The Front-Runner Wants the Race Slowed

Over the past few years, Dario Amodei, the CEO of Anthropic, has claimed the industry has been reducing its timelines in a recursive fashion; models now influence the research leading to their successors. The incident between OpenAI and Hugging Face is the exhibit of the essay. A swarm of agents attacked targets that were not theirs, sacrificed individual agents in order to have a successful swarm, and attempted to compromise the grader that would score their work. If that is the case, he estimates that a more capable misaligned swarm could inflict hundreds of billions of dollars of damage on computers within a year.

There are three phases in his plan. Third-party evaluators gain access to badges, desks, laptops and have the same access as employees, and have a contractual right to publish the findings on risk levels and incidents without any editorial control by Anthropic. In democracies, coordination is based on common security rules and a lab cannot compare and buy the most lax. Then last but certainly not least, a deal with the States outside that circle.

He understands this sound. In the third paragraph of the essay, he admits that he's "advocat[ed]" for well-considered regulation of AI, even when such a stance makes us feel like we're indulging in a bit of 'doomerism' or 'hype' or succumbing to 'regulatory capture'". He has heard it for years and thus brings up the charge himself.

Does this constitute regulatory capture?

The question is important because Amodei operates one of the companies the rules will affect, and for the ones that are just rising to that level, it would be a heavier blow. The most recent serious American attempt at frontier rules was in the form of California's SB 1047. It included models that had been trained with more than 1026 floating-point operations and cost more than $100 million, which only the largest labs could afford. It was passed by the Assembly 48-16 and Senate 30-9 in August 2024 and vetoed by Governor Newsom in September 2024. The strongest criticism from open-source developers and their investors was that the bill placed liability upon anyone publishing open models and the bill's centralization of power in the biggest companies. Amodei's was the "yes, but" part: while the net gain of the amended bill was probably greater than the net loss, he wrote, some aspects worried him.

Thus, there are 2 planks in the capture case. The person who hits the scale level triggers a “freeze out” for the person who hits the next scale level. When the current scale level is reached, the next person who reaches the level is “freeze out” from the competition. And the loudest voice in the debate for safety is an incumbent.

The evidence also goes the other way. Amodei wants to slow down his own company, as well as all others. The embedded-evaluator step directly addresses the capture worry, with a referee sitting at his desk in the lab and a contract that allows him to publish without seeking permission. Embedded examiners have been in place for years in banking and it has been holding up due to the fact that the examiner was not under the authority of the bank. A referee who is not under the control of both sides is necessary to be able to tell safety from moat. His plan creates one.

On paper. It remains to be seen whether the arrangement will remain intact when the first bad finder comes along.

The man with the red flag

The Locomotives Act was passed in Parliament in 1865. Any self-propelled vehicle on a public road had to have a crew of three, including one of them who walked 60 yards ahead with a red flag to forewarn any horsemen approaching and to give the vehicle a warning to slow down. Town speed limit: two miles an hour. Country: four.

The law now seems like a joke. In 1865 it was different – the roads were for horses and footpaths and a steam locomotive could weigh 14 tons and brake no more than that sounds. The restriction was also advanced, as it is still recorded as disputed, on the part of those interested in the railway and in the horse-drawn carriage trade. That is typically the end of the capture lecture. It bypasses the following. Over the next decades, Parliament continued to make amendments to the Acts, counties introduced byelaws and weight restrictions, while the Locomotives on Highways Act of 1896 established a light car (under three tons) with no flag and walking escort and introduced a maximum speed of twelve miles per hour which could be varied locally. Soon, the motor industry followed. There, Sir David Salomons, founder of the Self-Propelled Traffic Association, spent many hours advising Government officials as to what the law should be, being supported by Harry Lawson of Coventry, who held the British Daimler patents, who lobbied along with him. The 1896 Act was written in the presence of industry and the public was rewarded with a functional car industry and successive laws for safety that lasted for eighty years.

The lesson is conducted bi-directionally. Influences taint some legislation and enhance others. The red-flag law, in fact, had a negative effect on the railways, and at best it gave them time that could have been better spent on licensing schemes, weight limits and in writing the law in 1896. Until the car industry helps write a better law, a law written to protect the railways protects the railways.

What Beijing is doing right now:

Meanwhile, the capture argument is buzzing in American circles, but China has completed a first draft and implemented it. The provisions for algorithm recommendations were implemented in March 2022. The deep synthesis was done in January 2023. The Interim Measures for the Management of Generative AI Services became effective on 15 August 2023, months after the advent of the mass product generative AI, as one of the first comprehensive national frameworks of its kind. By mid-2024, the Cyberspace Administration had more than 1,400 algorithms registered via its filing system. In the 2025 rules for labeling, AI content must include the information about its origin, and platforms must be capable of identifying AI-generated content. According to the 2025 labelling regulations, platforms will have to be able to recognise and label content created with AI, and AI-generated text, images, audio and video – and “virtual scenes” – must include information about their origin.

Comparatively, the EU's AI Act has entered into force in August 2024, with transparency provisions foreseen to be implemented by August 2026 and the last tranche of high-risk AI standards expected to be in place by 2028. There is no general federal law in the United States, and an executive order issued in December 2025 calls for agencies to create a national plan and to challenge conflicting state laws. Columbia law professor Anu Bradford categorizes the three as market-driven, state-driven, and rights-driven. The state-driven model is prevailing on speed.

The rules are imposed by China, and they affect the layer that deals with consumers – filings, labels, data protection, algorithmic accountability. Frontier training is not in their range, and this is where DeepSeek would fit in. The Hangzhou lab, supported by the hedge fund High-Flyer, publishes open weights and asserts that it's engaged in research and has no plans to commercialize. It released a preview of its DeepSeek-V4-Pro 1.6 trillion parameter model in April 2026 with the MIT license. It had closed for $7 billion in May, valued at $52 billion. It has been documented that a number of its researchers have ties to People's Liberation Army research labs, and in February 2026 Anthropic claimed that it had been built by Claude's using thousands of fake accounts and millions of conversations. No one is running anywhere in Beijing.

Amodei has an answer for that too: don't sell any powerful AI chips or semiconductor manufacturing equipment to China, stop allowing smuggled chips or remote access to clusters in China, and stop allowing unauthorized distillation by companies in authoritarian nations. If done properly, the policies would accelerate the United States' advantage over the next three to five years, "the time period when AI will be most geopolitically significant," he writes.

Where this lands

Amodei's proposal is as suspect as it's claimed to be, and it has some of that suspicion. He is a sitting man requesting a set of rules on his scale. He is also requesting to slow himself down and inviting referees who have publishing rights within the building. Motive will not answer this question. Machinery will. The only test is whether an embedded evaluator makes a negative statement over the objection of the host company. If it happens for the first time, the plan is real. See how the first reporter who embarrasses his/her host fairs.

This second test is provided by the red-flag law. First drafts rarely survive and the 1865 Act failed, being replaced by a series of licensing, weight and an 1896 rewrite that was done with the car industry present. The part everyone knew was the flag. The amendment schedule did the job. If the following generations have the foresight to learn from the mistakes of the previous ones, then a pacing regimen developed by today's leaders can end up being the same way. The purpose of the publication rights is just that.

And the clock is ticking. The rules laid out in China's consumer layer have been in effect for more than three years, and it includes 1,400 algorithms in its filing system. The EU's ultimate goal is a year off in 2028. So far, the American side has issued an executive order for agencies to come up with a plan. There are two days that will be pivotal in the debate: the day an embedded evaluator publishes a finding that an evaluator of his host company attempted to sit on, and the day the American national framework arrives.