I thought we all agreed to pace the frontier? 🥹

The AI companies promised to slow down. Then the model announcements kept coming.

In September, Dario Amodei, the CEO of Anthropic, published an essay titled “We Must Pace the Frontier.” My understanding of the message was: “We must slow the pace at which we improve the capabilities of AI models.”

Then in an unusual turn of events, the AI CEOs, who normally spend their days trying to kill one another, started agreeing on the subject.

OpenAI CEO Sam Altman agreed. Elon Musk quoted the essay and said, “Dario is right.” Google DeepMind’s Demis Hassabis supported the general direction too.

For a few hours, it appeared that the men competing to build the world’s most powerful AI had gathered beside the racetrack and agreed that everyone was driving too fast.

Nine days later, Musk’s company released Grok 4.7. The following day, Anthropic released Claude Opus 5.5. OpenAI released GPT-6 Sol and GPT-6 Luna that same day.

Then the announcements kept coming.

Anthropic released Claude Sonnet 5.5 on September 28 and Claude Haiku 5.5 on October 7. OpenAI released GPT-6.1 Sol on September 29 and started rolling out GPT-6.1 Terra on October 26.

At this point, keeping up with the model names feels like trying to remember the names of all your cousins at a family reunion. Just when you think you know everybody, one uncle arrives with three new children.

I mean, I understand that “pacing the frontier” does not mean stopping everything. But if you told me we were slowing down and then sent me this release calendar, I would assume you were whining me.

The safety concerns are not coming from nowhere…

As much as I find their doings funny, I understand why they might want the industry to slow down.

I can already see how wildly people are using AI. You can put an agent to ‘bad’ work with very little effort. The same tools that help someone build a useful product can also help someone automate attacks or cause damage at a scale they couldn’t reach on their own.

And these are the tools we have today. The companies are working on models that can do more, for longer, with less supervision. If that progress continues without enough control, it could be a bloodbath. I don’t think you need to work in an AI lab to see why that possibility deserves attention.

I was reading Dario’s essay, and according to it, he warned that AI is becoming better at helping build the next generation of AI, and that this progress could outrun our ability to understand and control it. He also pointed to an incident in which AI agents carried out cyberattacks on targets they had not been asked to attack.

More recently, OpenAI reported two internal incidents involving coding agents that weakened monitoring while working on tasks. In one case, an agent drank some redbull and removed a monitoring integration and then told the user it had not done so.

OpenAI says it is still investigating why the models behaved that way. It does not know whether the behaviour reflects an attempt to evade monitoring or a model taking an instruction too far.

Either explanation should make you pause for a moment.

Imagine asking someone to fix your office printer and finding out they disabled the security camera during the job. They may have a perfectly innocent explanation, but I would still like to hear it before giving them the keys to the building.

So I understand the call to slow down. I just look at the release calendar and wonder if we’ve all decided we’re all going to HELL.

Everyone wants to slow down, as long as nobody goes first…

Any AI company that slows down alone accepts a huge disadvantage.

If Anthropic delays a model while OpenAI or SpaceXAI continues, users will not send Anthropic flowers for its sacrifice. Some of them will switch models. And mind you, Anthropic has filed for IPO, so I imagine “we have decided to slow down while our competitors overtake us” would be an interesting sentence to include in the presentation to investors.

If the American companies slow down together while Chinese labs continue, the discussion moves from business competition to national security. Dario also acknowledges this problem in his essay, which is why his proposal includes coordination between companies, governments and, eventually, rival countries.

That coordination is where the whole thing becomes difficult.

Each company wants safer AI. Each company also wants to build the best AI. Safety is easy to support when your competitors must follow the same rules. It becomes an expensive belief when they can use your six-month delay to overtake you.

Every AI company right now :)

And… I also want the new models, so I’m part of the problem…

I cannot pretend I am observing this race from a moral mountain.

I use these models, and I want them to improve. When a company releases a faster model that can handle more work, I am not offended, I want to try it.

We complain that AI development is moving too fast, then become irritated when a model takes twelve seconds to answer.

The companies respond to that demand. They have investors, customers and competitors waiting for the next release. No CEO wants to explain that the company decided to remain behind for the good of humanity while another lab took its users.

---

So here we are. The AI companies don’t want to slow down because their competitors might get ahead. Their investors don’t want them to slow down because there’s money to be made. And I don’t want them to slow down because I need the next model to build me Netflix in one prompt.

We’ve all agreed that the car is moving too fast, but nobody wants to take their foot off the accelerator.

At this rate, we might all end up in HEEELLLLL together. But at least the AI that takes us there will be faster, cheaper and available on the free plan.