AIANALYSISPODCAST
Dario Amodei says the AI backlash is a trust problem, not a messaging one. Critics read his regulatory stance as a call for the industry to write its own rules. Whoever writes them, the public has a stake.
Who should have the power to shut down a frontier AI system: the company that built it, the industry that profits from it, or a public agency? In August 2026, the people building the most powerful AI models in the world are publicly fighting about that question, and the answer is being written into the next round of U.S. AI policy.
Anthropic CEO Dario Amodei opened the exchange with a two-part X thread reframing the AI backlash not as a messaging problem but as "fundamentally a crisis of trust." TechCrunch covered the post the same day. Days later, All-In co-host David Sacks pushed back in his Fortune column, arguing that Amodei's preferred regulatory setup is essentially a "DMV for AI," a public licensing body that would, in Sacks' view, give the government veto power over what frontier labs can ship.
The fight is not about whether frontier AI should be regulated. It is about who writes the rules, who pays for the auditors, and who gets to pull the plug.
The two models under discussion are not new. The financial industry has lived with both for decades. A "FINRA for AI" would be a self-regulatory organization, an industry-funded body that writes its own conduct rules, inspects its own members, and refers enforcement to a public agency. The brokerage industry has operated this way since the 1930s, and FINRA itself is the most-cited template. A "DMV for AI" would be closer to a state motor-vehicle agency, where the public writes the rules, the public pays the inspectors, and the public decides who gets a license. Both can require safety audits. The difference is who controls them.
Amodei has not publicly demanded either label. Sacks is reading between the lines, and his reading is contested. Investor Gavin Baker posted that "Anthropic Believes They Could Be the ONLY Company Left in the World," then walked the line back after Anthropic's Sholto Douglas corrected him. On the All-In E286 episode, Baker reframed it as a pattern argument consistent with Amodei's public messaging rather than a literal claim. The pattern: Anthropic's CEO talks about banning public release of trained model parameters, screening Chinese investment in U.S. AI labs, and the risk that future systems could rewrite their own training without human help, all in the same breath as the cure for cancer. Critics hear an industry asking to certify itself.
The legal-authority question is the part the public rarely sees. If the U.S. does create a licensing body for frontier AI, it could sit inside an existing agency (NIST for technical standards, the FTC for consumer protection, or the Department of Commerce for export controls), or Congress could charter a new body, the way it created the SEC in 1934 and the Consumer Product Safety Commission in 1972. Each option has a different answer to the shutdown question. NIST writes voluntary frameworks. The FTC can punish after the fact. A new charter could grant pre-deployment veto.
Right now, only Anthropic is on the record. Sacks' column is a critic's read of Amodei's intent, not a comparison of the four labs' positions. The other three frontier labs have not publicly answered the question. The All-In episode treated the regulatory-structure question as a binary because Sacks framed it that way. The actual map is wider.
The public-interest stakes are concrete. An industry-funded SRO can move faster, hire specialized talent, and update technical standards without waiting for a congressional cycle. A public agency is slower but answers to voters, files its rulemaking in the Federal Register, and can be sued when it overreaches. The question is which trade-off the country wants for a technology that, by Amodei's own framing, is on track to "actually cure cancer," and, by his critics' framing, could also recursively improve itself past human control.
Amodei's two-part thread warned that local backlash to AI data centers is coming. On the same episode, co-host Chamath argued via X that the data-center buildout now props up 200 to 300 basis points of annual U.S. GDP. Whoever writes the rules for the models will, very soon, also be writing the rules for the towns the data centers sit in.
The next move is reporting out where OpenAI, Google, and Meta land on the FINRA-vs-DMV question, and whether any of them is willing to be the second lab on the record. Sacks and Baker have already taken their shots. Amodei has already answered. The other frontier labs have not.