A sharp argument over AI regulation has moved into public view, with Anthropic CEO Dario Amodei defending regulation as a way to limit the biggest AI labs while critics argue that the same rules could give those labs more power.
The debate is also about open AI models. Supporters say openness spreads access and prevents control by a small group. Amodei's answer is that open weights do not remove concentration if the most important advantage still belongs to whoever has the most computing power and AI chips.
How the argument started
The dispute began after investor Gavin Baker said on the All-In Podcast that Amodei had told people inside Anthropic that the company could one day be the only private firm left in the world, with only governments beside it. Anthropic researcher Sholto Douglas rejected that description, calling it "totally false."
Douglas argued that concentration of economic power is exactly the outcome Anthropic worries about. He described the AI market as intensely competitive because major companies are trying to build models that are smarter and cheaper.
Baker then framed the disagreement in simpler terms. If AI is dangerous, one response is to limit its spread and keep it in the hands of a small number of companies and political actors. Another response is to treat concentration as the greater danger and distribute AI as broadly as possible.
In Baker's view, trusting concentrated power to act benevolently is a historically weak bet. That concern sits at the center of the criticism aimed at Anthropic: the fear that safety arguments can become a pathway to regulatory advantage.
Why open AI advocates object
Former Meta chief researcher Yann LeCun sided with open AI in the exchange. His argument was that AI, like the printing press and the internet, expands human intelligence by widening access to knowledge.
LeCun also argued that societies need many AI systems with different values, much as they need a diverse press. The point is not only technical access. It is also about preventing any single view of what counts as good or bad AI from becoming dominant.
That position makes open models a political and economic issue, not merely a software licensing issue. If model weights are freely available, more groups can inspect, adapt, and build on them. In the open AI view described in the source, that broader access is a safeguard against centralized control.
Baker also said the wider industry has largely moved toward that side of the debate. He argued that almost every major company except Anthropic has signed Jensen Huang's letter backing open models.
Amodei's case for regulation
Amodei entered the discussion directly on X, in what the source describes as only his fourth post of the year. He rejected the idea that the choice is simply between regulation and regulatory capture.
His point was that some parts of Silicon Valley treat regulation as if it automatically means industry capture. Amodei called that too simple. He argued that regulation can also reduce corporate power and benefit ordinary citizens when institutions are built around ideas rather than individuals.
According to Amodei, Anthropic designs its regulatory proposals to slow the leading labs and give smaller companies more room. He pointed to California's SB53 law as an example, saying it fully exempts companies below a revenue or training-cost threshold from its requirements.
That detail matters because it is the core of Amodei's defense. He is not presenting regulation as a blanket burden on the entire AI market. He is arguing that the largest developers should face more friction because they are the actors closest to the frontier.
The chip problem
The strongest difference between Amodei and his critics is over whether open AI models solve the concentration problem. Amodei says they only solve part of it.
His reasoning is that freely available model weights do not make advanced AI equally available if running, improving, or competing with those models still depends on massive computing resources. In that view, market power moves from model access to infrastructure access.
That means the center of gravity shifts toward whoever controls the most computing power and the most AI chips. Amodei argues that this sends power back toward the big labs, because they are the players most able to obtain and use those resources.
He also said AI is centralizing by its very structure, mainly because of scaling laws rather than regulation. That is a different claim from saying regulation causes concentration. It suggests concentration is already built into the economics and engineering of frontier AI, with regulation serving as a possible counterweight rather than the source of the problem.
Sacks pushes back
Former White House AI adviser David Sacks challenged Amodei in a nine-part reply. Sacks said Amodei did not directly dispute Baker's report, even though he could have done so.
Sacks also rejected Amodei's description of the criticism as too broad. In his view, almost no one argues that every regulation is capture. The issue is whether a specific regulatory design lets an industry obtain rules that mainly serve its own interests while the public benefit remains diffuse.
To make that point, Sacks referred to Nobel laureate George Stigler's definition of regulatory capture. He argued that Anthropic understands this mechanism and has hired several former AI officials from the Biden administration while developing a large government operation.
Sacks was especially critical of Amodei's call for a federal agency to review top AI models before release. He mocked the idea as an approval office for AI and argued that such a review process would create long waiting lines. He also said it would weaken the US against China, which he said would not adopt the same requirements.
The result is a debate with no easy middle ground. One side sees open AI models as a necessary defense against centralized authority. The other argues that openness does not matter enough if the real bottleneck is access to chips and compute. Between those positions sits the central question now facing AI policy: whether rules can restrain the biggest labs without making them even harder to challenge.