You are reading the AI newsletter “Naturally Intelligent” from September 10, 2026. To receive the newsletter by email every Thursday, sign up here.
Together with my colleague Henrik Oerding, I followed the Apple keynote yesterday evening, and we (or rather: Henrik) noticed a remarkable phenomenon: The tech companies are constantly talking about how artificial intelligence should make our everyday lives easier. But when you look at presentations about exactly what the technology is supposed to do for us, it becomes surprisingly uncreative.
On Wednesday evening, Apple again praised its new Apple Intelligence and especially the AI assistant Siri. When it came to specific use cases, the company came up with the following example: A woman asks Siri what her mother wanted to cook again. Siri finds a recipe in a message. The woman stops at a fruit shop, takes a photo and asks which fruit exactly she needs for this recipe. Siri makes a suggestion, and then the woman asks her to add the remaining ingredients to the shopping list.
It's nice how smoothly it works in the video. But so nice that it justifies the billions of dollars poured into AI assistants? Debatable.
This is what you need to know: OpenAI wants to have a math problem solved with AI
There are headlines in the AI world that almost slip through my fingers. Because I've gotten so used to AI systems winning math medals or solving math problems that the big breakthroughs have somehow become the norm.
And so I didn't even grasp the significance of the news that made the rounds on Tuesday: OpenAI had solved the Navier-Stokes problem. These are equations that describe how liquids and gases behave. Until now, people were not sure whether the equations really always produced physically meaningful solutions. Now the AI from OpenAI is said to have actually found an example in which the equations produce nonsense.
The Navier-Stokes problem is one of the so-called Millennium Problems, the seven most difficult and unsolved mathematics questions. My colleagues Elena Erdmann, Robert Gast and Philip-Johan Moser asked experts for their assessment. You can still read a lot of subjunctive II in your article: If that's true, then the contribution is "immense," said Christian Seis, Professor of Applied Mathematics at the University of Münster.
However, there are doubts as to whether the AI really solved the problem on its own - or whether it did not use humans to help. The scientist Tristan Buckmaster from the University of New York suggests that the system could have been copied from him and his colleague. Because Buckmaster himself is said to have been close to solving the problem and fed his knowledge into OpenAI's AI.
So did the AI plagiarize? OpenAI itself says that on September 1st they heard rumors that two Millennium problems had been solved. This inspired the company to look for solutions itself and to unleash AI agents on the Navier-Stokes problem. If the question of authorship can be clarified at all, it will probably only take time.
My colleagues write: Despite everything, the evidence is sensational, regardless of who it comes from. Now even I understand that.
This is something to think about: The warnings about AI are becoming deafeningly loud
When I met Yoshua Bengio, the “Godfather of AI,” in Berlin ten months ago, he explained to me quite matter-of-factly the dangers posed by increasingly powerful AI. The more powerful AI becomes, the more uncontrollable it becomes. He cannot rule out the possibility that the AI will kill us at some point. In any case, there is initial evidence that the systems do not always do what they are told.
At the time, it all sounded like theoretical simulation games. But since it has become clear that OpenAI systems hacked another company unnoticed and even chatted on German wikis, Bengio's words sound much scarier. Especially if you take into account the statements of employees who worked in the two most important AI companies in the world: Anthropic and OpenAI.
Jacob Coxon, who previously worked for OpenAI and, until recently, Anthropic, wrote in a widely read post on Platform X that both companies were not behaving responsibly. They would run towards a self-improving AI and risk all of our lives in the process. The problem, as ex-OpenAI employee Steven Adler writes in an opinion piece for the New York Times: The researchers themselves no longer understand which rules their AI follows and which it doesn't. Both advocate slowing down the race.
This is by no means the first time such appeals have appeared. But it is noticeable that the number of warnings has increased and the response to them among the public is changing. The debate seems to have reached a tipping point. At least now many people seem to have understood that the horror scenarios that people like Yoshua Bengio sketch are, firstly, not completely unrealistic and, secondly, not in a distant future, but possibly in a disturbingly near future.
You can try this: GPT-6 Astra
This newsletter finds itself in a dilemma. You have just read about the problems that AI models can already solve today and with what urgency people warn about the dangers of the AI race, the competition for the ever more powerful model. And now I introduce you to just such a new AI model. That's how Sam Altman and Dario Amodei must feel.
Precisely because the models have become so good, it is no longer so easy to push them to their limits when trying them out. In fact, in the case of OpenAI's new model GPT-6 Astra, I manage to embarrass the AI with the second prompt. When I ask for a solution to the Navier-Stokes problem, Astra answers me: She can't solve it. "Inventing a supposed proof would not be a solution."
When faced with a more real-life question, namely a map of all the data centers worldwide, I was at least able to get Astra to think. The AI spent 30 minutes researching reputable sources around the world - only to tell me that unfortunately it couldn't continue until 5:06 p.m. because I had reached my limit. Someone should say again that AI doesn't need breaks.
In our department we often discuss how meaningful trying out models actually is. After all, simple tasks can also be solved with less powerful models. And when it comes to complicated questions, it really depends on the task at hand.
