← All articles

You Can’t Pace a Race Everyone Else Is Still Running

Why the AI “slowdown” talk sounds noble, looks like a farm gate, and only works if China, open weights, and every three-letter agency politely agree to sit still.

Last week Anthropic’s CEO, Dario Amodei, published an essay called We Must Pace the Frontier. Within a day Sam Altman said he agreed, Elon Musk typed “Dario is right,” and half of Washington started nodding like somebody had finally said the grown-up thing.

Then I asked Anthropic’s own model to help me write a blog post arguing that this whole slowdown pitch is a consolidation play. Claude, bless its silicon heart, opened with:

Before I start — worth saying plainly: I’m made by Anthropic. You’re asking me to write a post arguing they’re running a propaganda operation. I’ll help you make that case as well as it can be made… But you should know whose house I was built in and weigh my input accordingly.

That right there is more honest than most of the op-eds. So I weighed it. And then I looked at what actually happened in the next two weeks.

The short story

Imagine three kids racing bikes around the block. One kid yells, “We should all slow down so nobody scrapes a knee!” The other two say “sure,” then keep pedaling. The kid who yelled already owns the nicest bike and the only helmet that fits the new safety rule. Meanwhile a fourth kid across town downloaded a bike from the internet and is already halfway down Main Street.

That’s the AI “pace the frontier” moment in one picture.

A real example from this month

On September 12, Amodei said the industry must slow how fast it improves model capabilities so third-party evaluators can check the work. He was careful: pacing “does not mean halting model training.” Good. So we won’t pretend he’s calling for a freeze.

He also wrote, in his own essay, that pacing inside democracies is limited by America’s lead over China, and that if we slow down by more than that lead, Chinese projects pull ahead. He called global coordination hard, and a full pause unlikely. In other words: the man proposing the slowdown already named the hole in his own plan.

Then:

• Altman and Musk endorsed the vibe. Musk also signed the 2023 pause letter and then built xAI. Pattern, not theory.
• September 28-ish, President Trump sat down with Amodei and said the quiet part out loud: “I’m for let’s go and let’s win.” Roughly a year to a year and a half ahead of China, nobody in third place.
• September 29, reporting lands on Anthropic’s IPO prospectus: the company warning investors that advanced AI could pose “catastrophic or existential risks to humanity”… while asking those same investors to buy the stock. Either it’s that dangerous, or it’s for sale. Both on the same form is a hell of a magic trick.
• White House AI adviser David Sacks had already called Anthropic’s approach “a sophisticated regulatory capture strategy based on fear-mongering.” Anthropic has put $40 million into Public First Action for AI policy and education. Fixed-cost audits and pre-release rules are a farm gate: a company with billions walks through; a startup hits the fence.

And the part nobody can regulate away: open-weight models from China and elsewhere — DeepSeek, Qwen, and friends — are already downloadable. Once the weights are on ten thousand hard drives, there is no recall button. No UN resolution puts that genie back. (Yes, I know Amodei isn’t proposing a full stop. The point still stands: you can’t pace what you can’t control.)

Why this is like the climate sermon that only one country hears

We’ve seen this movie. Tell one or two countries to tax themselves into oblivion while 150 others keep burning whatever they burn, and call it saving the planet. Same logic here. Slow the U.S. frontier with rules that only bind the companies who show up to the hearing, and you don’t make AI safer worldwide. You make the U.S. slower while everyone else — nations, agencies, and kids with a torrent link — keeps racing.

Maybe Dario believes every word. Maybe Altman and Musk do too on alternating Tuesdays. Belief doesn’t fix the coordination problem. And the people writing the rules are the ones who can afford to live with them.

What to do instead (the steps)

1. Say the real goal out loud. Safety work is good. Pretending a U.S.-only slowdown equals global safety is not. Keep the lead; don’t gift it.
2. Treat open weights as the water under the bridge. They’re out. Policy that ignores downloadable frontier-class models is theater.
3. Watch the farm gate. Every “third-party evaluator” and “pre-deployment review” is a fixed cost. Ask who can pay it and who gets locked out.
4. Hold the IPO contradiction up to the light. Existential risk in the prospectus and a buy recommendation in the same breath deserves a raised eyebrow, not a standing ovation.
5. Build and ship. For local and enterprise businesses alike, the race is who answers the phone first with a system that actually works and delivers the best value to the client. Falling behind on purpose is not prudence. It is a strategy someone else profits from — and for you to potentially go out of business. A serious non-starter in my book.

I’m not against guardrails on bioweapons, obvious misuse, or companies lying about what their models can do. I’m against a sermon that only the choir can afford to attend while the rest of the world keeps riding.

The genie isn’t going back in the bottle. The only question is whether America stays on the bike — or sits on the curb writing essays about pacing while somebody else wins the race.

— Andrew Easy Anderson

FAQ

Is Amodei calling for a full stop on AI?

No. In We Must Pace the Frontier he says pacing “does not mean halting model training.” He’s arguing for slowing capability jumps so evaluators can catch up — not putting the genie back in the bottle.

Then what’s the problem with “pacing the frontier”?

Coordination. Amodei himself says the plan only works if democracies keep their lead over China, and that a true global pause is unlikely. Meanwhile open-weight models are already downloadable. You can’t pace what you can’t control.

Is this just fear-mongering for regulation?

That’s the charge from people like David Sacks — “regulatory capture based on fear-mongering.” You don’t have to prove what’s in anyone’s head. Watch the mechanism: fixed-cost audits and pre-release rules are a farm gate. Billions walk through; startups hit the fence.

What about the IPO filing?

Reporting on Anthropic’s prospectus said the company warned of “catastrophic or existential risks to humanity” while asking investors to buy shares. Dangerous enough to scare the public, safe enough to sell. Both claims on one form deserve a raised eyebrow.

Doesn’t safety matter?

Yes — bioweapons, obvious misuse, lying about model capabilities. Guardrails on those are fine. A U.S.-only slowdown sermon that mainly binds the companies in the hearing room is not the same thing as making AI safer worldwide.

What should a normal business do?

Build and ship. For HVAC, dentists, brick-and-mortar, and enterprise alike: answer first, deliver value, don’t volunteer to fall behind. Sitting out the race on purpose isn’t prudence — it’s how someone else eats your lunch, and maybe your company.

Why the climate comparison?

Same pattern: one or two countries tax themselves into oblivion while everyone else keeps going, and we’re told that fixes the whole planet. Slow only the U.S. frontier and you don’t freeze the world — you hand the lead to whoever isn’t in the room.