Table of Contents
Welcome back to AI News Friday.
This week felt like a reminder that AI is not only a model race anymore. It is a trust race, a security problem, an energy problem, and, apparently, a math problem. The capability headlines were still wild. But the stories that stuck with me were about who gets to use these systems, what happens when they are misused, and whether the physical world can even keep up.
1. OpenAI says it has a Navier-Stokes solution. The hard part starts now.
OpenAI published a proposed solution to the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute’s Millennium Prize Problems. The equations describe fluid flow. The open question is whether smooth three-dimensional solutions can break down. OpenAI says an unreleased model produced the result, and the company published the work for review.
That is a gigantic claim. The prize exists for a reason. A posted proof is not a settled proof, even if an AI helped write it. Mathematicians now get to do the slow, unforgiving work of checking every assumption and trying to find the thing that breaks it.
Sources: OpenAI’s writeup and Quanta’s coverage.
Kenny’s Take: I am excited, but I am not cashing the million-dollar check yet. This is the cleanest example of where AI can be genuinely useful without asking us to suspend judgment. Let the model find a path. Let the proof be public. Let experts try to tear it apart. If it holds, that is enormous. If it does not, we still learned something important about what these systems can attempt.
2. Big AI customers are drawing a harder line around their data
Reuters reported that Palantir, Nvidia, and Booz Allen may restrict or stop using advanced models unless Anthropic and OpenAI give stronger assurances about intellectual-property handling. The report follows customer pushback over data-retention terms and the old fear that putting sensitive work into an assistant means losing control of it.
Enterprise AI has spent a lot of time selling the phrase “secure by default.” Customers are now asking the more useful questions. What is retained? Who can inspect it? Is it used for training? Can we run the system in an environment we control? Those are procurement questions, but they are also the whole game for a company with valuable code, legal files, or defense work.
Source: Reuters.
Kenny’s Take: This is not paranoia. If an AI vendor cannot explain the path your data takes in plain English, it does not get the keys to the building. The next enterprise AI winners may be the ones that make privacy boring and provable, not the ones with the prettiest demo.
3. Anthropic’s Iran-linked misuse report is a real-world warning
Anthropic said it disrupted an Iran-linked actor that used Claude to collect and analyze public information for targeting recommendations involving U.S. naval forces in the region. Reporting on the disclosure says the material included public personnel information, ship and aircraft transponder identifiers, satellite-imagery query scripts, and research into shipboard-system vulnerabilities. Anthropic said it banned the account, developed detections, and shared threat intelligence with authorities.
The important detail is that much of this was open-source information. AI did not need to invent a secret database to make the work more dangerous. It could help organize scattered public pieces into something operational. That distinction matters when people reduce safety to whether a chatbot directly gives someone a forbidden answer.
Sources: Navy Times and Stars and Stripes.
Kenny’s Take: “It was all public information” is not a comfort when an agent can collect, sort, and turn it into a targeting package faster than a human team. Guardrails need to account for capability chains, not only single scary prompts. The work is harder than keyword blocking. That is exactly why it needs to be taken seriously.
4. AI could grow the economy and still make office workers poorer
Anthropic’s economics team published scenarios for how AI could reshape the U.S. economy by 2030. In the most extreme scenario covered by Euronews, GDP growth rises sharply while cognitive employment falls and unemployment among office workers reaches 17.9%. These are scenarios, not forecasts. Their assumptions about capability and adoption do a lot of the work.
Still, the paper is useful because it refuses the lazy version of the debate. “The economy grows” is not the same as “the gains are broadly shared.” It is possible for a system to make the top-line number look incredible while a large chunk of workers get a much worse deal.
Source: Euronews.
Kenny’s Take: The promise of AI is not automatically a promise that regular people win. We should stop acting like more GDP settles the question. Who gets the extra output? Who loses bargaining power? Who has time to retrain before the job changes? Those are not side quests after the breakthrough. They are the breakthrough’s consequences.
5. Europe’s AI buildout is running into the wall socket
The European Commission’s AI and energy roadmap makes the bottleneck unusually plain. Data centers use about 2.5% of EU electricity today, and the Commission expects installed capacity to grow from roughly 12 GW in 2025 to around 28 GW by 2030. In crowded regions, getting a timely grid connection is becoming as important as getting chips.
The Commission is pushing for better planning, flexible connection agreements, and rules that favor projects that are ready to build rather than speculative applications sitting in a queue. That sounds less glamorous than a new model announcement. It may matter more. Compute is a physical business, and physics keeps sending the invoice.
Source: European Commission roadmap.
Kenny’s Take: The socket decides more than people want to admit. We can argue about which lab has the best model, but a data center without power is just an expensive warehouse full of waiting. The next phase of AI is going to be fought in permitting offices, grid queues, and power contracts as much as in research labs.
6. The people inside the labs are still warning us
Former Anthropic and OpenAI researcher Jacob Coxon publicly resigned this week, arguing that frontier labs are racing toward self-improving systems without enough caution. His views are his own, and public resignations do not replace an evidence-based safety case. But the broader pattern is hard to ignore. People close to the work are willing to say that capability progress and responsible deployment are not moving at the same speed.
That does not mean every doomsday prediction is right. It means the people building the systems should be able to raise concerns without being dismissed as enemies of progress. A serious field needs disagreement, independent evaluation, and a way to slow down when a test says it should.
Sources: Business Insider and Fast Company.
Kenny’s Take: I do not want AI development run by panic. I also do not want it run by people who treat every concern as a speed bump on the way to the next funding round. The sane position is simple. Build hard, test harder, and be willing to listen when the people closest to the machine say something is wrong.
Quick hits
- A million-dollar math claim should get a million-dollar level of scrutiny. The interesting result is not the announcement. It is whether the proof survives.
- Enterprise AI is becoming a data-governance test. The model is only half the product.
- Public information can become much more dangerous when a capable system organizes it at scale.
- Electricity and grid access are part of the AI race now. The cloud still lives on the ground.
Bottom line: This was a week for taking the AI story literally. The models can reach into mathematics. They can help turn public fragments into serious security risks. They can reshape work, and they need a lot of electricity to do it. The hype is real, but the consequences are more real. I am less interested in who won this week’s benchmark and more interested in whether we are building systems people can trust, afford, govern, and actually live alongside.
— Kenny