Finn Brooks: Okay Clara, I have a confession — I started preparing for this episode and I genuinely got stuck on one sentence and kind of couldn't move past it for like two days.
Clara Bennett: Which sentence?
Finn Brooks: Pat Gelsinger, former Intel CEO, now partner at Playground Global, tells CNBC on July 12th that AI demand is 'almost unlimited.' That's the sentence. Because 'almost unlimited' is — I mean, that's not a quote from a hype cycle, that's from a guy who ran one of the most important semiconductor companies in the world for years. And he's saying it on the same day that Michaël Trazzi is leading over 100 people past Anthropic's building, past OpenAI, past xAI, all the way to Dolores Park, holding signs asking CEOs to commit to a conditional pause.
Clara Bennett: The simultaneity is doing real work there. It's not just ironic timing.
Finn Brooks: Right — but the part that doesn't fit for me is Trazzi specifically. He's not, like, a person who stumbled onto AI fear from the outside. Former AI researcher. Now filmmaker. He's with Pause AI and QuitGPT on this. So the march isn't ignorance about the technology, it's — it's people who understand it saying stop, and the people funding it saying 'we literally can't stop, the demand won't allow it.'
Clara Bennett: And then you have to add the Daniel Moreno-Gama arrest. Same day — 20 years old, allegedly tried to burn down OpenAI's headquarters and Sam Altman's home. Which is the extreme edge of exactly the same frustration, just without the organizing principle.
Finn Brooks: Dude — same day as the march?
Clara Bennett: July 12th. All of it. And that's the stress test this episode is actually about — not who's morally right, but whether the world where enterprises are shifting to 'valuemaxxing' and the world where Stop the AI Race is marching through San Francisco are even speaking a language the other can receive.
Finn Brooks: But hold on — that framing makes it sound like a disagreement that just needs better communication. Like if the executives and the marchers got in a room, they'd find common ground. That's not what this is, right?
Clara Bennett: That's exactly the thing I want to stop and pin down. Because here's the plain version of it: this is not a disagreement. Two people in a burning building — one is screaming 'stop adding fuel,' the other is on the phone ordering more fuel because the customer list just got longer. They are not having a debate. They are not in the same conversation.
Finn Brooks: Wait — that's a different claim than 'they disagree about risk.'
Clara Bennett: Right. And the proof is the 72% number. Approximately 72% of AI researchers believe this technology poses serious long-term risks — and the industry is still projected to spend over $300 billion on development in 2026. That's not a community that disagrees about danger. That's a community that knows the danger and is scaling anyway. Which means the protesters and the executives aren't arguing about facts. They're not even operating on the same — I mean, the inputs are shared, the outputs are completely divergent.
Finn Brooks: And that's what makes Trazzi specifically so significant here — because he's in that 72%. Former AI researcher. He's not arguing that the risk is real, he already knows it's real. The march isn't a warning, it's a — no, actually, it's closer to a resignation letter.
Clara Bennett: That's a useful reframe. And the Stop the AI Race demand — that CEOs commit to a conditional pause — structurally assumes one major lab moves first and others follow. But Pat Gelsinger is simultaneously telling CNBC demand is 'almost unlimited,' enterprises are shifting to valuemaxxing. Those two things cannot both be true and result in a pause. The demand framing assumes a sequencing that the incentive structure actively forbids.
Finn Brooks: So the protesters are addressing CEOs whose investors are literally being rewarded right now for moving faster.
Clara Bennett: That's the new thing. Not that both sides exist — that tension has been there for years. What's new is that on July 12th, both sides made their single strongest statement, and neither one was a response to the other. It's a knowledge-action gap, not a values gap. And that distinction matters because you close a values gap by persuasion. You close a knowledge-action gap by changing the institutional structure — which nobody has yet.
Finn Brooks: But that framing — the knowledge-action gap — I think it's actually letting the labs off too easy. Like the take I keep seeing is 'these executives are choosing profit over safety, full stop, moral failure.' And I want to push back on that because — no, actually, it's worse than that. It's not a choice. Stopping unilaterally would just hand the lead to someone else. That's not cynicism, that's just the payoff matrix.
Clara Bennett: Prisoner's dilemma, you mean.
Finn Brooks: Exactly — but here's where I want to stress-test it, because Google, OpenAI, and Anthropic all dropped competing governance frameworks in the same mid-2026 window. FARO co-regulation, federal blocking authority, independent evaluation mandates. Three different proposals. If this were a clean prisoner's dilemma, they'd have the same incentive structure. But Anthropic's position is genuinely different from xAI's.
Clara Bennett: That's a real crack in the framing, yes. Asymmetric incentives break the classic version. But — and I think this is the part that actually holds — the competing governance proposals are themselves the dilemma. Whichever framework wins legislative adoption reflects one competitor's preferences. So cooperating on safety rules is still defection in disguise.
Finn Brooks: Oh — oh that's a nastier version of it.
Clara Bennett: Now add the Cold War arms-control analogy — TIME ran analysis asking whether that model works here. It doesn't. Nuclear treaties functioned because you could count warheads. State actors, discrete weapons, verifiable numbers. AI capabilities are distributed across hundreds of companies and open-source repositories. There is no count.
Finn Brooks: Which — wait, that connects to something. 'The Essential Convergence' report, the global compact one on extreme AI risks — cyber escalation, biological threats, loss of human control — it's basically saying current frameworks aren't just insufficient, they're the wrong category of tool entirely.
Clara Bennett: Right. And the verification problem gets even sharper when you look at the actual model approvals — the Sol and Mythos cases show that even existing oversight may be unverifiable in practice. That's something we need to get into because it makes any pause agreement structurally hollow before it starts.
Finn Brooks: So the bad take — 'executives are villains' — misses that the payoff matrix is designed so cooperation loses. The structure produces defection, not the character of the people inside it.
Clara Bennett: And that structural hollowness — that's exactly what Mythos and Sol demonstrate. Anthropic's Mythos model gets flagged in April 2026 for unprecedented cyber capabilities, White House briefly bans it from public access, then eventually clears it. That whole sequence happened inside a voluntary NSA-led pre-deployment evaluation program. Voluntary. The word is doing enormous damage there.
Finn Brooks: Wait — the NSA review was voluntary? Like the labs could just... opt out?
Clara Bennett: Structurally, yes. And Mina Narayanan at Georgetown's Center for Security and Emerging Technology said analysts lacked full visibility into whether the safety review for Sol was even adequate. That's not a critic outside the process — that's a senior researcher trying to evaluate the process itself and hitting a wall.
Finn Brooks: Okay that's — no but that breaks something important. Because if the government's own review of Sol is opaque to Georgetown analysts, then a pause agreement is just... a promise with no receipt.
Clara Bennett: Put yourself in this position. You're an enterprise AI safety officer. Your job — your literal job title — is to evaluate a foundation model before your company deploys it. The government ran a pre-deployment review on that same model. You ask what they found. And the answer is: we can't fully show you. The voluntary review was never, I mean — senior researchers described it as never structurally designed to handle models at current capability levels.
Finn Brooks: So her job is to evaluate something that the government evaluated without her being able to see the evaluation. That's — that's not a gap, that's load-bearing opacity.
Clara Bennett: Now here's what I want to flag as the thing to actually watch. Pat Gelsinger said demand is 'almost unlimited' — but then named energy as the constraint. Power grids. Water cooling. If that's the real bottleneck, the verification question may eventually land somewhere unexpected — not at the labs, not at the NSA, but at whoever controls the infrastructure.
Finn Brooks: Wait — you're saying the grid might be the only enforceable lever? Because you can physically see a data center's power draw.
Clara Bennett: You can meter it. Which means — and this is the uncomfortable implication — the most credible pause mechanism available right now isn't a treaty between labs. It's an infrastructure chokepoint that nobody designed for governance.
Finn Brooks: But that's the thing that I — the funding side is where it matters. Like if private capital is the engine, the pause decision will always, structurally, be a competitive disadvantage for whoever takes it first. Which means the 72% — the researchers who actually know the risk — they're not in the room where that decision gets made.
Clara Bennett: No. They're not.
Finn Brooks: So does the answer have to be upstream of all of this? Not who oversees the labs, not the NSA voluntary review — who funds them?
Clara Bennett: I don't have an answer to that. The 72% number just sits there.
