Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The AI Ad-Hoc Prior Restraint Era Begins
TL;DR
The White House already blocked Anthropic from widening access to Mythos — Zvi says Anthropic wanted to expand Project Glasswing to more companies, including amid pressure from the EU, and the White House informally said no, creating a de facto veto over frontier model deployment.
This looks like the start of an ad hoc 'prior restraint' regime for AI — the administration is reportedly considering an executive order and AI working group to review models before release, which Zvi argues may be necessary for true frontier systems but is dangerous if done informally and arbitrarily.
The core complaint isn't just regulation — it's arbitrary regulation — citing Neil Chilson, Dean Ball, and others, the video argues that 'because the White House said so' favors insiders, blocks planning, invites corruption, and can be worse than a formal statutory framework.
The government may not need a new law to get real leverage over labs — with Anthropic, OpenAI, Google, Microsoft, and xAI entering pre-release screening agreements with AISI/CAISI and federal procurement on the line, Jessica Tillipman says model review may effectively become the 'price of admission' for defense and government work.
Zvi thinks some kind of frontier model review was always likely once capabilities became undeniable — his line is that with exponential progress, the choice is often 'too early or too late,' and once policymakers internalize cyber and catastrophic-risk implications, a permissioning system becomes politically hard to avoid.
The best-case path is narrow, rule-bound, and limited to the biggest frontier models — he repeatedly says a carefully designed regime with clear procedures, real thresholds, and minimal abuse could be justified, but fears the U.S. is stumbling instead into a broad, improvised system built in panic.
The Breakdown
Mythos gets stopped at the gate
Zvi opens with the big claim: the White House told Anthropic not to expand access to Mythos under Project Glasswing, and Anthropic complied. He frames that as the government already deciding who can access a frontier model and on what terms, even though the legal basis is murky. The sting is that this may be low-harm in this specific case, but the precedent is the real story.
From one-off intervention to full 'prior restraint'
The next jump is the alarming one: reporting says the White House is considering a review process for AI models before release via executive order and an AI working group of officials and tech executives. Zvi's take is nuanced but sharp — for the most capable frontier models, some version of this may actually be necessary, especially with cyber risk and a possible 'hackastrophe' where attackers gain the edge. But he thinks the U.S. is heading toward the worst possible version: improvised, panicked, and easy to abuse.
The backlash, jokes, and policy whiplash
He runs through reactions from Neil Chilson, Dean Ball, Jessica Tillipman, Taylor Barkley, and others, and the mood is equal parts horror and gallows humor. Dean Ball jokes about writing absurd names like Jimmy Apples and Neil Gorsuch directly into statute 'for all time,' which captures how unserious and chaotic the current process feels. Zvi also corrects a New York Times claim about Britain: the UK's pre-release reviews are voluntary through the AI Safety Institute, not a mandatory pre-approval system.
Why this was always coming
Zvi argues this was never really avoidable if frontier AI kept advancing. His line is memorable: when you're dealing with an exponential, the choice is often 'too early or too late,' and once it's no longer plausibly too early, it's definitely too late. He says years of rhetoric painting any precursor to model review as tyranny poisoned the well, making it harder now to build something thoughtful before crisis politics takes over.
AISI/CAISI becomes the practical choke point
Then the focus shifts to the new pre-release screening agreements: Anthropic, OpenAI, Google, Microsoft, and xAI are now on the list, with Meta conspicuously questioned. Right now, Zvi says, these tests are supposedly informational, but the obvious next step is that failing a review — or getting a quiet 'don't release this' — starts carrying consequences. Jessica Tillipman adds the key insight: procurement leverage may let the government achieve much of a review regime without ever passing a standalone law.
The Pentagon, chain of command, and Anthropic tension
He also covers the coming White House AI policy memo, especially language saying labs working with the Department of Defense must not 'interfere with the military's chain of command.' Zvi thinks there is a reasonable interpretation and a dangerous one: either it's about not trying to override military decision-making, or it's a demand for unconditional obedience. The Anthropic feud hangs over this whole section, especially with Pete Hegseth reportedly calling Dario Amodei an 'ideological lunatic.'
One genuinely good idea: always have multiple model providers
Zvi does single out one policy move as plainly right: agencies should maintain access to multiple AI providers rather than depend on one. It's practical resilience — backups matter, providers can become unavailable, and different models fit different jobs. In a video full of warnings about concentration of power, this is one place where the government instinct lands cleanly.
The endgame: do it wisely or stumble into abuse
The closing argument is that parts of government have finally grasped the security implications of frontier AI and are now predictably 'freaking the hell out.' Zvi says controlled access can quickly become a tool for insider advantage, corruption, and broad social control unless it's narrow, systematic, and clearly rule-bound. The internet-will-sort-it-out approach was only viable if capabilities plateaued — they didn't — so now the real task is limiting scope and convincing policymakers not to build the dumbest version of the regime that may be coming.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
