Adarsh's Guide to Cybersecurity, AI and CAREER Advancement

I’ve read a lot of “AI safety” essays that amount to a company patting itself on the back while changing nothing. Dario Amodei’s “We Must Pace the Frontier” isn’t that. It’s worth reading closely, because the specifics matter more than the headline.

What He’s Actually Asking For

Dario Amodei runs Anthropic. Anthropic builds Claude. And this week, Amodei published an essay — “We Must Pace the Frontier” — arguing that his own industry is moving faster than its ability to keep the thing it’s building under control.

He’s not calling for a pause. He says so directly: no stopping training, no freezing progress. Progress “will still seem fast.” What he’s actually asking is narrower, and honestly more interesting — slow down the rate at which capability grows, so safety work has a chance to catch up before the next leap.

Three steps. Let’s go through them, because the reactions to this essay tell you more than the essay itself.

Step 1: The One Thing Anyone Actually Committed To

Independent evaluators. Permanently embedded inside the lab. Employee-level access — not a quarterly checkbox audit, but someone inside the walls, all the time, watching.

Anthropic committed to this unilaterally. No conditions. Sam Altman said OpenAI would “do the same” within hours. This is real. This is the part of the essay with teeth.

Step 2: The One That Hits a Wall Immediately

Common safety standards across competing labs. Sounds reasonable — until you remember that coordinating with your competitors on anything is exactly what antitrust law exists to stop.

Amodei’s own essay admits this. It would need a legal carve-out to even be lawful. That carve-out doesn’t exist. Nobody’s holding their breath for it either.

Step 3: The Ask With No Mechanism at All

International coordination — including with China. Amodei draws his own line here: any restraint among democratic labs is bounded by how far ahead the U.S. stays. He’s not asking anyone to slow down into a disadvantage. He wants tighter chip and distillation controls running alongside the pacing, not instead of it.

Translation: two-thirds of this plan requires someone else to move first, and no one has.

Watch What Happened Next — Not What Was Said

Altman: “I agree with Dario that we need to pace the frontier… we will do the same.” Musk backed it. Nadella welcomed “deliberate pacing” and published his own AI code of conduct.

Everyone agreed with the one step that cost something concrete and specific. Everyone agreed in spirit with the two steps that require rivals to give something up. Spot the pattern?

My Take

I’d rather have the evaluators than not. Real access, real presence, real potential for someone outside the building to say “this isn’t safe” and mean it — that’s a genuinely different posture than “trust us,” and I don’t want to undersell that.

But strip away the framing and this essay commits exactly one company to exactly one thing. The rest is an invitation dressed up as a plan — sincere, maybe, but structurally empty until someone else’s incentives change. And nobody’s incentives have changed yet.

So don’t watch what anyone says next. Watch whether the evaluators actually show up. Watch what access they really get. And watch whether anything they find is ever allowed to see daylight — because that’s the only part of this plan that was never optional for anyone to agree to.

Everything else is just a really well-written invitation.