Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters

The episode opens in the show's established register rather than with news. Jason Calacanis introduces himself as "the world's greatest moderator," Chamath Palihapitiya answers with a flat run of "Great. Great. Great." and the room detours into a garbled seventeenth-century French anecdote about prisoners in Paris being offered freedom on condition they marry a prostitute and move to Louisiana. The joke lands on David Sacks, who is told it explains "a certain of your proclivities." None of this is substantive, but it establishes the working method: an in-group comedy track laid over a docket that is, this week, unusually heavy.

David Friedberg returns after a week away; the verdict on his substitute is that "Brad was great" and the guest took "a little victory lap." Then the show pivots hard: "We got a full docket today. Lots of stories." The two anchors are Demis Hassabis's proposal for an AI self-regulatory body and the reported bid for PayPal. That pairing matters, because both stories are treated as tests of one question: whether incumbent institutions can absorb a new technology without strangling it.

The proposal, published as an X article, calls for a US-led international AI standards body modeled on FINRA, the Financial Industry Regulatory Authority. The mechanics are specific, and most of the subsequent argument depends on them: federally overseen but industry funded and run by independent technical experts; frontier labs submit models 30 days before release; participation is voluntary at first and mandatory at some later point; assessments cover cybersecurity, national security, biological threats and other high-risk domains; benchmarks refresh quarterly; and the body can coordinate a slowdown in development if conditions demand it.

The reception is presented as broad: Elon Musk called it thoughtful, with endorsements listed from Sam Altman at OpenAI, Jack Clark at Anthropic, Sundar Pichai, Satya Nadella, Jack Dorsey of Block and the Carlson brothers. Jason's summary of that guest list sets up the episode's running gag, that he had called this outcome on air before it was fashionable.

The clip he plays is the substance of that claim:

The whole industry is gonna need to be regulated, and I think the industry needs to regulate themselves. That's the key to this. We need to have a set of tests that Google, Microsoft, Amazon all agree to, Elon.

He extends the argument to certification: models should be self-certified by the industry before the government, "which doesn't understand the models," is asked to certify them. The precedents offered are the MPAA and the video game industry, and self-certification is called "the simplest thing in the world to do." The weakness is that the clip asserts simplicity without saying who writes the tests or how a bad certifier is punished, which is exactly the gap the segment then tries to fill.

Friedberg supplies the definitional work. An SRO is not a purely independent body: FINRA and the National Futures Association exist inside financial markets and were created so institutions could set their own rules, check each other, and keep the system safe, because they all trade risk with one another. The industry therefore has a self-interested reason to police itself and a self-interested reason to avoid slowdowns that make markets inefficient. The AI analogy holds, he argues, because there are many players all trying to advance the technology, and none wants an outside body dictating a fixed test list.

His strongest point is about regulatory half-life. California's AI legislation, he says, made little sense when written and mapped even less well onto the technology a year later. An SRO can adjust how tests are run, who runs them, and which independent experts are involved, with federal oversight but not federal control. FINRA and the FAA ultimately report to Senate and House committees, which gives those committees a lever over bodies meant to do the actual work. The AI version would bring in people who know how to assess cyber, bio and weapons risk, and because they already know how to run software evals, they could move faster than a newly chartered agency.

The concession embedded here matters: he agrees the industry now recognizes that "there needs to be some degree of oversight and checkpoints here." That is a retreat from the position that no external checkpoint is needed, and it is the opening Sacks later widens.

On buy-in, Friedberg says both OpenAI and Anthropic signed up, with the caveat that Dario Amodei may not have endorsed it personally while "Dario's president" and Sam Altman did. That detail goes unverified, and it matters, because the later argument about Anthropic's intentions depends on reading its endorsement as tactical rather than sincere.

Sacks frames his position as a surprise even to himself: he told Hassabis he could potentially support the proposal, speaking only for himself and not for anyone in government, because an SRO would be "infinitely better" than a new agency that "would rapidly become a DMV for AI." The phrase he attributes to Amodei is an "FAA for AI," and his objection is capability and tempo: government lacks the expertise to evaluate models, and "The criteria are changing too rapidly." The predicted failure mode is a queue that starts at one month and ends at many, costing the US the AI race.

He then lays out five conditions, the most concrete contribution anyone makes in the segment.

1. Broad representation inside the AI industry, explicitly including startups and open source rather than just the three largest labs, to make regulatory capture harder. 2. Review limited to true frontier models, meaning models that represent a step change in the state of the art, so lesser models are not held out of the market. 3. Catastrophic risk only, defined as cyber and CBRN (chemical, biological, radiological, nuclear), excluding disinformation and "microaggressions" so the body does not become a speech regulator. 4. Voluntary first, proving itself before being legally enshrined. 5. A substitute for a new agency, not an addition to one.

Condition two carries the sharpest anti-capture argument: incumbents could use review "as a way to tie up lesser models," and if a model is not at the frontier there is no incremental catastrophic risk to justify delay. Condition five is load-bearing, since he says plainly that "If it's just additive, then it defeats the purpose."

He then does the arithmetic on the alternative. FAA type certification for an entirely new aircraft design takes five to nine years, amending a certificate takes three to five, and the Boeing 737 MAX took about five. That is permission-based regulation: nothing flies commercially without approval. Swapping a months-long release cadence for a years-long queue, he argues, hands the race to China, which will not abide by the rules. Between an FAA for AI and a DMV for AI, he prefers the SRO, provided it is kept "honest and pure."