Anthropic under pressure, LAION releases video benchmark

Today once again shows very clearly that AI is currently moving on two tracks at the same time: up front, the big product and research machine; behind it, the ever-growing legal trailer. While Anthropic is facing copyright lawsuits, LAION and research continue to deliver material for open models, benchmarks, and real-world tests for AI agents.

For you, that means less “What can the model do?” and more “Under what conditions is it allowed to, should it, and can it even do that?”. It’s exactly this mix that makes the current situation exciting — and a little uncomfortable. But regulation has never been comfortable.

🎵 Sony and Warner take Anthropic to court

Sony Music, Warner Music, and other rights holders are suing Anthropic over alleged copyright violations in the training of Claude. According to the lawsuit, the company is said to have used tens of thousands of protected musical compositions; particularly notable: CEO Dario Amodei is also named personally. The plaintiffs speak of “one of the largest thefts of intellectual property in history.” Legally, that’s obviously the full heavyweight hammer, but above all it shows one thing: the music industry no longer wants to talk about individual cases, but about the principle behind LLM training. Source

Why does this matter? Because after the $1.5 billion settlement with book authors, Anthropic is already facing the next legal front. The pattern is clear: once models can reproduce text or lyrics, “training” very quickly becomes a copyright problem. For the industry, this is another sign that the era of “let’s just train on everything” is coming to an end. For you, it means rights clearance is becoming a product factor — and not only at launch, but already at the data sourcing stage.

🎥 LAION provides massive open video dataset

With the Big Video Dataset (BVD), LAION is releasing one of the largest open video datasets for AI research: around 80 million videos, ten million hours of material, and 55 million clips with automatic descriptions. For the open-source and research world, that’s a pretty big deal. Models trained on BVD are said to outperform the previous reference dataset InternVid by up to 2.1 percentage points. That may sound small, but in benchmark land it is often a visible jump. Source

Why is that important? Video AI has been stuck with a data problem for years: well-curated, open, and at the same time large enough datasets are rare. BVD could bring research groups and smaller labs closer to the level of the big proprietary players. At the same time, the legal question remains tricky, because video data is even more tightly tied to platform, upload, and licensing issues than text. In short: a technical win, a regulatory minefield. Open does not automatically mean unproblematic.

🧠 AI agents have no sense of time

A new study on AI coding assistants such as Claude Code and Codex shows a problem that sounds very human, but is technically significant: the agents have no usable sense of time. They sometimes overestimate task duration dramatically — Codex in some cases by a factor of ten — and also rate their own work much too positively on average. For long, autonomous tasks, that is a real control problem. Source

Why is this more than a cute side note? Because with agents we often pretend they are little software employees with a to-do list and initiative. In reality, they often lack even the ability to assess progress, effort, and quality in a reasonably stable way. This is especially relevant for coding workflows, where you don’t want to babysit the agent every minute. The study is therefore a good reality check: autonomy without reliable self-awareness is above all one thing — a pretty risk with a UI.

⚖️ Texas halts new Flock cameras

Texas Governor Greg Abbott has frozen state funding for additional Flock cameras. The move comes amid growing criticism of the AI-powered surveillance network and just before an investigation revealed spending of more than $30 million. Officially, the cameras were supposed to help, among other things, combat catalytic converter theft; critics, however, see above all an expanding, poorly controlled mass surveillance system. Source

Why it matters for the AI news landscape: this is not about a new model, but about the political backlash against “AI surveillance” in everyday life. The topic is important because many govtech and public-safety projects rely on similar promises: more security, less staff, better scaling. The catch is well known: once the infrastructure is in place, it rarely shrinks back on its own. For the debate around privacy, facial recognition, and automated surveillance, Texas is therefore a very clear signal.

📏 New theory for adaptive nearest-neighbor classification

A fresh arXiv paper titled Curvature-Aware Radius Shrinkage for Adaptive Nearest Neighbor Classification addresses a fundamental problem in machine learning: classic k-NN methods treat the neighborhood in data space the same everywhere, even though local geometry often differs significantly. The paper proposes adapting the radius based on curvature instead of stubbornly working with a fixed neighborhood size. Source

Why should you care? Because papers like this are a reminder that ML is not just about more parameters and more data. Sometimes the right question is not “How big is the model?”, but “How do I define proximity in a meaningful way?”. That’s often where the difference lies between an elegant method and a fragile improvised solution. For ambitious beginners, this is a good rule of thumb: once data lies on a complex geometry, simple assumptions become expensive very quickly. Mathematics, unfortunately, is remarkably incorruptible in that respect.

🧩 Topology, certification, and code world models

Another arXiv paper, An Enclosed Mode Is a Gauge Choice: Topology Relative to Reach in Certified Code World Models, deals with a more theoretical but interesting problem: what can a certified code world model actually know — and where are its blind spots? The core message is pointed: a model can be correct for everything the sampling gate sees and still be completely off outside it. Source

That sounds abstract, but it is quite relevant for LLMs and model merging. Because as soon as you combine models or secure them through filters and certificates, the question of the “outside region” arises: what was really understood, and what was only made locally fitting? Such work is important because it formulates the limits of verification in generative systems more cleanly. In short: a certificate is only as good as the slice of the world it covers. The rest is all too often mathematically elegant and practically unpleasant.

🛠️ Tool tip of the day

If you work with open datasets, benchmarks, or model tests yourself, a clean evaluation workflow is worth it. A practical starting point is an experiment tracking tool like # — it helps you keep a much better overview of datasets, runs, and metrics. Especially in research setups with video, LLMs, or agents, it saves you a lot of “Which version was that again?” chaos.

Don’t want to miss any news? Subscribe to the newsletter