Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The White House Floated AI Model Vetting, Then Backed Off in 72 Hours
TL;DR
The White House floated pre-release AI model vetting, then retreated almost immediately — after the New York Times and Politico reported a possible executive order and Kevin Hassett compared it to FDA drug testing, Susie Wiles and other officials walked it back within roughly 72 hours as “partnership,” not regulation.
The hosts see a real split inside the administration — one camp is spooked by frontier capabilities like Anthropic’s “mythos model” and AI-enabled cyber risks, while another fears regulation will slow U.S. innovation and hand an advantage to China.
Their core objection isn’t oversight itself — it’s politicized oversight — Paul says he could imagine expert, apolitical model review in theory, but in today’s climate he sees “zero chance” it would stay scientific rather than turn into winner-picking or ideological filtering.
They use the FDA vaccine-study story as a warning about AI governance — citing a New York Times report that studies on COVID and shingles vaccine safety were blocked from publication despite involving millions of patient records, Paul argues that if government can politicize something this objective, AI model review would be even messier.
Dean Ball’s essay becomes the most concrete proposal on the table — despite opposing almost all AI regulation, Ball argues catastrophic risks like cyberattacks on hospitals, banks, and power plants will inevitably pull the state in, so the best option is an intermediary private institution that sits between frontier labs and government.
The hosts think some version of “soft nationalization” is probably coming anyway — if the government doesn’t directly control labs, they suspect it will exert pressure through procurement, national security, or even its own model-building efforts, with references to DARPA, In-Q-Tel, Intel, and the book “The Pentagon’s Brain.”
The Breakdown
The 72-Hour Whiplash on White House AI Vetting
The segment opens with a mini-drama: reports say the Trump administration is considering an executive order to create a federal review process for powerful AI models before public release, with Anthropic, Google, and OpenAI already in the room. Then Kevin Hassett goes on Fox Business and likens it to FDA drug testing — and that’s when the blowback hits.
From “FDA for AI” to “No, No, Just Partnership”
By that night, chief of staff Susie Wiles posts that the White House is not in the business of “picking winners and losers,” and officials start insisting Hassett was taken out of context. By Friday, Bloomberg reports the administration is instead preparing agencies to partner with AI companies on cyber defense, notably stopping short of requiring government approval for frontier models.
Why the Hosts Think the Administration Is Internally Split
Paul says the back-and-forth feels like the administration testing messages in public: Tuesday it sounds like model approval, by Thursday it’s suddenly draft language and collaboration. His read is that some advisers are genuinely alarmed by frontier-model misuse — cyberattacks, infrastructure threats, hostile nations — while others see regulation as a direct threat to innovation and U.S. competitiveness.
The Bigger Problem: Lawmakers Don’t Really Get Frontier AI Yet
He’s especially worried about Congress operating at what he calls a “beginner level” understanding of AI — maybe they’ve used ChatGPT, but they don’t grasp agentic systems, reasoning models, or self-improvement. That matters because they’re being asked to make national-security and economic decisions about systems they can’t realistically project even one or two years forward.
The Vaccine Study Example as a Warning Shot
Paul then pivots to a New York Times report that the FDA blocked publication of studies showing serious side effects from COVID and shingles vaccines were very rare, despite the work involving millions of patient records and millions in public funds. His point isn’t to relitigate vaccines; it’s that if even relatively objective science can get politically filtered, AI model evaluations — full of judgment calls about bias, outputs, and ideology — would be far easier to manipulate.
Dean Ball’s Case for Catastrophic-Risk Regulation Only
The most substantial section is a close read of Dean Ball’s essay “Before Leviathan Wakes,” which Paul recommends even though Ball starts from an aggressively anti-regulation position. Ball says he opposes nearly all AI policy ideas — from algorithmic discrimination rules to pauses and bans — but makes one exception: state action to manage catastrophic risks like cyberattacks on hospitals, banks, power plants, and potentially biological threats.
The “Intermediary Layer” Instead of Soft Nationalization
What Ball supports, and what the hosts find useful, is not direct nationalization but “sustainable methods of perpetual interference”: private institutions positioned between the state and frontier labs. The idea is to give government some control and the feeling of control, without letting politics decide which CEO or model gets approved.
Why They Think the Government Will Move Either Way
The conversation ends on a blunt note: Mike says he doesn’t see a plausible path where nationalization pressures don’t increase, and Paul says his baseline assumption is that the government is already building — or will build — its own models. They cite DARPA history, the book “The Pentagon’s Brain,” In-Q-Tel, and even Intel to make the point that none of this is conspiratorial; the national security apparatus has been playing this game for decades.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
