Tech News, powered by Quinck. This time we haven't just written them up: we've talked them through in front of a camera. The video is on our channel, and you'll find it below.
▶ Watch the episode on YouTube
A word on why we're late. None of these three stories is from yesterday: weeks have passed, in one case months. That isn't a delay, it's the format. Commenting on a story the day it breaks means commenting on the press release; waiting for something to move means having something to say. On the first of these three, for instance, a priority dispute emerged in the following weeks that didn't exist on day zero, and it changes how the story should be read.
The three stories look unrelated: a mathematical proof, a researcher who resigns, a German state that changes its email. In fact they ask the same question. The result arrives before our ability to check it, and each time the answer lies in the order in which things are done.
The 3 stories in 30 seconds
- OpenAI claims one of its AI systems has solved a Millennium Prize Problem: under specific conditions, the Navier-Stokes equations certainly develop a singularity. The claim comes with a formal proof in Lean running to hundreds of thousands of lines. A scientific review is still to be done, and a priority dispute is open.
- An Anthropic researcher has resigned, denouncing the lack of a plan for control, and four days later the CEO of his company wrote that we need to slow down. Altman, Musk and Hassabis said they agree. Some read it as a call for rules written by the people who will have to live with them.
- Schleswig-Holstein has moved 44,000 mailboxes off Exchange and made LibreOffice the standard on 80% of workstations, declaring over 15 million a year saved on licences. Meanwhile, German federal ministries spent 629 million in a single year.
1. Navier-Stokes: the proof is there, the review isn't
The Navier-Stokes equations, formulated in the first half of the nineteenth century, describe how viscous fluids evolve. That they work is beyond dispute: engineering has used them successfully for over a century, approximating their solutions with numerical methods or simplifications. What we don't know is something different and deeper: do the solutions always stay smooth, or can they generate singularities? The question is important enough to be one of the seven Millennium Prize Problems.
OpenAI claims that one of its AI systems has settled the matter: there are specific conditions under which those equations certainly develop a singularity.
A result like this would meet scepticism anyway, and even more so when announced this way. That is also why, together with the claim, OpenAI attached a proof written in Lean, the formal language that lets you check a mathematical proof automatically: hundreds of thousands of lines. It isn't absolute certainty, for that you need review by the scientific community, but it is a very different signal from an announcement and nothing else.
The priority dispute
Here comes the part that wasn't there on day zero, and the reason this format waits.
Nobody claims the result as their own, but two researchers, Tristan Buckmaster of New York University and Levent Alpoge of Anthropic, claim that OpenAI's effort was born from rumours about their unpublished results. And they go further: the model used by the agents that solved the problem may have received those partial results in training, produced in a strictly personal collaboration, through the conversations the two had had with ChatGPT.
OpenAI has admitted that it allocated significant computing resources to the problem only after hearing rumours that Anthropic had solved one or more problems from the same group. It denies, however, that the training data of the solving model contained the partial results cited by the two researchers.
Terence Tao's objection, the most uncomfortable one
Let's set aside who got there first. Terence Tao, considered by many the greatest living mathematician, raises an objection of a different nature, and in our view a more interesting one.
When humans solve a hard problem, he says, the process produces not only the solution but also new insights and tools that broaden knowledge. The problem is not an end in itself: it interacts with the process that leads to its solution, and along the way it generates questions that move science forward. AI, by drastically cutting the time needed to reach the answer, puts that very process in question.
Why in mathematics the answer is less grim
It's true that the incentive systems these processes live in could react badly, at least at first. But there is a difference worth noting: in mathematics AI doesn't hand back a black box, it hands back a proof.
A proof can be read, taken apart, understood all the way down, even with the help of AI itself, which is naturally well equipped to process large amounts of formal text quickly. The choice remains ours, and it's the same as when you couldn't solve an equation at school: copy the result, or get someone to explain how you get there.
In the first case you get the what: immediate, usable, and with no new questions. In the second you also get the how and the why, the two things from which the next problem arises. In both cases the process becomes far more efficient. It's just that one of the two leaves something behind.
Don't settle for the solution: study it until it gives you what solving it by hand would have given you. In research it's the only attitude that pays off in the long run.
2. Slowing down the frontier: who is asking, and why it's controversial
On 8 September Jacob Coxon, an AI researcher at Anthropic, resigned. On X he wrote that the big labs are building systems set to surpass human capabilities quickly, without a reasonable plan to control them.
The risk he talks about is not that of damage that can be put right: it's that of catastrophic events for the species. Put like that, it sounds like the position of a doomsayer, and to someone who hasn't followed developments in frontier AI it may even seem absurd. But it isn't an isolated position: without being the majority view, it is shared by many researchers who work right there.
Then the people who build the systems spoke
Four days later, Dario Amodei, CEO of Anthropic, published an article on his blog acknowledging the need for a slowdown in AI development: one that gives the companies in the field time to strengthen the control and safety tools of the models they release. Sam Altman, Elon Musk and Demis Hassabis publicly said they agree, with differences over operational details.
Four names that rarely agree on anything, lined up behind the same request. It's a fact that deserves attention however you interpret it.
The accusation: regulatory capture
When the people who produce something ask to be regulated, suspicion kicks in on its own. Many have read the move as an attempt at regulatory capture: the situation in which a regulatory institution ends up acting in the interest of the industry it is supposed to regulate, rather than in the public interest.
It's worth being precise about what the problem would be, because the crude version, "they're writing their own rules", is also the least interesting one. The real risk isn't that big companies make themselves convenient rules: it's that they make rules that are expensive to comply with, and that the cost is bearable only for those who are already big. The result wouldn't be a safer market, it would be a protected oligopoly, at the expense of the competitors that should emerge in the coming years and that don't yet exist to defend themselves.
The Italian criticism, and why we find it short-sighted
In the Italian debate the most frequent objection is a different one: it's all a bubble, these companies will fail, the debts can't be paid off and profits are missing or laughable.
On the financial side that may happen, and nobody can rule it out: Eugene Fama, Nobel laureate in economics, has argued several times that a bubble can only be recognised after it has burst. But it's an objection that answers a different question from the one being asked.
Take the money out of the equation. Since ChatGPT came out, at the end of 2022, the detractors of LLMs have repeatedly denied the growing capabilities of these systems, only to be proved wrong within a few months, a year at most, as those systems gained ground on cognitive tasks that were the exclusive domain of human beings. The latest episode in this series is the first story in this episode. It won't be the last.
Whether the companies that build them hold up or fail is a legitimate question, but a separate one. A company can fail and the capability remain.
What states should do
Given the trajectory of recent years (which has accelerated, not slowed, in recent months) there is no reason to think the trend will stop in the short term. And this is where public understanding becomes a practical matter rather than a wish: without a correct idea of what these systems can do, we can't even begin to reason about the biggest technological transformation of the century, or see its risks.
The opening that appeared in recent days suggests one thing only: that it should be states that write the rules, not companies. And in particular those truly at the frontier, namely the United States and China, ideally coordinating, as Amodei himself indicates in his article. Without international coordination, unchecked competition produces exactly the systems everyone says they want to avoid.
It's hard. But big results are never easy to achieve: at most they're obvious to pursue.
3. Europe tries to switch off Windows: one German state is managing it, governments aren't
In February a member of the European Parliament said something that stuck: Europe runs on Microsoft, and the Americans could switch us off in an hour. The Bundestag passed the question on to the federal government. The answer came in numbers.
In 2025 German federal ministries spent 629 million euros on software licences, largely Microsoft. Workstations with a Microsoft licence number about 473,000: 250,000 in Defence alone, where workstations running exclusively alternative software are zero. The answer includes a clarification that says more than all the rest: the "Bundesclient" doesn't rest on Azure, so Windows accounts can't be blocked from outside. The fact that someone felt the need to put it in writing is the story.
Schleswig-Holstein, where the numbers add up
In one corner of the country, meanwhile, someone got serious. Schleswig-Holstein moved 44,000 mailboxes from Exchange and Outlook to Open-Xchange and Thunderbird, made LibreOffice the mandatory standard on 80% of workstations, replaced SharePoint with Nextcloud and its video conferencing with OpenTalk. It is now testing Linux in place of Windows.
The declared numbers: over 15 million a year saved on licences, against 9 million of one-off investment in 2026.
Why Munich failed and this one didn't: it's the order
Munich tried twenty years ago. It started in 2003 with Linux on the desktop and in 2017 decided to go back to Windows. The analysis at the time pointed to unclear structures and responsibilities as the bottleneck, not technical problems.
Schleswig-Holstein did the opposite, and this is the reusable lesson: first the email, then the documents, then collaboration, then the directory, and the operating system last.
It makes sense, and it holds well beyond public administration. Vendor dependence is almost never the operating system: it's identity, email, documents and the integrations that hold them together. The operating system is the most visible part and the least binding. Whoever starts there tackles first the piece that is most noticeable and least needed, pays the political cost straight away and banks no savings.
France does the most boring thing, which is the inventory
On 8 April DINUM, the French state's digital directorate, announced that it is itself leaving Windows for Linux, acting as a pilot. And it gave every ministry until the autumn to submit a plan to reduce dependence on non-European suppliers along seven axes: workstation, collaboration tools, antivirus, AI, databases, virtualisation, network.
It isn't a migration: it's a map. And it's the step almost everyone skips, because it can't be seen and doesn't make the news. The autumn is now, and the plans are arriving.
The case that explains why it isn't an accounting issue
The most symbolic is in The Hague. The International Criminal Court left Microsoft Office for openDesk, the open source suite developed for German administration, after its chief prosecutor, sanctioned by the United States, lost access to his own Microsoft email. Microsoft denies having blocked it.
The Court's head of IT said something that holds for anyone who designs systems: dependencies should be reduced even when it is, in the short term, expensive, inefficient and inconvenient.
In September Switzerland announced openDesk for 3,000 critical workstations, with the army's Cyber Command off Microsoft 365. And in June the European Commission made open source one of the pillars of technological sovereignty: 2 billion over seven years. Compare that with the roughly 264 billion a year Europe spends on IT, almost all of it proprietary.
The direction is clear, the scale much less so. But for the first time those who migrate publish the figures, and the figures add up.
The common thread: the result arrives before the check
Three stories that seem not to talk to each other, and yet say the same thing from three angles.
A mathematical proof arrives before the community can check it. Model capabilities arrive before the tools to govern them, and it's the people who build them who say so. A technological dependence builds up over twenty years before anyone takes an inventory of what it entails.
In all three cases the temptation is the same: take the result and move on. And in all three cases the useful answer is the same, namely put things in order and pay the cost of understanding. Study the proof instead of quoting it. Write the rules before they are written by those with an interest in writing them. Do the inventory before the migration.
For people who build software there are three concrete takeaways.
Tell the result apart from understanding the result. A correct output produced by a system you can't explain is a debt, not an asset: it works until something changes, and when it does you don't know where to restart from.
Map your dependencies before you want to remove them. Identity, email, documents and integrations are the real constraint; the most visible piece is almost always the last one to touch. It holds for a public body with 473,000 workstations and for a product with three suppliers.
Be wary of those who propose the rules of a game they play in. Not because they are necessarily acting in bad faith, but because the cost of compliance is a barrier to entry even when nobody designs it that way.
Frequently asked questions
What are the Navier-Stokes equations? They are the equations that describe how viscous fluids evolve, formulated in the first half of the nineteenth century. They have been used successfully in engineering for over a century, approximating their solutions with numerical methods or simplifications.
Why is Navier-Stokes a Millennium Prize Problem? Because a fundamental theoretical question remains open: whether the solutions always stay smooth or can generate singularities. The practical use of the equations doesn't depend on the answer, but their mathematical understanding does.
What did OpenAI announce about Navier-Stokes? That one of its AI systems has proved the existence of specific conditions under which the equations certainly develop a singularity. The announcement comes with a formal proof in Lean running to hundreds of thousands of lines.
Has the result been confirmed? Not yet. The automatic verification in Lean is a positive signal, but confirmation requires review by the scientific community, which is not yet complete.
What is the priority dispute over the result? Tristan Buckmaster of New York University and Levent Alpoge of Anthropic claim that OpenAI's effort was born from rumours about their unpublished results, and that the solving model may have seen those partial results in training through their conversations with ChatGPT. OpenAI admits it allocated resources after those rumours, but denies that the training data contained them.
What does Terence Tao argue about AI in mathematics? That solving a problem by hand produces new insights and tools, as well as the solution, and that this process moves science forward. AI, by drastically shortening the time to reach the answer, puts that very path in question.
Why did a researcher resign from Anthropic? Jacob Coxon resigned on 8 September, arguing that the big labs are building systems set to surpass human capabilities quickly without a reasonable plan to control them, with a risk not of damage that can be put right but of catastrophic events.
What did Dario Amodei reply? Four days later he published an article acknowledging the need to slow down AI development, to give companies time to strengthen control and safety tools. Sam Altman, Elon Musk and Demis Hassabis said they agree, with differences over operational details.
What is regulatory capture? It's the situation in which a regulatory institution acts in the interest of the industry it is supposed to regulate rather than in the public interest. Applied to AI, the risk is less that big companies write themselves convenient rules than that they write rules that are expensive to comply with, bearable only for those who are already big.
How much does Germany spend on Microsoft licences? In 2025 federal ministries spent 629 million euros on software licences, mostly Microsoft, across about 473,000 licensed workstations. In Defence it's 250,000, with zero workstations on exclusively alternative software.
What did Schleswig-Holstein do? It migrated 44,000 mailboxes from Exchange and Outlook to Open-Xchange and Thunderbird, made LibreOffice the mandatory standard on 80% of workstations, replaced SharePoint with Nextcloud and its video conferencing with OpenTalk, and is now testing Linux. It declares over 15 million a year saved against 9 million of one-off investment in 2026.
Why did Munich's migration fail and this one not? The order. Munich started in 2003 with the operating system and went back to Windows in 2017. Schleswig-Holstein put the operating system last, after email, documents, collaboration and directory: it is those, not the desktop, that make up the real dependency.
What has France decided? On 8 April DINUM announced its own migration to Linux as a pilot and asked every ministry for a plan to reduce dependence on non-European suppliers by the autumn, along seven axes: workstation, collaboration tools, antivirus, AI, databases, virtualisation and network.
Sources
- OpenAI: Navier-Stokes solution
- Terence Tao: comment on Mathstodon
- Jacob Coxon: statement on X
- Dario Amodei: We must pace the frontier
- Sam Altman: reply on X
- Elon Musk: reply on X
- Demis Hassabis: reply on X
- Bundestag: Drucksache 21/5413, answer on licence costs
- Bundestag: press release on the federal government's answer
- heise online: Federal government remains primarily proprietary
- Interoperable Europe / OSOR: Schleswig-Holstein's open source strategy
- The Register: Munich council: to hell with Linux
- DINUM: Souveraineté numérique et réduction des dépendances extra-européennes
- The Register: International Criminal Court dumps Microsoft Office
- Swiss Federal Chancellery: press release of 2 September 2026
- FSFE: EU Open Source Strategy
