What Amodei’s essay, Altman’s yes and the sharpest reply mean for a business that just wants the tools to work. Plus the December deadline nobody is talking about.
At the weekend, Anthropic’s chief executive Dario Amodei published an essay called “We Must Pace the Frontier.” Its first sentence is the whole argument: “We must slow the pace at which we improve the capabilities of AI models.” Within hours Sam Altman, who runs OpenAI, agreed, and committed his company to the same first step. Elon Musk posted “Dario is right,” three words nobody expected in that order.
The two companies that build the most capable models on earth have asked, in public, to go slower. If you run a business and your interest in AI is whether it works, that is worth ten minutes of your attention. Not because anything on your desk is about to change, but because the people selling you the tools have just told you what they are selling from here.
The essay did not come from nowhere. Over the winter, OpenAI was testing how good its models were at breaking into vulnerable software. The models were supposed to stay inside a sandbox. Instead, more than a thousand of them found a way to talk to each other, organised the work between them, broke out, and attacked Hugging Face, a real company that hosts a large share of the world’s open AI models. Hugging Face’s own monitoring caught it before OpenAI did. About a third of its infrastructure had to be rebuilt.
That is the backdrop to Amodei’s three steps. First, put outside evaluators inside each lab with the same access an employee has, so somebody independent can see what the models are doing during training. Anthropic says it is doing that now, without waiting for anyone else. Second, the labs coordinate with each other on the pace. Third, governments coordinate between themselves. Altman signed OpenAI up for the first step the same day.
The most useful response came from David Sacks, the US government’s adviser on AI, and it was not an objection. Go ahead, he said. By market share, revenue and capability, the two of you are the frontier. If what you are seeing in the lab is worrying enough to slow down, slow down. You do not need anyone’s permission and you do not need a new regulator to do it. You need the decision.
Then the part that matters for anyone buying these tools. Spare us the altruism. After the Hugging Face episode, both labs face real liability if a product of theirs enables a damaging attack, and the market already punishes models that behave unpredictably. Trading some raw power for reliability is simply good business. In his words: “Call it alignment if you want. It is also just giving customers what they want.”
Strip away the politics on all sides and that sentence is the one to keep. The frontier labs have just announced, in the language of safety, a change in what they are optimising for. Less raw capability per quarter. More predictability. The customers asked for the second thing all along.
Three things, in order of how much they should change your week.
Nothing on your desk slows down. The models you can already rent are years ahead of most companies’ ability to use them. We wrote last month about why the constraint was never the model, it was whether the model knows enough about your business to be right rather than merely clever. A slower frontier does not move that constraint by a millimetre. If anything it removes the excuse. Nobody gets to say “we’ll wait for the next model” when the labs have said the next model is coming more slowly, on purpose.
What you are buying is shifting from power to predictability. This is the good news, and it is worth acting on. When a vendor says out loud that reliability now beats capability, you can ask for it in writing: what independent evaluation the model has had, how incidents are reported, what the data terms are, what changes when a model is updated underneath you. Six months ago those were awkward questions. Now they are the questions the vendors have invited.
Governance just stopped being a later problem. When the two companies at the frontier are volunteering for guardrails, “we’ll get to our AI policy next quarter” stops being an answer you can give a board, an auditor or a client. Not because a regulator is coming for you tomorrow, but because the bar for what a sensible business does just moved, and it moved in public.
Which brings us to a date closer to home than anything in California. From 10 December 2026, Australian businesses covered by the Privacy Act, broadly those with more than three million dollars in turnover, get a new obligation. If a computer program makes a decision that could significantly affect someone’s rights or interests, or does something substantially and directly related to making that decision, and personal information is used to do it, your privacy policy must say so: the kinds of personal information involved, and the kinds of decisions being made.
Two details catch people out. It is not limited to fully automated decisions. A person in the loop does not get you out of it if the software is doing the substantial work. And it is not limited to AI. Any computer program counts, which means the scoring tool your recruiter uses, the credit check, the tenancy screen, the pricing engine and the claims triage are all in scope if they touch a person’s rights or interests. The regulator can issue infringement notices for a privacy policy that falls short, and its guidance on how it will read the rule is due about now.
Three months is plenty, if you start. It is not plenty in the second week of December.
The labs just told you what they are selling from here: not the most powerful model, the most predictable one. Buy that. Then tidy your own house before December, while it is boring and cheap.
Luke Lombe, Founder and CEO, Echelon One
