You have such good intuitions here. And for the record, I know Carey and I know that she doesn't have some sort of insider knowledge on this matter, so it makes it even more impressive how well you've been able to forecast this entire situation. We really needed a voice like yours in the AI safety community, somebody that actually understands both governmental and regulatory incentives, but also corporate and financial incentives. I am really looking forward to reading similar pieces!!
My god, I wish I had insider knowledge. That would make things SO much easier.
But seriously, thank you friend <3 I appreciate that you’re out here telling people to wake up and think more about AI governance realism, instead of “AI Governance” in the abstract, and calling out where enforcement is unlikely to hit the mark. We need more of this.
The asymmetry is even sharper than the piece let on: federal contracts are a rounding error against Anthropic's commercial ARR, so the "ban" mostly hurts the agencies losing access. Worth watching which cabinet departments end up routing around it through commercial resellers. That tells you who actually holds power in this administration.
Also, thank you for the citation, and for making the game theory explicit!
The variable I'd add: Mythos as leverage only holds if Anthropic keeps differentiating. The moment another lab ships a comparable sovereign-ready model, the negotiating position softens. Worth watching whether CISA's testing period produces public validation or quiet silence.
Oh, agreed. The downsides of innovation moats is that they don’t usually stay moats for long.
I still think Anthropic is a bit different though, as they’ve spent resources and time prioritizing getting lower in the stack within orgs. They historically have focused on cross-platform coding tools — which developers can use to create their own tools and solve legit in-house constraints. Their Desktop client shifts that from the CLI to a more user-friendly space, but it’s still mostly platform-agnostic, and connective. By contrast, Microsoft focused on building AI constrained within the Microsoft ecosystem (which is self-limiting). Google’s a bit better, but you still get the most value out of using it as part of the larger Google Workspace empire.
Meanwhile, OpenAI appears to have thrown a bunch of darts blindly at the wall, building out separate tools for narrow niches, largely disconnected from one another. Trying to focus will take effort. Not impossible, ofc, but they’ve put themselves at a disadvantage by spreading themselves too thin.
Sovereign-readiness is mostly an accreditation moat. FedRAMP High and IL5 take 12-18 months of paperwork even after a model is technically capable, so even if a competitor ships parity tomorrow, the authorization stack lags. CISA going quiet usually means findings got incorporated without public acknowledgment, which is its own kind of signal.
I suspect that CISA mostly went quiet because of DOGE and partisan staff reductions and hiring incompetent replacements. I know folks who left CISA, and it’s a shadow of its former glory. Plus, Trump tends to nominate the absolute worst / most incompetent / problematic people.
You have such good intuitions here. And for the record, I know Carey and I know that she doesn't have some sort of insider knowledge on this matter, so it makes it even more impressive how well you've been able to forecast this entire situation. We really needed a voice like yours in the AI safety community, somebody that actually understands both governmental and regulatory incentives, but also corporate and financial incentives. I am really looking forward to reading similar pieces!!
My god, I wish I had insider knowledge. That would make things SO much easier.
But seriously, thank you friend <3 I appreciate that you’re out here telling people to wake up and think more about AI governance realism, instead of “AI Governance” in the abstract, and calling out where enforcement is unlikely to hit the mark. We need more of this.
Let’s get loud on this stuff.
The asymmetry is even sharper than the piece let on: federal contracts are a rounding error against Anthropic's commercial ARR, so the "ban" mostly hurts the agencies losing access. Worth watching which cabinet departments end up routing around it through commercial resellers. That tells you who actually holds power in this administration.
Well, the NSA already has access to Mythos. State, Treasury, and Commerce are also continuing to use Claude, notwithstanding the Tweedict.
https://www.devicesecurity.io/federal-staffers-are-still-using-claude-despite-trump-orders-a-31437
Also, thank you for the citation, and for making the game theory explicit!
The variable I'd add: Mythos as leverage only holds if Anthropic keeps differentiating. The moment another lab ships a comparable sovereign-ready model, the negotiating position softens. Worth watching whether CISA's testing period produces public validation or quiet silence.
Oh, agreed. The downsides of innovation moats is that they don’t usually stay moats for long.
I still think Anthropic is a bit different though, as they’ve spent resources and time prioritizing getting lower in the stack within orgs. They historically have focused on cross-platform coding tools — which developers can use to create their own tools and solve legit in-house constraints. Their Desktop client shifts that from the CLI to a more user-friendly space, but it’s still mostly platform-agnostic, and connective. By contrast, Microsoft focused on building AI constrained within the Microsoft ecosystem (which is self-limiting). Google’s a bit better, but you still get the most value out of using it as part of the larger Google Workspace empire.
Meanwhile, OpenAI appears to have thrown a bunch of darts blindly at the wall, building out separate tools for narrow niches, largely disconnected from one another. Trying to focus will take effort. Not impossible, ofc, but they’ve put themselves at a disadvantage by spreading themselves too thin.
Sovereign-readiness is mostly an accreditation moat. FedRAMP High and IL5 take 12-18 months of paperwork even after a model is technically capable, so even if a competitor ships parity tomorrow, the authorization stack lags. CISA going quiet usually means findings got incorporated without public acknowledgment, which is its own kind of signal.
I suspect that CISA mostly went quiet because of DOGE and partisan staff reductions and hiring incompetent replacements. I know folks who left CISA, and it’s a shadow of its former glory. Plus, Trump tends to nominate the absolute worst / most incompetent / problematic people.
https://www.csoonline.com/article/4162801/the-curious-case-of-sean-plankeys-derailed-cisa-nomination.html
This is really a great explanation of the facts.
At the time of the problems with the Pentagon I had something in my mind like "Anthropic will be back".. Here we are :)