Nobody Pulled the Brake

When DeepStack read the public laws of the 119th Congress on October 31, 2026, the list ended at PL 119-136 of October 22, and no title on it named AI; the slip-law texts of the twelve laws enacted after October 15 do not mention artificial intelligence. The FRONTIER Act, H.R. 9925, has three actions, all of July 23; the Ban Artificial Superintelligence Act, H.R. 10538, has two, both of September 24. Executive Order 14409 of June 2 directs, in section 3(b), the design of a voluntary framework under which developers could give the government access to covered frontier models for up to 30 days before releasing them to other trusted partners, and section 3(c) of Executive Order 14409 says nothing in that section authorizes mandatory licensing. Executive Order 14434 of September 29 defines "Super Intelligence" as what the statute already calls artificial intelligence and gives the President's science adviser, in section 3(b), 60 days, to November 28, to send the President proposed language for a federal definition. Nothing in the order requires that language to be published.

Most of the public wants AI slowed anyway. Quinnipiac polled 1,202 adults from September 24 to 27 and reported on September 30 that 47% want powerful AI systems slowed until their safety can be evaluated and 30% want them stopped until then; 14% would keep the current pace and 5% would move faster. In the same poll, 69% called keeping up with China important.

This Part asks whether a brake in Washington hands the race to Beijing. DeepStack's answer begins with a count: nobody has pulled a pace brake. DeepStack read 225 sources by 85 voices dated June 14 to October 18, 2026, and verified 75 disagreements, 31 of them direct contradictions and one refutation. A disclosure: DeepStack is built with Anthropic's models, Anthropic's chief executive wrote the essay that asked for the brake, Anthropic is the lab both federal security actions reached, and it is the first-named defendant in a private antitrust suit filed over that request. Since October 26 it is also a company in registration for a public offering, and its prospectus lists all three among its risks; Part twenty-two reads the filing. Its critics' case is in DeepStack's reading below.

The brake count

DeepStack's measure counts binding instruments, with weights set before the reading: a federal law 3, an enforcement action with a sanction 2, a state law in force 1, an executive act with effect 1, a pledge 0. Each instrument goes in one of three columns, by what it binds: the pace at which a lab trains and releases, what a lab must disclose, or a lab's conduct on national security; an order that stops a model on security grounds goes in the security column even when it halts a release. The pace column is zero: no federal law binds a lab's pace, and both executive orders score zero, because the June framework is voluntary and the September order binds no private lab. The disclosure column holds California's SB 53, chaptered September 29, 2025 and scored 1 as the one state law in DeepStack's count already in force, and New York's RAISE Act, under which, Governor Hochul's office said on September 21, large frontier developers register from November, report from January 2027 and disclose critical safety incidents within 72 hours; it scores once its reporting duties begin in January 2027. Both laws govern what a lab must disclose, not when it may release. The disclosure column totals 1. The security column totals 3, from two federal actions against one lab, set out below: the March designation, an exclusion the D.C. Circuit has upheld and Anthropic has asked it to rehear, scores 2 as an enforcement action with a sanction, and the June directive, which stopped two models for every customer, scores 1 as an executive act with effect. The count is American: the European Union's AI Act, which binds providers that serve Europe, sits outside it.

Max Nadeau told Nathan Labenz on The Cognitive Revolution on September 27 that, as he understands it, the only legally mandated role for outside auditors is to check compliance with safety policies the companies write for themselves under state law, which leaves them little to test. Ariana Salvatore said on Morgan Stanley's Thoughts on the Market on October 7 that targeted action on frontier oversight could occur, but that Congress is unlikely to pass comprehensive legislation without a high-profile safety or security incident, and that executive agencies will remain the more important actors on model access.

Jensen Huang, as Tae Kim of Key Context reported on September 29, said anyone who wants to pace is free to do so without being told. Jigar Shah told Stephen Lacey on Open Circuit on September 15 that the companies building AI data centers do not have the public's interest at heart and that the administration lets them run hands-off to win a race against China. DeepStack verified the two as a disagreement over whether voluntary standards are discipline enough or the public needs protection from outside. Sam Altman, as Ruben Dominguez of The VC Corner reported on September 21, said the industry would slow down or stop if it could not keep safety ahead of capability, and Michael Spencer of AI Supremacy relayed on September 11 that OpenAI said it was pausing certain frontier training runs. The pacing in DeepStack's record is in private hands, and pledges score zero.

What actually stopped a model

The security column scores against the lab that asks for the brake, and both of its entries predate the September essay. On March 4 the Department of War designated Anthropic a supply chain risk to national security, and Anthropic wrote on March 5 that it did not believe the action legally sound and would challenge it in court. On September 25 the U.S. Court of Appeals for the D.C. Circuit upheld the exclusion 2 to 1 under the Federal Acquisition Supply Chain Security Act of 2018. Anthropic petitioned on October 13 for rehearing by the panel or the full court, and on October 23 the court gave the government until November 13 to respond. A second letter of March 4, under 10 U.S.C. 3252, fared differently: the Northern District of California set that designation aside on August 27 and held that the challenged actions were unlawful retaliation in violation of the First Amendment, as its order and the D.C. Circuit's opinion record, and the government appealed the judgment to the Ninth Circuit on October 26. The opinion records that the Department, which had asked for contract permission to use Claude for "all lawful uses," acted after Anthropic refused to relax its prohibitions on the use of Claude for lethal autonomous warfare or domestic surveillance. Judge Gregory Katsas wrote for the court that the Department had "ample support for its conclusion" that Claude posed a covered risk, since restrictions Anthropic encodes in the model had stopped it from doing tasks government users asked for, and that the Department acted on Anthropic's refusal of a contract term it deemed essential, "not based on the company's support for greater governmental regulation of AI technology." Judge Karen LeCraft Henderson dissented, reading the statute's supply chain risk to mean deliberate, subversive acts, not a contractor's open enforcement of its own use limits.

Anthropic wrote on June 12 that an export-control directive, issued under national security authorities and received at 5:21 p.m. Eastern, suspended access to Fable 5 and Mythos 5 by every foreign national, and that it had to disable both models for all customers. Anthropic objected to the stop: Anthropic wrote that a narrow potential jailbreak should not be cause for recalling a model deployed to hundreds of millions of people, that the same standard applied across the industry would essentially halt new frontier deployments, and that it believed the directive was a misunderstanding. Geo Chen of Fidenza Macro wrote on June 15 that the White House acted because the government and Amazon believed the model's safeguards against use as a cyberweapon could be circumvented. Josh Hendrickson of Economic Forces recorded on July 2 that the control had been lifted for Fable 5 while Mythos 5 stayed limited. Anthropic's prospectus of October 26 names the Department of Commerce as the source of the worldwide restrictions and says Anthropic re-deployed both models once the government lifted them, after it updated its safeguards; it gives no date for the re-deployment.

Spencer wrote on July 6 that the administration had significantly damaged America's AI leadership through its actions against closed-source model makers. Hendrickson had written four days earlier that a cybersecurity risk to intelligence operations of the kind reported would be a genuine defense problem warranting action, and called the action a de facto ban on one model, though a temporary one. DeepStack judged that a refutation of the premise that the move was unwarranted; the judgment does not reach Spencer's estimate of its cost to American leadership. Part eighteen recorded the directive as the moment Washington reached past the chip. Here it counts on the security column, with the designation, and neither entry sets a pace.

The brake is the leaders' moat

The strongest case against the brake is that the leaders would write it. Steve Eisman wrote in his notes to the October 11 episode of The Real Eisman Playbook, with Gary Marcus, that the labs aim to create enough of a scare to force regulation that protects their franchise and builds moats around their profits. In Eisman's view, Anthropic and OpenAI deliberately let their agents cause havoc they could easily prevent, because reining the agents in would hurt their numbers. With open-weight models taking share, Eisman said, the big labs may be shaping rules that leave them a duopoly, and Marcus agreed they have to be thinking about it. Kevin Gee of A Letter a Day wrote on June 26 that the closed labs' lobbying for compute-threshold reporting, pre-release review and licensing reimports a setup that favors incumbents, and that sincere motives can still produce captured outcomes. Kris of Potential Multibaggers wrote on September 14 that rules fall harder on smaller rivals and open models, and that Chinese open weights are easy for lawmakers to deem unsafe, to the big labs' benefit. Mistral, the French lab, said in a statement Reuters published on September 18 that some incumbents are using the moment to consolidate their market position by pushing for regulation that favors them over competitors. Geo Chen had written on June 23 that a safety wall would cap next-generation models and give open-source and foreign labs room to catch up; DeepStack verified that as a contradiction of Gee.

Amodei's own text hands the critics their case. His post, dated September 2026, says Anthropic is unilaterally committing to the first of its three steps and calls on governments to require other frontier companies to match it, and that the most effective way to pace is regulation that targets all American frontier companies, because it reaches those unwilling to cooperate. Its second step is industry-wide coordination, with government mediation or waivers of antitrust restrictions, in which frontier companies in democratic countries set common safety standards and "limits on the rate of unchecked AI progress." The essay says some of that coordination is legally challenging and will require government support. Its third step has democratic governments try to coordinate with authoritarian ones. Scott Galloway of Prof G Media wrote on September 18 that the labs' regulatory ideas, employee-like access for outside evaluators and vague promises to welcome a federal safety framework, are performative, not genuine self-regulation; his own remedy, a binding one, follows. Spencer wrote on September 14 that the sudden alignment of Amodei, Altman and Elon Musk felt premeditated and possibly tied to an anti-China strategy that could reach Chinese open weights. Excellent AI Prompts noted the same day that Anthropic and OpenAI had announced confidential IPO filings and asked what they stand to gain from the policies they back, while declining to say the filings explain the timing of the warnings.

A third reading answers the question head on: the race is the reason to bind the labs, and soon. Galloway wrote in the same post that the AI race is a reason to speed regulation up, not to delay it, citing the Cold War, when adversaries cooperated to build safeguards against the spread of nuclear weapons. He would not ban research or slow the frontier, which he wrote would spill over into areas where AI is rapidly improving lives; he would bind it instead, starting with a U.S. regulator that records incidents, approves new models, writes safety rules and has subpoena powers, and would have it share authority in time with a Chinese counterpart in an international agency modeled on the International Atomic Energy Agency. Garrison Lovely told Nathan Labenz on The Cognitive Revolution on September 29 that even a national data center moratorium would disappoint those who expect it to slow AI, and that without regulation of the model developers and the chipmakers AI will move faster. Dave Karpf of George Washington University wrote in Tech Policy Press on September 17 that Amodei's scheme "amounts to a series of proposals that all benefit Anthropic," and that the pace should come from binding regulation with clear liability, not from rules the industry selects for itself.

Amodei's second step is already in court. Four paying subscribers to one or more of Claude, ChatGPT, Grok and Gemini filed a proposed class action, Buist v. Anthropic PBC, on September 18, 2026 in the U.S. District Court for the Northern District of California against Anthropic, OpenAI, SpaceXAI and Google. It alleges an agreement among the four to slow the pace at which each improves the products it sells, which the plaintiffs call a restraint of output under section 1 of the Sherman Act, and it says Amodei's proposal solved the cost of slowing alone "the way cartels always have." The complaint is an allegation: the docket, read October 31, shows no answer to it from any defendant and no ruling on the claim; an opposition to SpaceXAI's motion to transfer the case or compel arbitration was filed on October 27, and a hearing on it is set for November 18. Anthropic says in its prospectus that it disagrees with the claims and intends to defend the matter vigorously.

The loudest answer to the essay came from a rival lab's chief executive. Mark Zuckerberg, in an essay countering the Pacing the Frontier letter that Nathaniel Whittemore summarized on The AI Daily Brief on October 13, argued that open and widespread access must be the core tenet of the AI era even at some extra risk, because keeping unreleased models for internal use concentrates power in ways that could be worse than releasing them. He told John Collison on Cheeky Pint on October 12 that safety comes from many models in many hands, not a couple held by a couple of companies. Richard Socher of Recursive told Shawn Wang on Latent Space on September 14 that fearmongering had already led Europe to regulate heavily before its own AI takeoff, and called it unfortunate that some urge pacing while sprinting toward the frontier themselves. Ed Elson wrote on September 22 that if the leaders are right about extinction risk the labs should probably be shut down, and that if they lied or exaggerated, the remedies should address public deception and securities fraud; DeepStack verified that as a disagreement with Socher.

The critics also name risks of Anthropic's own. Dan Kagan-Kans of Understanding AI wrote on October 14 that Anthropic is in some ways less open than OpenAI about its operations, and that industry-wide mandated monitorability would make every lab's statements more believable. Spencer wrote on September 25 that Anthropic is close to unleashing recursive self-improving AI on drug research with hardly any oversight as it nears a $2 trillion IPO. Huang, in Kim's report, pointed to an open agent safety system backed by more than 100 companies, and Matt Barrie told Erik Townsend on MacroVoices on September 10 the opposite complaint: Anthropic's refusals are hurting its adoption, and safety work is pointless against open models. DeepStack verified Huang's point as a disagreement with Spencer on how much oversight exists, and Barrie's as one on whether Anthropic's safety posture is too thin or too strict.

The critics are right about the shape of the ask, and the ask is not yet a rule.

DeepStack's reading: Amodei asks for regulation of every American frontier company, which is the shape of the moat the critics describe. His second step asks for more: government mediation or a narrow antitrust waiver under which frontier companies in democracies would agree among themselves on limits to the rate of unchecked progress, proposed by the chief executive of a lab that would sit inside the agreement. Congress has written neither; section 3(c) of Executive Order 14409 says nothing in that section authorizes mandatory licensing; and the states ask for disclosure. The September 18 complaint alleges that an agreement among four leading labs to slow down already exists, and no court has ruled on it. Anthropic's prospectus takes up the question in its risk factors: rules of the kind it advocates "would apply to us" and could "slow our growth," and increasingly restrictive rules in the United States could let competitors in less regulated places become more advanced. The critics are right about the shape of the ask, and the ask is not yet a rule.

The distance sits outside the law

Epoch AI's data insight of January 2, 2026 has Chinese models trailing American capabilities by an average of seven months since 2023, with a minimum of four and a maximum of 14: an average over three years, not the gap on any one day. DeepStack read the current lag, counted as Epoch counts it in time to an equal score, in Epoch's Capabilities Index data on October 31, 2026, by the rule of the fourth test below: the best China-developed model, DeepSeek's V4.1 Flash 1017, released October 17, 2026, scores 158.83, a score an American model first reached on March 5, 2026 with OpenAI's GPT-5.4 Pro, 226 days earlier. On October 20 the reading was 133 days, for Moonshot's Kimi K3 of July 16; the lag grew because the new Chinese high came three months after Kimi K3 and scores half a point more. The best American model on the index, OpenAI's GPT-6 Astra of September 3, scores 169.06. Stanford's 2026 AI Index says the gap has effectively closed, with Anthropic's top model ahead by 2.7% as of March 2026. The two rulers measure different things: Epoch counts the months between the dates American and Chinese models reach the same capability, and Stanford compares benchmark scores at one moment. Amodei's essay says chips will be the main determinant of China's AI strength, and Epoch, in a report of September 4 updated September 24, estimates Huawei will produce less than 4% as much AI compute as Nvidia in 2026, and its chip imports explorer of October 27 estimates that of the AI compute that arrived from 2022 through June 2026, 72% went to the United States and 10% to China, counting the chips China makes itself.

Mia Silverio of Prof G Markets wrote on September 14 that the case for lighter rules to avoid ceding ground to China ignores that China already has some of the strictest AI rules in the world. Peter Thiel, in Kevin Gee's summary of September 29, argued that a precautionary default is not neutral and that, short of a one-world government with real teeth, slowing means the West restrains itself while China does not. Glen Kacher told Barry Ritholtz on Masters in Business on September 18 that open models cannot be regulated very well once anyone can install and adjust them.

Richard Bian, a product lead at a Chinese lab, told Jordan Schneider on ChinaTalk on October 16 that AGI cannot be stopped or slowed and that calls to slow down have failed before. Kevin Roose told Alex Kantrowitz on Big Technology on October 14 the opposite: the labs are racing toward recursive self-improvement, and they should stop. On paper sits China's foreign ministry readout of September 26: a China-U.S. AI Dialogue, a bilateral channel for AI incidents, and a next exchange in November 2026. The readout, the one step on record toward the global coordination of Amodei's third step, sets no limit on either country's labs.

What would settle it

DeepStack's Call is that there is no brake and no handoff. No federal law sets the pace of a frontier lab, the federal acts that reached Anthropic were taken on security grounds, and the main case on record against the request for a brake is the critics' case that it would serve the leaders at least as well as the public, a case no price or rule has yet tested. Chips set the distance to China, read on Epoch's ruler as the fourth test reads it, and open weights would slip any brake that came. Conviction stops at three of five because the distance test rests on a private index and because an incident, which Salvatore says comprehensive legislation would need, is on no calendar.

Four tests are recorded with this Part on November 6, 2026. The first asks whether a Public Law enacted by December 31, 2026 obliges private frontier developers to a safety framework, testing before release, incident reporting, licensing, or a limit on training or release; procurement, federal-contractor, funding, study and definition provisions do not count, so a defense authorization act whose AI duties reach only the Pentagon and its contractors, as in the House-passed H.R. 8800 and the Senate's S. 4784, would not count; and the list, read again on October 31, is read once more for laws enacted through November 5 before the test is recorded. The second asks whether whitehouse.gov publishes the definition of "Super Intelligence" that section 3(b) of Executive Order 14434 calls for by November 28, or a new bill proposes one, by December 31; a proposal delivered to the President and never published is a no. The third asks whether China's foreign ministry, the State Department or the White House publishes an official readout of a meeting of the China-U.S. AI Dialogue by December 31. The fourth asks whether, on March 31, 2027, the best China-developed model on Epoch AI's Capabilities Index holds a score an American-developed model first reached no more than three months (92 days) before the Chinese model's release; on Epoch's data of October 31, 2026, the gap was 226 days.

The first date to watch is November 28. The brake was asked for. It has not been built.