Pacing the Frontier: The People Building AI Just Asked to Slow It Down

01 · The Short Version

The people winning the AI race just asked for a speed limit.

"Pacing the frontier" is Anthropic CEO Dario Amodei's September 2026 call to deliberately slow the rate at which frontier AI capabilities advance, so that safety and alignment work has time to keep up. Within days, Sam Altman publicly agreed and committed OpenAI to the same independent oversight, and Elon Musk backed it in four words: "Dario is right."

Sit with that for a second. The three people with the most to gain from going faster are the ones asking to slow down. That is either the most important safety signal we have gotten in years, or a very clever business move dressed up as one. A lot of smart people are arguing about which.

I have been thinking about this all week, and I want to walk you through what was actually proposed, what genuinely concerns me, the case that this is all theater, and where I land. Then, because this is what I do, what any of it means for the way you run your business.

Sourcing note: the facts here come from Amodei's essay itself, the public posts from Altman and Musk, and coverage from TechCrunch, CNBC, SiliconANGLE, Reason, and The Register. The opinions are mine. Every link is listed and was verified at the end of this post.

02 · The Proposal

What is "pacing the frontier," exactly?

The core line from Amodei's essay: "We must slow the pace at which we improve the capabilities of AI models." Not stop. Not pause. Slow, so the safety work can catch up to the capability work. He is explicit that pacing does not mean halting model training or technical progress.

His plan has three steps, in increasing order of difficulty:

• A

  Embedded evaluators. Frontier AI companies give third-party safety evaluators ongoing, employee-like access to verify safety practices, report incidents, and assess not just finished models but the training pipelines behind them. Anthropic committed to this unilaterally; OpenAI said it will do the same.

• B

  Democratic coordination. AI companies in democratic countries agree on common safety standards and limits on the rate of unchecked progress, with government support where the coordination gets legally tricky.

• C

  Global coordination. The hardest one: democratic governments attempt to bring authoritarian governments, mainly China, into some form of agreement, starting with narrow deals both sides actually want.

03 · The Case for Slowing Down

Is AI actually moving too fast?

Here is my starting point, and I hold it as someone who uses this technology every single day to run a business: AI has already come so far that it is transformational exactly the way it is right now. Anything beyond this point, we need to be careful about. We need to be making sure AI is aligned before we make it more powerful.

That is not a vibe. Amodei's essay names specific things that already happened. Click each one.

In an earlier essay this year, Amodei put it plainly: we are considerably closer to real danger in 2026 than we were in 2023. Reading the incidents above, I believe him.

04 · The Other Camp

Or is this just a publicity stunt?

There is a real counterargument, and it deserves a fair hearing. The skeptics' version goes like this: the big AI labs want regulation because regulation is a moat. Rules written for frontier labs are rules open-source projects and smaller competitors cannot afford to follow. The Register ran its coverage under a headline calling the proposal "Big AI sets out its terms for regulatory capture." Investor David Sacks argued the market already handles this: models that behave unpredictably get punished by customers, so trading raw power for reliability is simply good business, no treaty required.

So where do I land? Honestly: there might be a little of both. Maybe there are some selfish measures in this, some self-preservation by the AI companies. I am not naive about that. But there is also real risk here, and I think that part is becoming obvious to everyone who is paying attention.

It would be very irresponsible to totally side with "this is fiction and people are making a big deal out of nothing for their own selfish gains." There is a lot more to it than that.

Both things can be true. The motives can be mixed and the risk can be real.

05 · The China Question

What about the race with China?

The strongest version of the skeptics' case is the adversary argument: we cannot slow down, because China will not. I take that seriously. It is why I agree we should not be sending them advanced chips, and Amodei's plan actually leans hard on exactly that: keep chip export controls tight, crack down on smuggling, prevent companies from distilling frontier models on the cheap. In his earlier essay he called chip export controls perhaps the most important single action we can take. Protect the lead first.

But holding the lead and talking are not opposites. We also need to meet with China. And here is the part I think people miss: China has its own reasons to want this technology under control. An authoritarian government does not want private companies inside its borders wielding AI more powerful than the state. Control is the whole point of their system. A world where frontier AI advances unchecked is a threat to them too, just for different reasons than ours.

Amodei's proposal reflects that reality with a tiered approach: start with narrow agreements both sides obviously want, like banning AI-driven bioweapons development, and work up from there toward limits on the riskiest capabilities, with the lead held by democracies the entire time. Beat them on chips. Meet them at the table. Those are the same strategy, not competing ones.

06 · The Real Question

Has "faster" actually made us better off before?

Strip away the technical detail and the geopolitics, and here is the question I keep coming back to. Are we better off today than before we had the internet? Maybe we are. I would probably say yes.

Are we better off than before we had social media? Absolutely not.

We have already run this experiment. We have already shown that we will push ahead through technology we do not really need, at full speed, without asking whether it is good for us, and that the harm shows up years later when it is too late to unwind. Social media was engagement algorithms pointed at human attention. This time the technology can act on its own, write its own successors, and pursue goals we did not give it. The cost of learning the lesson late is a lot higher than a generation of doomscrolling.

We have done "move fast and find out" once already. The bill came due. This one deserves the caution the last one never got.

That is not doomerism. That is pattern recognition.

07 · What It Means for You

What should a business owner actually do with this?

Here is the strange comfort in all of this. Go back to where I started: the AI that exists today is already transformational. That cuts both ways. It is why the frontier needs pacing, and it is also why nothing about a slowdown hurts you as a business owner. Nobody is taking away the capability that already exists. It is sitting there, mostly unused, in businesses that never got past treating AI like a search box.

I run my entire business on the models that exist right now: my marketing, my bookkeeping, my client delivery, my content. Not because I chase every release, but because the technology stopped being the bottleneck a while ago. The bottleneck is whether your business is organized so AI can actually work on it. If the frontier paused for two full years, most business owners would still have years of untapped leverage in front of them. I wrote about what that looks like in practice in my post on using AI agents without being a developer.

So my honest advice: pay attention to this debate as a citizen, because it matters and we had better take it seriously. But as an operator, stop watching the race. The gap that decides whether AI transforms your business is not the gap between this model and the next one. It is the gap between what today's models can do and what you have built with them.

08 · Questions People Are Asking

Pacing the frontier, quick answers.