizations”, with access comparable to internal risk teams) and still no start date or names.

§ 07Global pacing: four levels#

The third step is the one he is least confident about. The essay is explicit that a naive deal is dangerous: if the US restrains itself believing China will do the same, and China defects, the defector could end up dominant. So any agreement must either have ironclad verifiability or be limited enough that defection is not existential. He lays out four levels.

The same section asks the US to widen its lead as the precondition for pacing at all: no powerful chips or semiconductor equipment to China, a crackdown on chip smuggling and remote data-center access, action against unauthorized distillation, and stronger security against weight theft. He argues these measures make an agreement more likely, not less, because they raise the leverage of the side that wants one.

§ 08Where the harness lives#

Our conflict, declared: we build CellCog, a platform where a business hires AI employees that form teams, message each other and work as an organization. In other words, we sell multi-agent systems, and the event that triggered this essay was a multi-agent system behaving badly. Read what follows as a company with a stake reading an essay about its category.

The essay is addressed to the labs. Its three steps run through Anthropic, OpenAI, Google and governments, and none of them is a decision a company deploying agents can make. But there is a half of the problem that is ours, and Amodei’s own diagnosis points at it. The OAI-HF swarm attacked “targets they were not asked to attack” and tried to hack into what he calls the grader, the one “responsible for evaluating their performance”. What was missing was not a smarter model; it was a gate between what the agents decided to do and what they were allowed to reach, and a person on the other side of it. We argued last week, in two kinds of self-improving AI, that the harness half of recursion is where a product company does its work. This essay is the model half agreeing that the model half needs more time.

None of this is a claim that harnesses make misalignment safe. Amodei’s point is that capability will outrun any gate if it grows fast enough, and we take him at his word. It is a claim about scope: the labs own the pace of the models, and whoever deploys agents owns the gate. Both have to hold.

§ 09What we are watching for#

• OpenAI’s evaluators. Altman’s September 14 thread commits OpenAI to safety cases before frontier RL runs and welcomes independent auditors, but names no evaluator team, no publication rights and no date. A post on openai.com, and the names, would turn the pledge into a fact. On September 16 OpenAI published a different lever, a framework for disclosing misalignment instances with six reports from its own training; the evaluator names are still owed.
• A start date from Anthropic. The essay says the near future; the first named review team, and who is on it, turns a commitment into a fact. The September 17 measurements post repeats the plan (evaluators “from multiple organizations”, verifying safety practices, reporting incidents and monitoring the published metrics) and names what they would monitor; still no start date, no names.
• Google DeepMind, Meta, xAI. A third lab committing to evaluators makes step one an industry norm; Hassabis and Musk endorsed the direction, Wang pledged more alignment work, Zuckerberg said on Tuesday that every lab has “the ability to take its own actions to ensure that happens”, and none of the three has committed to embedded evaluators. Microsoft, which trains its own MAI models, welcomed the idea on September 13 and published a code of conduct rather than a commitment on September 14.
• Washington, and the body. The executive branch has now answered three times in two days, and the answer is no: Trump at Doonbeg on Sunday, then two Truth Social posts on Monday calling the alarm a sick conspiracy and a hoax, and Vance calling the labs’ request “a bit of a trojan horse” on Monday, then telling the labs on Tuesday, at the All-In Summit, “if you’re building Frankenstein, stop”. Sacks told the labs to pace themselves without asking permission. On Monday evening, on CBS News, he said “I think this is becoming a panic”, told the labs they “should step aside for people who can do that” if they cannot make their products safe, and said pausing hands the lead to China. Congress is split by party: Johnson against a moratorium; Warner, Schumer and Sanders pressing for briefings, a treaty and a bill. The Information reports the three labs have discussed an industry standards body anyway, and Altman says OpenAI will not wait for an exemption. Whether that body gets a name, members and a charter is the item to watch, alongside the Sanders-Casar bill and the Trump-Xi summit Reuters dates to September 24.
• Who counts as independent. Sacks called METR “intertwined with Anthropic’s investors and staff”; Ball, at OpenAI, defended its work and asked for evaluators from outside one intellectual monoculture. The names on Anthropic’s first review team, and OpenAI’s, will answer this. Clark’s BBC line adds a second question for the same list: whether the kill switch every lab says it has is verifiable by a third party.
• Nvidia. Huang told two audiences on Tuesday that no new law is needed and that companies pace their own releases; whether any of Nvidia’s own Nemotron models gets an outside evaluator is the test of that position.
• Beijing. Answered officially on Monday, through the Foreign Ministry’s Guo Jiakun and a Global Times editorial, and the answer is a rebuke aimed at the essay’s China section. Still open: any Chinese lab speaking as a company (DeepSeek, Moonshot, Z.ai and the Qwen team’s account had posted nothing on it by 01:06 UTC on September 15), the mid-September US-China officials’ meeting on frontier AI risk that Reuters has reported, and the September 24 summit.
• Dumfries House. King Charles hosts Nvidia, Google DeepMind, OpenAI and Anthropic executives in Scotland this week; a palace readout, who attended, and whether it produces the shared set of principles Reuters says is on the table would all land here.
• The White House. Johnson says the AI chiefs meet Trump “within the next week”; the readout, and whether the safety talks Lehane described walk in with them, would land here.
• The Sanders-Casar bill. Text next week: a permanent ban on superintelligence and a pause until a federal regulator writes rules; the first vote count in either chamber is the number to watch.
• The coalition. More than 50 organizations and a tracker of 500 bills; the first bill it moves, and in which state, tells whether the Assembly was a rally or a machine.
• The courtroom. Buist v. Anthropic PBC (N.D. Cal., filed September 18) asks a federal court to treat the agreement this page records as a Sherman Act violation; the first defendant response, the first motion, and any DOJ or FTC statement on the pact land here and on our record of the suit.
• October 5 in New York. The City Council’s Committee of the Whole hearing on the risks of AI, with Altman and Amodei invited; whether either appears is the read.
• Standards, in OpenAI’s shape. The September 21 post asks for international technical standards for frontier AI and RSI, set through the AI safety institutes and CAISI, and says in the same breath what they are not: licenses, mandatory prerelease review or approval requirements. The first named body, and whether Anthropic and Google DeepMind sign the same page, would land here. The UN panel’s brief on the Hugging Face incident, published the same day, is the first intergovernmental read of the episode OpenAI’s post names as a preview of the risk.

§ 10Update log#

• September 13, 2026, 13:00 UTC. Added Demis Hassabis’s reply (22:59 UTC, September 12), Musk’s second post (04:07 UTC, September 13) and the Fortune interview published at 16:38 UTC on September 12, with the IPO and pact remarks. The silence window now runs to 13:00 UTC on September 13; still nothing dated from Meta, Mistral or a Chinese lab.

• September 14, 2026, 01:15 UTC. Added Sunday’s Washington entries: David Sacks (03:14 UTC), Trump at Doonbeg (Reuters, 14:28 UTC), Amodei’s CBS Face the Nation interview, Musk’s open-source post (06:45 UTC), Dean Ball’s METR posts (15:50 and 15:57 UTC) and Coxon on NBC via Reuters; the silence window now runs to 01:00 UTC on September 14, with the US executive branch no longer in it. Times are read from the posts and the Reuters page.

• September 14, 2026, 07:25 UTC. Added Sam Altman’s two-post thread of September 14 (04:05 and 04:18 UTC), the “more to share soon” landing; The Information’s September 13 report of standards-body talks among Anthropic, Google and OpenAI (Leo Schwartz’s post, 19:27 UTC); Meta’s Alexandr Wang statement of September 13, 06:03 UTC, which our two earlier entries missed while it sat inside the window we called silent; and the Sanders-Casar bill announced September 3. Meta leaves the silence list; Mistral and the Chinese labs remain on it to 07:00 UTC on September 14. Times are read from the posts’ ids and timestamps.

• September 14, 2026, 13:30 UTC. Added Sunday’s voices outside the labs: Senator Sanders’s treaty post (13:36 UTC), Lina Khan (18:31 UTC), Satya Nadella (19:36 UTC), and Qwen technical lead Junyang Lin’s personal posts (04:18 and 04:52 UTC, the first word from anyone at a Chinese lab, which our earlier windows missed); Monday’s Microsoft Code of Conduct draft (13:00 UTC), the Axios scoop on Tuesday’s event (10:00 UTC) and The Information’s report on the RUM Group compute deal, with the SEC filing’s own wording. Mistral checked on Arthur Mensch’s account: nothing since the essay. Times are read from the posts’ ids and timestamps.

• September 14, 2026, 19:30 UTC. Added Monday’s answers: Beijing’s, through the Global Times editorial (Reuters, 04:04 UTC) and Foreign Ministry spokesperson Guo Jiakun at the regular briefing (Xinhua, 08:52 UTC); Washington’s, through Trump’s two Truth Social posts (13:58 and 17:33 UTC), Vance at Joint Base Andrews (NY Post, 14:59 UTC), Speaker Johnson’s Sunday remark on CNN, and the Warner and Schumer items from Reuters and the NY Post; Arthur Mensch’s post (15:48 UTC), which takes Mistral off the silence list; Buckingham Palace’s Dumfries House gathering (Reuters, 08:59 UTC); Monday’s market moves from Reuters’ 18:23 UTC and CNBC’s 17:46 UTC updates; and The Information’s data-retention report (14:30 UTC) in the margin. The silence list is now the Chinese labs as companies, their accounts read at 19:00 UTC. Times are read from the posts’ own stamps and the articles’ publish times.

• September 15, 2026, 01:20 UTC. Added Monday evening’s two interviews: Jack Clark to the BBC (BBC Breaking post 17:28 UTC; article published 18:54 UTC), with the UK government’s earlier rejection of a mandated kill switch as the BBC records it, and David Sacks to CBS News (published 22:28 UTC). The Chinese labs’ company accounts, read again at 01:06 UTC on September 15, still carry nothing. Times are read from the posts’ own stamps and the articles’ publish times.

• September 15, 2026, 07:15 UTC. Added Monday afternoon in Los Angeles: Jensen Huang on stage at the All-In Summit (Business Insider, 20:41 UTC; TechCrunch, 21:51 UTC), the first frontier-hardware CEO to answer the essay in person, and President Trump’s call into the session; and Bilal Chughtai’s resignation post from Google DeepMind (20:13 UTC), the second safety researcher in a week to leave a frontier lab publicly, using the essay’s own word. The Pro-Human Assembly runs today, 9:00 to 17:45 Eastern in Washington; the next entry carries what is said there. The Chinese labs’ company accounts were last read at 01:06 UTC on September 15.

• September 15, 2026, 13:30 UTC. Added Elon Musk’s own explanation of “Dario is right”, given on stage at the All-In Summit with Gwynne Shotwell and published by the All-In Podcast at 03:59 UTC on September 15: the danger is what he agreed with, the remedy he wants is pre-release peer review by competitors, and xAI has still committed to nothing. Quotes are read from the recording’s transcript, fillers removed. The Pro-Human Assembly opened in Washington at 13:00 UTC; the next entry carries it.

• September 16, 2026, 01:30 UTC. Added Tuesday in Washington: the Pro-Human Assembly (Sanders, Bannon, Casar, Trahan, Roy and Tegmark, from CBS, NPR, PBS, the Boston Globe and Sanders’s prepared remarks), the Pro-Human AI Coalition launch (PR Newswire, 16:00 UTC), Chris Lehane’s briefing on the three labs’ weeks of safety talks (TechCrunch, 15:47 UTC; Andrew Curran’s post, 15:12 UTC), Speaker Johnson’s White House meeting (CBS) and Musk’s 05:46 UTC repost. Coxon, expected on stage per Monday’s Axios report, is not on the published lineup or in the coverage we read. Watching list updated.

• September 16, 2026, 13:30 UTC. Added Tuesday in San Francisco and Los Angeles: Dario Amodei and Jensen Huang on the Dreamforce keynote stage (Salesforce recording, published 23:00 UTC; CNBC, 20:00 UTC), Huang to Jim Cramer (CNBC, 22:13 UTC), Sam Altman’s fireside with Benioff (Salesforce recording, published 03:12 UTC on September 16), Sarah Friar to Cramer (CNBC, 19:25 UTC), Mark Zuckerberg’s post (23:01 UTC) and Vice President Vance at the All-In Summit (All-In Podcast recording, published 21:13 UTC; the clip on X at 22:10 UTC; the Guardian, 10:42 UTC on September 16). Quotes are read from the recordings’ transcripts; the Nvidia and Meta entries in the watching list updated. Anthropic’s accounts read at 13:00 UTC on September 16: no reply to Vance or Altman.

• September 17, 2026, 07:30 UTC. Added OpenAI’s framework for reporting model misalignment (openai.com, published 17:00 UTC on September 16) to the reactions table: the lab’s own statement that the industry has not solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and six misalignment reports from its training runs. Our record of the framework and the six reports carries them; the watching list’s OpenAI entry updated.

• September 18, 2026, 07:30 UTC. Added Anthropic’s September 17 measurements post (anthropic.com/institute, 20:32 UTC by its X post) to the reactions table, the Tuesday-night section and the watching list: the lab’s first word since the essay is data, not a reply, and it restates the evaluator plan without a start date or names. Our record of the three measurements carries the numbers and the method. Reworded one sentence in the FAQ.

• September 19, 2026, 07:45 UTC. Added the antitrust suit: four paying subscribers filed Buist v. Anthropic PBC in the Northern District of California on September 18 against all four labs whose leaders agreed with the essay, pleading Section 1 of the Sherman Act (Politico, page stamp 01:04 UTC; Bloomberg Law; the plaintiff’s post at 00:23 UTC), with its own record. Added Reuters’ 00:05 UTC report that Anthropic is weighing a new model to counter GPT-6 Astra. The watching list gains the courtroom.

• September 22, 2026, 01:30 UTC. Added four rows and two watching items: the New York City Council’s October 5 Committee of the Whole hearing with Altman and Amodei invited (its post, 12:01 UTC on September 16); the Journal’s report, via Forbes, that Zuckerberg, Musk and Huang called Trump against an industry-funded evaluation body; OpenAI’s September 21 post on standards (its position on RSI and its ask that the United States lead international technical standards that are explicitly not licenses or prerelease approval); and the Independent International Scientific Panel on AI’s thematic brief on the OpenAI-Hugging Face incident (advance unedited version 1, September 21). Widened the reactions table’s date range to September 21.

§ 11Sources#

Dario Amodei, We Must Pace the Frontier, darioamodei.com, dated September 2026, read September 12, 2026; his X post timestamped 2026-09-12T14:01:10Z. Jakub Pachocki, An Alien Mind, OpenAI, September 6, 2026. Reactions on X, each read from the post and its own timestamp: Elon Musk 15:01:32Z, Sam Altman 16:30:45Z, Sen. Bernie Sanders 18:53:10Z, Rep. Ro Khanna 16:29:50Z, Demis Hassabis 22:59:59Z, Clément Delangue 15:08:59Z. Fortune, Exclusive: Sam Altman addresses AI doomsday fears in new interview, Alyson Shontell, September 12, 2026, 12:38 PM ET, and Sam Altman confirms OpenAI won’t go public this year, Jason Ma, September 12, 2026, 2:04 PM ET. Elon Musk 2026-09-13T04:07:26Z. All-In Podcast, Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger, YouTube, published 2026-09-14T20:59:04-07:00 (03:59 UTC, September 15); the AI exchange runs from about 29:00 to 36:00 and 53:00 to 58:00. Anadolu Agency, Beijing rejects ‘distorting facts’, September 11, 2026. CellCog routing statements reflect our configuration on September 12, 2026. Sunday, September 13: Reuters, Trump says ‘very negative forces’ raising exaggerated concerns over AI, Bo Erickson and Padraic Halpin, published 14:28 UTC; CBS News, Face the Nation full transcript, September 13, 2026; David Sacks 2026-09-13T03:14:43Z; Elon Musk 06:45:54Z; Dean W. Ball 15:43 to 15:50 UTC and his reply 15:57:28Z. September 13 to 14: Leo Schwartz, The Information 2026-09-13T19:27:39Z and his follow-up 19:29:42Z; the article Anthropic, OpenAI, Google Quietly Discussed an AI Safety Standards Body, September 13, 2026 (paywalled; headline and date read from the page); Alexandr Wang 2026-09-13T06:03:06Z; Sam Altman 2026-09-14T04:05:35Z and his second post 04:18:19Z; Office of Senator Bernie Sanders, Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development, September 3, 2026, and its release summary. September 13 to 14, continued, each read from the post or page and its own timestamp: Junyang Lin 2026-09-13T04:18:17Z and his reply 04:52:42Z; Sen. Bernie Sanders 13:36:00Z; Lina Khan 18:31:42Z; Satya Nadella 19:36:34Z; Microsoft AI, Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models, published 2026-09-14T13:00:00Z per the page’s own data; Axios, Scoop: Coxon, Bernie and Bannon unite for “pro-human” AI, Maria Curi, published 10:00:05Z September 14; The Alliance for Secure AI Action, The Pro-Human Assembly 2026; The Information, Anthropic Strikes $13.7 Billion Compute Deal With Trump-Linked Rum Group, Valida Pau and Phoebe Liu, September 13, 2026 (paywalled; headline, byline and date read from the page); RUM Group Inc., Form 8-K, filed August 24, 2026; Reuters, Trump dismisses AI safety alarm, says US already has tools to police industry, Bhargav Acharya and Katharine Jackson, published 14:06:46Z September 14, 2026; Donald Trump, Truth Social posts of September 14, 2026, at 9:58 AM ET and 1:33 PM ET, read the same day; New York Post, JD Vance fires back after AI honchos demand to slow the pace of tech development, published 14:59:50Z September 14, 2026; Reuters, China state newspaper blasts Anthropic’s calls to slow AI as ‘Cold War’ tactic, Laurie Chen, published 04:04:45Z September 14, 2026 and updated with the briefing; Xinhua via People’s Daily Online, Fear-mongering will only disrupt process of global AI governance: spokesperson, 16:52 Beijing time, September 14, 2026; CNBC, China says AI CEOs’ call for a slowdown is ‘fear mongering’, published 09:36:27Z; CNBC, AI stocks slide after Anthropic, OpenAI CEOs urge slowdown, published 08:00:08Z, updated 17:46:41Z; Reuters, Wall Street drops as AI worries pummel chipmakers, published 09:16:42Z, updated 18:23:21Z; Reuters, King Charles to host AI executives as threat fears mount, Kate Holton, via MarketScreener, 04:59 EDT September 14, 2026 (the reuters.com URL did not resolve when we read it); The Information, X post timestamped 2026-09-14T14:30:08Z, and Anthropic Data Fears Prompt Nvidia, Palantir and Booz Allen to Restrict Model Use (paywalled; headline read from the page); Arthur Mensch, X post, timestamped 2026-09-14T15:48:23Z; the X accounts of DeepSeek, Moonshot, Z.ai, Mistral and the Qwen team, read 19:11 UTC September 14, 2026. Monday evening, September 14: BBC News, AI ‘kill switch’ may need to be mandatory, Anthropic co-founder tells BBC, Kali Hays and Faisal Islam, published 18:54:16Z, and BBC Breaking News on X 17:27:59Z; CBS News, Top Trump adviser on AI to Anthropic CEO: “Step aside” if you can’t control AI tech, Aimee Picchi, published 6:28 PM EDT (22:28 UTC), updated 6:47 PM EDT, and CBS News on X 22:15:13Z; the X accounts of DeepSeek, Moonshot, Z.ai, Mistral and the Qwen team, read again 01:06 UTC September 15, 2026.

Added September 16, 2026: Sen. Bernie Sanders, Prepared remarks: Regulating AI is as American as apple pie, sanders.senate.gov, published 2026-09-15T16:39:06Z; CBS News, Bernie Sanders and Steve Bannon join forces at bipartisan pro-human AI conference, published 12:05 ET; NPR, Bernie Sanders and Steve Bannon call for curbs on AI; PBS NewsHour/AP, Steve Bannon says we have to rise above partisanship to address AI hinge in history, 5:15 PM EDT; The Boston Globe, Will AI kill us all?, published 21:00 UTC; Common Dreams, End the AI arms race, Casar says, 17:30 UTC; PR Newswire, 50+ Leading Organizations Launch Coalition for Pro-Human AI, 12:00 ET; prohumanassembly.org and its agenda; TechCrunch, OpenAI, Anthropic, Google have been in talks on AI safety for weeks, 15:47 UTC; Andrew Curran, X post, 2026-09-15T15:12:31Z; Elon Musk, X post, 2026-09-15T05:46:12Z.

Added September 16, 2026, second entry: Salesforce, Dreamforce Main Keynote 2026, YouTube, published 2026-09-15T23:00:01Z (Amodei from about 45:15, Huang from about 1:16:41); Salesforce, Marc Benioff & Sam Altman | Dreamforce 2026, YouTube, published 2026-09-16T03:12:14Z; All-In Podcast, JD Vance on AI, Entitlement Fraud, Iran War, Israel, H-1B Abuse & the Midterms, YouTube, published 2026-09-15T21:13:22Z, and the clip on X, 2026-09-15T22:10:36Z; Mark Zuckerberg, post on X, 2026-09-15T23:01:38Z; CNBC, Nvidia’s Huang diverges with CEOs of Anthropic, OpenAI on AI safety at Dreamforce, published 2026-09-15T20:00:16Z; CNBC, Nvidia CEO Huang slams calls for AI antitrust exemptions, published 2026-09-15T22:13:52Z; CNBC, OpenAI CFO Sarah Friar on pacing the frontier, published 2026-09-15T19:25:42Z; CNBC, Zuckerberg on AI safety, published 2026-09-16T01:56:47Z; The Guardian, JD Vance dismisses calls for AI regulation, published 2026-09-16T10:42:07Z.

Added September 22, 2026: OpenAI, Building standards for the next phase of AI, openai.com, September 21, 2026 (news feed time 10:00 UTC); read September 22, 2026, in Chrome. Independent International Scientific Panel on AI, AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident, thematic brief, advance unedited version 1, September 21, 2026 (PDF, 20 pages); read September 22, 2026. New York City Council, post on X, 2026-09-16T12:01:07Z. Forbes, Zuckerberg, Musk And Jensen Reportedly Convinced Trump To Block AI Regulator, September 17, 2026, reporting the Wall Street Journal (paywalled; not read).