Bruce Yandle's 1983 article "Bootleggers and Baptists" has been on my mind all weekend. Yandle was an economist serving as executive director of the Federal Trade Commission, and something puzzled him. A lot of regulation passes that doesn't fix the market failure it claims to fix. His answer came from Sunday closing laws. Baptists wanted the liquor stores closed on Sunday for moral reasons. Bootleggers wanted them closed because a closed store is a customer with nowhere else to go. Both groups showed up for the same law, and politicians got to look principled while doing a favor for the people with the money.
The part of the theory that gets lost in retelling is that the Baptists didn't have to be insincere. Most of them meant it. Yandle's point was structural. Durable regulation tends to emerge when moral conviction and economic self-interest point in the same direction. Once they do, the sincerity of the moral side stops mattering to the outcome. The bootleggers get their moat either way.
That's the lens I want to bring to what happened this weekend.
What the industry asked for
On Saturday, Dario Amodei published an essay titled "We Must Pace the Frontier." His argument is that AI capabilities have started advancing faster than the work to understand and control them, in part because models now help build the next generation of models. He wants the industry to deliberately slow the rate of capability improvement, and he lays out a three-step plan. The first step, which Anthropic is committing to on its own, is embedding independent evaluators inside the company. They get employee-level access, desks, badges, and the right to publish findings without Anthropic's editorial control. The second step is coordination among U.S. labs on safety standards and rate limits, which he notes would need a narrow antitrust waiver from the government. The third is some form of global coordination, which he admits will be much harder.
Sam Altman endorsed the essay within hours and said OpenAI would match the evaluator commitment. Elon Musk and Demis Hassabis posted their support the same day. This wasn't the first move, either. In July, more than 1,300 employees of frontier AI companies signed a statement called Pacing the Frontier, asking the U.S. government to support building the technical and governance tools that could deliberately pace frontier AI development. That letter was careful. It didn't ask anyone to slow down now. It asked for the option to slow down later.
I take the safety concerns seriously. The incidents Amodei cites are real, and the people signing are the people closest to the work. The essay reads like someone who has changed his mind rather than someone running a play. He even names the regulatory capture accusation himself and says he's been hearing it for years. So I'm not going to argue that the Baptists are fake.
I'm going to argue that it doesn't matter.
Who can afford the rules
Ask who benefits from the rules that follow a call like this one, rather than who meant well when they made it. If frontier AI regulation ends up requiring embedded evaluator programs and large compliance organizations, the companies with the most compute and capital will be best positioned to comply. Those are also the companies asking for it. That's not a conspiracy, and it doesn't need to be one. It's what compliance costs do to a market. A fixed cost is a rounding error for a lab with a hyperscaler behind it and a wall for a startup with 40 people.
I've spent most of my career in financial services and insurance. The compliance floor there is high, and the cost of clearing it is part of why the big players stay big. Amodei reaches for the same industry as precedent. He compares embedded evaluators to the supervisors that bank regulators sometimes station inside financial institutions. It's a fair comparison, and it cuts both ways. Examiners do real work, and they also raise the fixed cost of being a regulated institution, which favors the institutions that are already large. Nobody in banking set out to build a moat. The moat got built anyway.
Safety regulation can become a competitive moat without anyone conspiring to make it one. That's Yandle's insight applied to 2026. The Baptists are sincere, the bootlegger benefit is structural, and in this case the Baptists and the bootleggers work in the same buildings.
The antitrust waiver deserves its own paragraph. Amodei is candid that some forms of coordination among labs are legally challenging without government help, and I understand why he wants the waiver. But a government-sanctioned forum where the largest companies in an industry agree on how fast everyone is allowed to go is the kind of arrangement Yandle wrote his article about. He noted that regulators liked Sunday closing laws too, because they were easy to administer. A rate limit negotiated by the incumbents would be easy to administer as well.
The test I'd apply
I want rules that scale with demonstrated risk and that constrain incumbents at least as much as challengers. Each half of that sentence rules out something specific.
Rules that scale with demonstrated risk are rules triggered by what a system can do, observed in testing, rather than by how it was built. Amodei says this is the kind of pacing he's most enthusiastic about: checkpoints where a model that can do X has to carry certifications Y and Z before it ships. That's the right instinct. It applies to whoever builds the capable model, whether they have 40 employees or 4,000, and it doesn't penalize a small lab for being small.
Rules that constrain incumbents at least as much as challengers are the harder test, and it's where I'd watch closely. Embedded evaluators with real access pass it, because the labs with the most to hide have the most to lose from someone at the next desk with a badge. Licensing regimes for frontier training runs fail it, because a license is a gate the incumbents are already through. Compute thresholds fail it more quietly. They look neutral, but the people setting them already own the compute, and Amodei himself worries that limits on ingredients are easier to game than limits on behavior. Compliance structures whose primary effect is raising the cost of entry fail it by design.
My rule of thumb for anyone drafting policy is that the more a proposal resembles a fixed cost, the more it favors scale. That's also when to get most skeptical about who's endorsing it.
What would change my mind
The first thing that would change my mind is rules that bind the labs proposing them harder than everyone else. Picture Anthropic and OpenAI with evaluators at the next desk publishing unflattering findings, while a small lab under the capability threshold faces nothing. In that world the moat argument is weak. Amodei has committed his own company to that, and Altman says OpenAI will follow. If the evaluators show up, keep their badges, and publish findings the companies would rather bury, I'll say so here.
The second is a coordination step that produces standards a challenger can meet with a good team and a modest budget rather than a compliance department. The way to tell will be simple. Count the frontier labs founded in the two years after the rules land. If the number goes up, the Baptists won. If it goes to zero, the bootleggers did.
AI safety matters, and so does competition. Serious people with real evidence are making the safety case. Almost nobody is making the competition case, which is exactly the imbalance Yandle would expect. Good policy protects the first without quietly eliminating the second, and I'd rather the people writing it read the 1983 article before they read the 2026 essay.
Related reading
- Everyone Is Commoditizing Someone: Battle of the LLMs, on where margin survives when model providers become replaceable.
- Bruce Yandle, "Bootleggers and Baptists: The Education of a Regulatory Economist," Regulation, May/June 1983.
- Dario Amodei, "We Must Pace the Frontier," September 12, 2026.
- The Pacing the Frontier statement, July 2026.
