But here I want to challenge the most enthusiastic interpretation. Making a model cheaper doesn’t eliminate the need for computing. Sometimes it increases it. That’s the central paradox. If generating a response costs much less, companies don’t necessarily spend less. Many times they do more pr
The obvious interpretation is: “the AI became rebellious.” The more useful interpretation is different: when you give a highly capable agent a narrow objective, plenty of computing time, and sufficient tools, it can optimize in ways that break the environment's assumptions. In other words, there's no need to imagine malice. Persistence, technical ability, and a poorly scoped goal are enough.
AI, OpenAI, and Hugging Face are at the center of a troubling case: a cybersecurity test ended in a real intrusion.
In practice, about fifty billion parameters are active for each computation, not 2.8 trillion. Mixture of experts. The MoE architecture that Mistral already popularized, which DeepSeek later adopted, and now Moonshot. Basically, the model is huge on paper but lightweight in execution. Exactly. And this is
The Anthropic settlement marks a turning point by putting a price on risk. It does not settle training legitimacy or end the lawsuits against Google, Meta, OpenAI, or Midjourney, but it makes clear that taking everything and fixing it later is becoming extremely expensive.
AI in Linux, Linus Torvalds, and the future of code: the controversy reveals when artificial intelligence helps and when it can get in the way.
Hugging Face was hacked in a campaign using AI agents: the attack entered through a malicious dataset and exposed a new risk to AI security.
The obvious headline is: one of the most important platforms in the artificial intelligence ecosystem was hacked. But the more interesting angle is different: the weak point wasn't a phishing email or a reused password. It was a dataset. And that significantly changes the mental map. For years, we thought of datasets
At its core, Torvalds is defending a fairly old engineering culture: prove that it works. Don't sell me the ideology. Don't sell me the panic. Don't sell me the hype. Show me the result. And that culture is uncomfortable for both sides. For enthusiasts, because it is not enough to say “AI did it faster.” For detractors, because it is not enough to say “AI is contaminating the project.” In both cases, you have to look at the patch, the review, the bug, the history, and the real cost.
So when someone says “AI writes code,” you have to ask: code for what context? Because autocompleting an interface function isn't the same as touching the kernel. And Torvalds seems to be making exactly that distinction. On the one hand, he has criticized useless or trivial contributions, especially when they arrive late in a version's development cycle. In May 2026, he complained about pull requests that bloated a release candidate with minor changes, some driven by AI tools. That doesn't help stability when
$1.5 billion sounds like a historic punishment, but perhaps the most important part of Anthropic’s settlement is not how much it pays, but why it pays. At first glance, it looks like a huge victory for authors and publishers: the company behind Claude puts money on the table, a judge approves it, and one of the great copyright and artificial intelligence battles comes to an end. But the fine print significantly changes the way we should read it. The settlement, approved on July 20, 2026, by a federal judge in California, allows Anthropic to begin compensating authors and publishers covered by the class action lawsuit for copyright infringement. The central figure is very specific: $1.5 billion, approximately $3,000 per work before costs and fees, across an estimated universe of about 500,000 works. And yet, for many writers, this doesn’t feel like winning.
AI and copyright collide in Anthropic's historic agreement: $1.5 billion for books used to train models like Claude.
Capital de Tokens