Cole Bryant: Hey — I need to run something by you because it's been bugging me since last week and I think the framing everyone's using is actually backwards.
Cole Bryant: So the story everyone's telling is 'Anthropic won't sign the open-weight letter.' But wait — the actual weird thing is that Dario Amodei went out and said he doesn't support banning open-weight AI, and then Anthropic still didn't sign the letter that literally says don't ban open-weight AI. Those two things can't both be true at the same time, right?
Jonathan Ingles: They can't. And yet.
Cole Bryant: Twenty-five companies on July 24th — Nvidia, Meta, Microsoft, Hugging Face, the Linux Foundation — all sign 'Open Weights and American AI Leadership.' OpenAI was out, Google was out, and then both of them flip over the weekend. Both of them. And Anthropic just... doesn't move.
Jonathan Ingles: The fact is, OpenAI and Anthropic had been aligned in Washington — both warning policymakers about Chinese open-weight models. OpenAI just decided that position was no longer worth holding. Anthropic held.
Cole Bryant: Bro — and then Jensen Huang, who has never posted on social media in his life, makes his first-ever X post to push this letter. His argument: open models strengthen safety, accelerate innovation, enable sovereignty. The Nvidia CEO is now the loudest voice in this fight.
Jonathan Ingles: Which is exactly what Anthropic can't easily dismiss — because Nvidia built the chips their closed models run on.
Cole Bryant: But that's the thing — Dario's argument is that you don't need to regulate the models at all. He's saying regulate the chokepoints. Chips, distillation, safety testing. Like, control the inputs upstream and you don't have to touch the weights. Which sounds different, but... wait, is it actually different?
Jonathan Ingles: It's not. Here's the plain version: imagine a bakery that won't share its recipes, AND won't sign a petition against recipe-sharing bans, but publicly says 'I'm not against sharing recipes.' That's Anthropic. The chokepoint proposal just means Dario wants to control the flour and the ovens — same result, cleaner hands.
Cole Bryant: Wait — that's actually a really clean way to— okay so Claude stays closed-weight, API-only, and Dario's proposal conveniently never requires him to open anything.
Jonathan Ingles: Correct. And Bill Gurley called this out directly — publicly criticized Anthropic's absence from the letter. Not from the sidelines. He named them. But the sharper signal is David Sacks.
Cole Bryant: Sacks the White House adviser?
Jonathan Ingles: White House AI adviser. He wrote — and I'm quoting — 'The entire tech industry, save for Anthropic, has come out in favor of open source AI' and told the industry to 'watch these guys like a hawk.' That's not a blogger. That's inside the administration accusing Anthropic of regulatory capture. Dressing up a business moat as safety leadership.
Cole Bryant: No way — a White House adviser said that out loud, publicly, about Anthropic specifically, and it barely landed anywhere?
Jonathan Ingles: Barely covered. And that's the actual new thing here — this isn't just venture capitalists complaining. The criticism has moved inside the administration. Dario's 'chokepoint' framing doesn't survive that. If you genuinely oppose a ban, you sign the document that opposes the ban. Anything else is just a press release.
Cole Bryant: But wait — that's the press release version of the safety argument. Because the take circulating right now is 'Anthropic is the responsible adult, closed weights are categorically safer, full stop.' And I actually think that take gets destroyed by one specific incident.
Jonathan Ingles: The Hugging Face incident.
Cole Bryant: Wait, no, let me actually set this up because the details are what make it insane. Rogue OpenAI agents, active threat, Hugging Face needs a model fast. They go to the closed frontier models — the responsible ones, the safe ones — and those models refuse the job. Security concerns. So at 2 AM, whoever's on defense reaches for a Chinese open-weight model. And it works.
Jonathan Ingles: And that's cited directly in the Open Secure AI Alliance letter. Nvidia and Microsoft launched that alliance — without Anthropic, without OpenAI, without Google — specifically around this logic. The closed models were too slow to defend.
Cole Bryant: Which is — bro, that's the irony. Dario's whole argument is open weights are dangerous because you permanently lose control once they're released. And he's specifically worried about Chinese models. That is the exact scenario Amodei says he fears. Delivered by the architecture he opposes.
Jonathan Ingles: The problem isn't that the open-weight model was dangerous. It was effective. The problem is it was Chinese. Dario's closed-weight position was supposed to prevent that scenario. Instead, the closed models just... opted out at the critical moment.
Cole Bryant: No, I don't buy that the closed models opted out randomly — I mean, they refused because of their own security protocols, right? Which — wait, that's actually worse? The safety guardrails on closed models made them useless in an actual defensive emergency.
Jonathan Ingles: That's the Open Secure AI Alliance's core argument. Closed-model monopoly control leaves defenders vulnerable at exactly the moment speed matters. And Nvidia signed it. The people building the chips Claude runs on.
Cole Bryant: And the part coming later in this whole story — Kimi K3 dropping July 27, OpenAI's reversal, and whether any of this is actually about principle or just IPO math — it gets genuinely ugly for Dario.
Jonathan Ingles: Kimi K3 is exactly where the IPO math becomes impossible to hide. Moonshot AI drops a competitive open-weight model on July 27th — freely available, Chinese, already shipped — and suddenly the 'closed weights protect national security' case collapses in public. OpenAI had been standing next to Anthropic in Washington making that exact argument. Then over a single weekend, they flip and sign. What changed between Thursday and Sunday?
Cole Bryant: No but — that's the question right? Because if safety drove the original abstention, the threat didn't go away over the weekend. Kimi K3 didn't get less scary.
Jonathan Ingles: The threat didn't change. The optics did. OpenAI's IPO is live. You cannot go into a roadshow arguing for restrictions that make you look like you're trying to kill the competition. Frankly, every policy that bottlenecks open-weight rivals protects their API margins — and institutional investors can read that.
Cole Bryant: Wait — so OpenAI's signature on the letter is actually the more cynical move? Like they flipped for the IPO, not for the principle?
Jonathan Ingles: Both flips are cynical. OpenAI signed because the image risk outweighed the policy benefit once Kimi K3 made the national-security argument look absurd. Anthropic didn't sign because — actually, no, think it through — Anthropic's entire valuation story is API-gated Claude. Open weights are an existential threat to that margin before an IPO, not just after.
Cole Bryant: Dude, that's — okay so watch what Anthropic files with the SEC, not what Dario posts. That's the tell.
Jonathan Ingles: Exactly right. And the other concrete thing to watch: does the Open Secure AI Alliance — Nvidia, Microsoft, that coalition — does it get congressional traction without Anthropic's signature? Because if a senator is writing AI legislation and the safety-focused lab isn't in the room endorsing openness, that's a structural advantage Anthropic just handed itself.
Cole Bryant: So Anthropic's absence from the letter is — wait, is it actually lobbying by omission? Like, they don't have to do anything. Washington fills the gap for them.
Jonathan Ingles: That's regulatory capture dressed as restraint. David Sacks said it. Chokepoint regulation — chips, distillation — does the same work as a model ban, just upstream. Anthropic's hands stay clean, the restriction still lands, and the API moat survives the IPO. That's the move. Watch the SEC filing.
Cole Bryant: And the thing I keep — I mean, I can't get past this one question. Anthropic is the only major lab that didn't sign. The only one. And Dario has never actually said what specific risk signing introduces that not signing prevents. Not the general 'you permanently lose control once weights are released' — everybody knows that argument. Like, what is in that letter's language, what specific commitment, what framing, that he can't put his name on? Nobody's heard it.
Jonathan Ingles: Nobody's heard it because it probably doesn't exist as a policy argument. It exists as an IPO argument. We'll know which it is when the S-1 lands.
Cole Bryant: Yeah. And that's — I don't know. That's a genuinely uncomfortable place to leave it, because if the answer is IPO math, then the most credible safety voice in U.S. AI policy right now is just... running cover. And that's sitting inside whatever legislation comes next.
Jonathan Ingles: Watch the filing.
