Finn Brooks: Juniper, okay — I've been sitting on this since Thursday and I need to say it out loud: the Trump administration just handed open source its biggest marketing moment in history, and they have absolutely no idea they did it.
Juniper Vale: Thursday? What happened Thursday?
Finn Brooks: Wait, not Thursday, I mean it started June 12, but I fell down the rabbit hole Thursday. The U.S. Department of Commerce issued an emergency directive. Forced Anthropic to suspend foreign nationals from accessing Claude Fable 5. No exemptions — like, not for Britain, not for Canada, not for anyone. No appeals process. Just: off.
Juniper Vale: Okay that I did not know — zero exemptions for allies?
Finn Brooks: Zero. And then — and this is the part that got me — they lifted it. Access came back to customers within about 24 hours of the lift order. So the whole thing lasted weeks and the resolution took one day. That's not security policy, that's compliance theater, and every developer on earth watched it happen in real time.
Juniper Vale: I mean — and that's the part that actually matters, right? Because if your business just got cut off from a frontier model with no warning and no appeals, you start asking whether closed models are even a safe bet.
Finn Brooks: Exactly! And meanwhile the Trump administration had already squeezed the pre-release window — like, Anthropic and OpenAI went from 90 days of government pre-release access down to 30. Closed-model releases now pile up compliance review steps that add weeks. But open-source releases on public repositories? Largely outside the new oversight scope entirely. They accidentally made open source the path of least resistance.
Juniper Vale: So the question I actually want to dig into is: is there a coherent strategy underneath this, or are we watching policy that's contradicting itself in real time?
Finn Brooks: Okay but here's where it gets weird — the Trump administration's own AI Action Plan, the one called 'Winning the AI Race,' literally advocated for open-source AI as a geostrategic move. Like, that's in the document. So they're restricting closed models and *promoting* open-weight releases at the same time. That's not contradiction, that's — wait, actually maybe it is contradiction.
Juniper Vale: That's the thing I want to slow down on, because I think the plain-language version of this gets lost. Closed AI — think of it like a SaaS subscription. The company holds the keys. They can lock you out, like we just watched happen with Claude Fable 5 on June 12. Open-weight AI is like getting the actual source code. Once it's downloaded, no one takes it back. No kill switch.
Finn Brooks: Right — but the part that doesn't fit is *who else* has the source code now.
Juniper Vale: Exactly. And that's the China-shaped consequence nobody wants to say out loud. The experts are flagging this directly — if U.S. restrictions reduce the supply of accessible American closed models, and Meta's Llama downloads are spiking, and Mistral's checkpoints are spiking, those same weights are available to anyone globally. Including China, which is already offering cheaper open-source models.
Finn Brooks: No, I don't buy that they didn't see that coming.
Juniper Vale: I mean — I'm genuinely not sure. Because if you *did* see it coming, you'd have to explain why the same action plan promoting open-weight models also mandates that federal agencies only buy 'ideologically neutral' LLMs. Those two things are pulling in opposite directions. One decentralizes capability, the other tries to control it.
Finn Brooks: Wait — so the administration is basically telling the world to go open-source, and then being surprised the world includes Beijing?
Juniper Vale: That's the trap door. You escape the kill switch — you escape what happened to foreign developers locked out of Claude Fable 5 overnight — but you land somewhere with no compliance floor either. The oversight that made closed models feel restrictive is also the thing that was theoretically checking them.
Finn Brooks: And that's the trap door that Google just blew wide open — like, Gemma 4 drops as Apache 2.0, thirty-one billion parameters, runs on a *single* H100, and it's sitting third on the Arena AI open-source leaderboard. That's not a hobbyist toy. That's frontier.
Juniper Vale: Wait — a single H100? That's one GPU.
Finn Brooks: One GPU. And there's a Mixture-of-Experts variant — twenty-six billion total parameters but it only activates three-point-eight billion at a time during inference. So you get the quality, you spend a fraction of the compute. It is — I mean, that's actually the architecture doing the work nobody's explaining in the headlines.
Juniper Vale: Okay, so the 'open source can't compete at the frontier' argument just — it doesn't hold anymore.
Finn Brooks: That's Finn's partial win, yes! But — wait, picture a government IT procurement officer in Paris. She's just signed off on Mistral for sovereign AI. Fourteen-billion-dollar European company, open-weight model, Arthur Mensch literally stood in front of the French parliament and said 'closed AI traps your business' and 'Europe has only two years.' She felt good about that decision. And I can't shake what happens next.
Juniper Vale: And then the inference call routes through Azure's Dublin region.
Finn Brooks: Through Azure's Dublin region — which means U.S. CLOUD Act jurisdiction. Mistral controls the weights, totally. But not the compute, not the legal exposure. Microsoft, Google, Amazon — that's who's running the actual inference.
Juniper Vale: So 'open' and 'sovereign' are not the same thing — and I think that distinction is the one that gets flattened every time Mensch gives a speech.
Finn Brooks: Completely different things! A developer in Berlin downloading Mistral's weights gets genuine access parity — that part is real — but the moment inference hits Azure, you're back under U.S. jurisdiction. The kill switch didn't disappear, it just moved.
Juniper Vale: And whether Europe's legislative response is actually moving fast enough to close that gap — that's the part we need to get into, because the numbers there are not encouraging.
Finn Brooks: The EU Technological Sovereignty Package — that's the legislative answer, right? June 2026, there it is. And I want to believe it. I genuinely do. But Sachiko Muto, who chairs OpenForum Europe, called it 'the equivalent of a New Year's resolution to get fit.' Like — that's not a friendly critic. That's the chair of the pro-open-source advocacy group saying your sovereignty bill has no legs.
Juniper Vale: Aspirational and slow-moving were the actual words from the procurement side.
Finn Brooks: Slow-moving against what clock, though? Because Mensch said two years. Two years to build independent infrastructure before dependency locks in. That's not an open-ended runway — that's closing.
Juniper Vale: I mean — and this is the calibrated version I keep landing on — open source redistributes the kill switch. It doesn't eliminate it. The weights are free, but the compute isn't. And a legislative package that's basically still aspirational in June 2026 is not going to fix that gap inside Mensch's two-year window.
Finn Brooks: Wait, so the actual verdict is — winning the weights war doesn't win the sovereignty war.
Juniper Vale: That's exactly it. And nobody reconciled that with the other side — the Trump administration's AI Action Plan pushed open-weight models for geostrategic reasons, but the security crackdown on models like Anthropic's Mythos pulled the opposite direction. Those two positions are still just sitting there unresolved.
Finn Brooks: Mythos — the cybersecurity exploit concern — that's what spooked the tighter controls on closed models. And then the same administration is like, yes, but also, open source, good, strategic.
Juniper Vale: There's no reconciliation. That's — yeah, that's the incoherence. And until Europe's compute layer is actually European, the June 12 directive proved it: one jurisdiction, no exemptions, no appeals. That's still the world we're in.
Finn Brooks: New Year's resolution to get fit — while the gym is in Virginia.
Juniper Vale: That image is going to stick with me — the gym is in Virginia. I mean, that's where we started, right? Finn saying the Trump administration handed open source its biggest marketing moment and had no idea. And I think that's still true — except now I'd add: the developers who fled closed models to escape a kill switch might have just downloaded a distributed version of the same problem.
Finn Brooks: No, that's — yeah. Okay, hot take officially revised. I was describing open source winning as a victory. OpenAI and Anthropic under government vetting, Gemma 4 on Apache 2.0, Mistral's weights on Hugging Face — that whole shift, I called it a win. It's a consolation prize. A real one, genuinely useful, a kid in Lagos can run Gemma 4, that part matters. But it's not sovereignty. The distributed kill switch is just harder to see.
Juniper Vale: And harder to see makes it more dangerous, not less. That's — I think that's the sharpest version of where we landed. Anyway. Thanks for dragging me down this rabbit hole.
Finn Brooks: Worth every minute of the Thursday spiral.
