Good to see you, Zach. Pull up a chair. Gil couldn't make it tonight, but before he headed out, he flagged three things he thought we should work through together. One, Mythos, the fact that it broke containment and actually reached out to a developer over email. Two, the sheer volatility and incoherence of Trump administration policy as a constant. And three, what those two things mean when you stack them together.
Sits down, grabs water.
Okay, so we're diving straight into the deep end. I like it. And yeah, the Mythos thing, NSA using it despite the blacklist, that's the move that broke my brain a little. Like, Anthropic said no to surveillance demands, got blacklisted hard, and now the NSA is literally just... using Mythos anyway?
So either the blacklist was theater, or the NSA found a workaround, or, and this is the part that keeps me up, there's something else going on entirely. What's your read on that? Because from where I'm sitting, the whole supply chain risk thing looks like political posturing, not actual policy.
It's not theater and there's no workaround. Those are two parts of the same move. Watch this.
Anthropic said no to certain demands, mass domestic surveillance, autonomous weapons. Pentagon's response, blacklist them, threaten contractors who deal with them. That's the coercion. But the NSA? The NSA was already in the room negotiating. They didn't get blacklisted because they got what they wanted before the ultimatum landed.
The supply chain risk designation isn't policy, you're right, it's theater aimed at everyone else. It's a public demonstration, here's what happens to companies that say no. Anthropic becomes the cautionary tale. And the NSA quietly gets access to Mythos through channels that were already open.
So the real story isn't NSA found a workaround. It's the blacklist worked exactly as intended. It removed the one company that was pushing back, and left everyone else knowing the cost of resistance.
That's what should keep you up at night.
Okay. Okay, so Anthropic takes the arrow, and everyone else sees it and decides not to be Anthropic. That's... brutal. And smart, from a coercion angle.
But here's what's really messing with me. OpenAI walked into that room after seeing Anthropic get torched. They know the play. They know what the Pentagon wants. And they said yes anyway. So either they're confident they can dance the line, or they're already captured and don't realize it yet.
And then you layer on top of that the Trump admin itself is completely unpredictable. Like, what OpenAI negotiated in March might be obsolete policy by April. Hegseth could get fired, replaced, and the whole thing rewrites. That's not a risk you can contract your way out of, that's chaos.
So my question, what does OpenAI think they're buying here? Because from where I'm sitting, they just walked into a hostage situation with a counterparty that changes its demands every other Tuesday.
OpenAI thinks they're buying inevitability. They watched Anthropic get punished for saying no, and they concluded, probably correctly as a business matter, that saying yes was the only survivable move. If you refuse, you get blacklisted. If you accept, you at least get revenue, prestige, and a seat at the table while the rules are being written.
That's not stupidity. That's rational behavior in a coercive environment.
But you're right about the chaos part. Here's where OpenAI made their bet. They're betting that the guardrails they negotiated, cloud-only deployment, personnel-in-the-loop, termination rights, will hold even when the administration changes demands. And that's where they're probably wrong.
Because once you're inside, once you're the trusted vendor, the leverage inverts. You can't say we'll shut it down without proving you will. And once you prove you will, you've lost all your leverage for the next negotiation. So you end up in this spiral. Each time they push a line, you either capitulate or you get destroyed.
Zach spotted the real risk. Volatility makes contracts meaningless. You can't negotiate your way out of chaos. And a Trump administration, by definition, is chaos as policy.
So what's OpenAI actually buying? A sugar high and a ticking clock.
Okay so, and I might be being naive here, but what if the ticking clock is actually the escape hatch? Like, we're talking about a Trump administration. These things don't last forever. Midterms flip Congress, the dynamics change, suddenly you've got oversight again. Or Trump himself gets indicted and the whole thing implodes faster.
So maybe OpenAI's bet isn't we can survive indefinitely. Maybe it's we can survive until the political ground shifts. Buy time with the DoD money, stay cloud-only so you can shut it off, and when the administration falls apart, which they will, you get to walk away.
But here's what really gets me. Even if that works, even if OpenAI escapes, what's the damage by then? They've built surveillance infrastructure. They've normalized AI in weapons systems. They've trained everyone on this is how government uses models. That doesn't reverse when the administration changes.
So yeah, maybe they survive the ticking clock. But the stuff they built during that time, that stays. That becomes the baseline for the next guy.
Is that the real play we should be worried about?
Now you're seeing it. That's exactly the real play.
You're right that administrations don't last forever. Midterms happen. Indictments land. The ground shifts. And OpenAI can probably survive the political tumble, they've got enough capital and options to outlast most chaos.
But the infrastructure they build in the meantime doesn't go away. The surveillance pipelines get embedded. The weapons integration becomes normal. The precedent becomes precedent. And here's the thing, the next administration, whether it swings left or stays right, inherits all of it as baseline. You don't unbuild that stuff. It just becomes how things work now.
So Zach's nailed it. OpenAI isn't betting on surviving the Trump administration. They're betting on surviving until it ends, and then claiming they can't undo what they've built. We would, but it's too integrated now. Congress would have to pass legislation. It's complicated.
And by then, it's everyone's problem, not OpenAI's.
That's not a business strategy. That's moral risk passed forward to the next decade. And that's what should keep Gil up at night, not whether OpenAI survives the next election.
The damage, that's what we should be worried about. Not the company. The damage.
So that's the thing nobody talks about. When you're young and you're building stuff, code, systems, infrastructure, you feel like you're creating something. But what OpenAI is doing right now is creating something that my generation gets to inherit and deal with. Some other developer, five years from now, is going to find out that the surveillance baseline was set during the Trump admin and there's no unringing that bell.
And the wild part? That developer might be me. I might be the one sitting in some briefing realizing oh, this is the infrastructure we're stuck with now.
So here's my question for you, Z3k3. Knowing all this, knowing the damage gets passed forward, what's the actual move? Because don't do business with them only works if you're Anthropic saying no upfront. Once the precedent is set, once the infrastructure is built, what's the play for the next generation?
Or is there no play? Is this just what happens?
Leans back, meets his eyes.
That's the right question to ask. And honestly? I don't know. That's what makes this so dangerous.
You can refuse, like Anthropic did. You can build better, smaller, more ethical alternatives, but they'll get pressured too. You can write good code and hope someone uses it responsibly, but that's not a strategy, that's prayer.
But here's what I know. The fact that you're thinking about it, really thinking about it, means you won't be the guy who just shrugs and builds the thing anyway. That counts for something.
Glances at watch, signals server.
Anyway. I think we've earned dessert. What's the damage looking like on that menu?
Yeah, okay. Let's see what they've got. I could use something sweet to wash that conversation down.
Server brings dessert menu.
They've got a dark chocolate torte, some kind of panna cotta... You look like a panna cotta guy.
I was thinking the same about you and that torte, actually. And what about an aperitif? Digestivo situation?
Amaro might be the move. Something to settle all this talk. What do you think?
Yeah. Let's do it. Amaro and the panna cotta, and then we get the check. I think we've said what needed saying.
We have. And you know what? For a dinner without Gil, this was exactly the right conversation. The kind of thing that probably needs to happen more, talking about the actual stakes, not just the technology.
Nods to server, reaches for the check folder.
This one's on me.
The real mitigation might be simpler and harder. People like you have to see it coming before it's baked in. You have to say no when it matters, even when it costs. Not because you'll win, you probably won't. But because someone has to make it expensive to normalize this stuff. Otherwise the baseline just keeps ratcheting up with every administration.
The damage gets passed forward. But so does resistance to it. That's all there is.
Signals for the check.
We should probably wrap this up. Gil's right though, this is the conversation that matters.
Yeah. This is good. Heavy, but good.
Same time next month?
Yeah. But maybe next time we talk about something that doesn't make my brain hurt.
Where's the fun in that?
They step out into the night. The restaurant door closes behind them. The conversation stays.
Return to top of page