Juniper Vale: Tell me something — when you saw the Time magazine piece drop, what was your first actual reaction? Not the analysis, the gut reaction.
Finn Brooks: Honestly? I laughed, and then felt bad about laughing. Because Anthropic — the company that exists because its founders left OpenAI over safety concerns — just formally walked back its pledge to pause development if things crossed a danger threshold. That's not irony, that's like, load-bearing irony.
Juniper Vale: And then the Future of Life Institute's AI Safety Index landed, and the picture got sharper — not just Anthropic, but OpenAI, Google DeepMind, and Meta all weakened or eliminated those pause commitments. Every one of them. No company hit a top grade. Anthropic ranked first with a C+.
Finn Brooks: A C+! That is the best! That is the ceiling!
Juniper Vale: It's the ceiling, yeah. Ina Fried at Axios framed the whole FLI finding as a widening gap between model power and stated safety commitments, and the FLI panel used the phrase 'moving the goalposts.' Max Tegmark said companies are sprinting while the frameworks lag.
Finn Brooks: And here's where my brain broke a little — Anthropic drops the safety pledge, and then, like, within weeks, the U.S. government pulls Fable 5 and Mythos 5 offline. June 12th, export-control directive. Because Amazon flagged that Fable 5 could walk someone through exploiting a vulnerability.
Juniper Vale: A normal software bug. Not a catastrophic scenario — a bug. Treated as a national security event.
Finn Brooks: So the government is saying these models are dangerous enough to pull, and the company is simultaneously saying we can't keep a binding promise to pause if they get dangerous. That's not two separate news cycles — that's the same story running into itself. And that's what we're trying to figure out today.
Juniper Vale: But here's what I want to nail down, because I think the 'same story running into itself' framing — it's right, but it lets us skip the part that's actually new. Because people already suspected companies were soft-pedaling safety. What's new is that the Future of Life Institute put a grade on it. Like, an actual letter grade. And the best grade in the industry is a C+.
Finn Brooks: Right — but the part that doesn't fit is Max Tegmark isn't some outside critic. He's the FLI president. The institute that *built* the index is the one saying everyone is sprinting.
Juniper Vale: Exactly. And the FLI panel's phrase was 'moving the goalposts' — which is specific. It means the criteria for what counts as a dangerous threshold got quietly redefined. Not eliminated publicly, redefined. Think of it like a smoke detector that's hardwired to go off at a certain temperature. The pledges were that trigger point. The companies just raised the temperature until the detector basically never goes off.
Finn Brooks: And the best smoke detector in the building still only gets a C+.
Juniper Vale: That's it. That's the whole thing.
Finn Brooks: Okay wait — so it's not just Anthropic quietly dropping a pledge, it's Anthropic, OpenAI, Google DeepMind, and Meta all moving in the same direction at once, and the index just made that visible in a way that press releases never did. That's — I mean, that's the new thing, right? Not the betrayal, the *measurement*.
Juniper Vale: The measurement. Before this, it was vibes and press releases. Now there's a named index, a named score, and the voluntary safety regime has a grade. The grade is failing.
Finn Brooks: And nobody announced that the threshold moved. That's what 'moving the goalposts' actually means — the building didn't get safer, they just quietly recalibrated when the alarm goes off.
Juniper Vale: But that's the part the 'goalposts moved' story misses — because people are now reaching for government-gated access as the fix. Like, okay, internal pledges collapsed, but the government is stepping in, so there's still a guardrail. And I want to name that take directly, because I don't think it holds.
Finn Brooks: Wait, you think that take is wrong?
Juniper Vale: Look at the timeline. OpenAI previews GPT-5.6 — they're calling it Sol — on June 26, restricted to a handful of U.S. government-cleared partners, customer by customer. They framed it as the first model shipped 'when permitted, not when built.' That sounds principled. And then Anthropic's Mythos 5 gets pulled June 12, export controls, real intervention — and those same controls are eased June 26 and lifted entirely July 1. That's nineteen days.
Finn Brooks: Nineteen days. That is — okay, that is not a principled gate, that's a revolving door.
Juniper Vale: Dean Ball — who was inside the Trump administration on AI — described the whole federal posture as going from 'implausibly libertarian to increasingly draconian and opaque' in a matter of weeks. That's not a governance framework. That's whiplash.
Finn Brooks: Okay but — devil's advocate — isn't Dario Amodei at least consistent here? Like, he's calling for tighter open-source controls, same month. That's not a contradiction, that's a position.
Juniper Vale: That's — actually, no. He's calling for restrictions on open-source development the same month Anthropic drops its own binding internal pledge. He wants a gate on everyone else's models while removing the gate on his own company's behavior. Andy Konwinski wrote an entire essay about exactly this — safety language being used to lock down frontier AI and concentrate power. That essay came out right after roughly a hundred researchers met in San Francisco on June 30th.
Finn Brooks: And LeCun called the whole closed-lab model 'medieval obscurantism' — compared it to the Ottoman empire banning the printing press for two hundred years. I mean, that's a scorched-earth analogy but it's not wrong about the power concentration piece.
Juniper Vale: So 'safety' is doing a lot of work right now. It means pause internally when things get dangerous — until it doesn't. It means government gates — until those lift in nineteen days. It means restrict open-source — except your own models. The word is just covering whatever is convenient for whoever controls the frontier model at that moment.
Finn Brooks: And the pressure underneath all of that — DeepSeek, Z.ai, the cost math U.S. companies are already running — that's the part that actually explains why every pledge is bending, and we're going to get there.
Juniper Vale: And that pressure you're naming — it's not theoretical anymore. CNBC is reporting that DeepSeek and Z.ai are pulling actual U.S. business contracts. Not experiments, contracts. Because the performance gap is closing and the cost gap is not.
Finn Brooks: Okay that — the cost gap not closing, that's the thing. Like picture a startup CTO, it's Sunday night, she's staring at an Anthropic invoice and an API key from DeepSeek that does eighty percent of the job at twenty percent of the price. What does she do?
Juniper Vale: She switches. Or she at least pilots it. And that's the mechanism — not ideology, just the invoice.
Finn Brooks: And nobody at DeepSeek or Z.ai is pledging to pause at a danger threshold. They're not even playing that game. So every time Anthropic holds to a safety commitment, it's a cost they're eating that the competition just — doesn't.
Juniper Vale: Which is exactly when voluntary pledges have historically collapsed. Not when companies get evil — when the competitive math turns against them. Capability races are precisely that moment.
Finn Brooks: Wait so — I mean, does that mean the voluntary regime was never going to survive contact with actual competition? Like was FLI grading something that was already dead?
Juniper Vale: I think it survived until the margin pressure arrived. And now we're watching what fills the gap — binding regulation, some international coordination, or a two-tier system where U.S.-cleared models are one track and everything else is just... the other track that everyone quietly uses anyway.
Finn Brooks: And given that Mythos 5 export controls lasted nineteen days, I'm not holding my breath that the first track holds either.
Juniper Vale: And that's actually what I keep sitting with. Not the nineteen days, not the C+, not even Anthropic dropping the pledge the same month Fable 5 and Mythos 5 got pulled. It's — I mean, both mechanisms failed at the same time. The internal pledge collapsed, and the government gate lasted less than three weeks. So who actually holds liability when a frontier model causes real harm? Because I genuinely don't know who the answer is supposed to be.
Finn Brooks: Yeah, I don't either. And I don't think anyone has a credible answer right now.
Juniper Vale: That question's just — it's open. I think we have to leave it there.
Finn Brooks: Thanks for working through it with me. Genuinely.
