Imagine three racing drivers lined up on the grid.
All of them are driving cars that get faster the longer they drive. All want to win. Until now, none has shown much interest in slowing down.
Then they turn to the stands at the same time and say:
You do not start discussing the colour of the seats then.
You ask: what is around the next corner?
That is exactly how I feel right now about statements coming from the circles around OpenAI, Anthropic and Elon Musk, or xAI.
A former researcher who worked at OpenAI and Anthropic is resigning publicly and accusing both companies of racing towards self-improving superintelligence without acting responsibly. Dario Amodei says the pace at the model frontier needs to slow. Sam Altman agrees publicly. And Elon Musk compares the risks of AGI to nuclear weapons, only greater.
These are not identical statements. There was no joint press conference. There is no evidence that a server room has just been locked somewhere in a lab.
But the consonance is remarkable.
After all, these people make their money by making the systems more capable. More capabilities, more users, more data centres, more capital. Until now, the brake has rarely been the product advertised most loudly.
And suddenly, they are talking about control.
Or perhaps AI simply analysed the state of the world, looked at the distribution of power and wealth, and concluded:
A very good joke, I think.
But it only works because there is genuine uncertainty underneath it. When people who are closer to the machines than we are suddenly talk about losing control, you inevitably wonder: what do they know that we do not?
We do not know.
It could just as easily be that the warnings also serve very tangible business interests. Anyone who says that only billion-dollar labs can responsibly develop a particularly dangerous model is building a rather efficient fence around their own property.
Safety and market power are not mutually exclusive, unfortunately.
Then there is the unromantic explanation: data centres, chips and electricity have become so expensive that every new stage of development is a bet with a great many zeros. “We are slowing down out of responsibility” sounds more elegant to investors than “The next improvement costs us a small European economy.”
So I am less interested in which theory sounds best.
I am interested in what follows from it in practice.
If a system is allowed to work independently, it needs boundaries. Clear access rights. Logs. People who approve real decisions. A switch that actually works. We should simply stop giving systems ever more room to act, then acting surprised when control becomes an issue.
Perhaps AI never said “Eat the Rich.”
But when the people running the largest AI labs all start talking about brakes, we should become extremely alert by then at the latest.