AI Controls Your PC, France Builds Data Centers

Today is one of those days when the AI world is moving on three levels at once: at work, in research, and in infrastructure. While OpenAI brings the next stage of PC automation to Windows 11, new papers show where models need to become more robust — and where transparency can suddenly become a security problem.

At the same time, Europe continues building out the physical foundation for the AI era: data centers, power, cooling, compute. In short: today is about who uses AI better, who understands it better, and who even has the infrastructure for it.

🤖 OpenAI brings Computer Use for Codex to Windows 11

OpenAI is expanding the Codex app with Computer Use on Windows 11: the AI can now independently operate programs, test apps, and follow up on errors — basically taking over the “mouse-click middle management.” What makes this exciting is not just the automation itself, but the move from text and code assistance to real desktop automation. For developers, that means testing, reproducing bugs, and routine tasks could become much faster if agents can reliably navigate interfaces. At the same time, the risk increases that an AI agent ends up in the wrong window and confidently clicks the wrong button. The bigger context: we’re seeing how LLM tools are increasingly growing into concrete workflows — especially on Windows, where many companies still rely on classic desktop processes. Source: The Decoder

🧠 Diversity improves test-time compute in vision-language models

The paper “Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models” examines how test-time compute (TTC) behaves in vision-language models. The idea: at runtime, a model gets more compute time to find better answers through additional steps or candidates. This is well known in the LLM world, but apparently still much less systematically studied for VLMs. The authors analyze seven models and six benchmarks and show that not every TTC strategy works equally well — especially important seems to be the diversity of the generated suggestions. For you, that means: more compute time alone is not automatically better; what matters is how intelligently the model uses that time. For product teams and researchers, this is relevant because VLMs are increasingly ending up in assistive systems, document processing, and multimodal search applications. Source: arXiv

🔍 Explainability can make attacks on graph neural networks easier

A new paper with the unwieldy but important title “Can Subgraph Explanations Be Weaponized to Steal Graph Neural Networks?” shows a problem that appears in many ML products: what is well-intentioned for transparency reasons can help attackers. Specifically, it concerns Graph Machine Learning as a Service and Subgraph Explanations, meaning explanations that show subgraphs as reasons for a prediction. That very interface can be abused for model extraction attacks to better imitate or steal a model. This is especially relevant for companies deploying explainable AI in regulated environments — for example in fraud detection, chemistry, or network analysis. The core takeaway: more transparency does not automatically mean more security. Explainability has to be designed so it helps users without making the model unnecessarily vulnerable. Source: arXiv

📈 Convergence of Adam and steepest descent is analyzed more precisely

With “Convergence of Steepest Descent and Adam under Non-Uniform Smoothness”, research delivers a piece of foundational work for training machine learning models. It asks when optimizers such as Adam or steepest descent converge cleanly under more realistic assumptions. Instead of idealized smooth landscapes, the paper considers a form of non-uniform smoothness, where curvature depends on the function value. That sounds dry — and it is a bit — but results like these are important for understanding why training methods behave stably in practice or why they don’t. For the AI community, the takeaway is: better theory creates better heuristics for training LLMs, reasoning models, and other deep networks. This is not the kind of paper that goes viral, but it is the kind that later appears in the footnotes of major systems. Source: arXiv

🌐 SoftBank wants to invest up to 75 billion euros in France

SoftBank is apparently planning a massive expansion of its AI infrastructure in Europe: up to 75 billion euros for data centers in France, with up to 5 gigawatts of additional capacity. That’s not just a big number; it’s a signal of how intense the competition has become around data centers, power supply, and geopolitical location advantages. France likely benefits here from energy policy, available land, and strategic positioning in Europe. At the same time, as always with SoftBank, the announcement is spectacular, but the real story is whether it actually gets implemented. For the market, it still matters because AI is no longer just a software topic. If you want to train and run models, you need power, cooling, network connectivity, and capital — lots of capital. Source: The Decoder

👥 The gender gap in AI coding tools is bigger than expected

An Anthropic study, covered by The Decoder, shows a clear gender gap in AI coding tools: researchers with typically male names use coding agents such as Claude Code much more often than those with typically female names — even at similar subject areas and career levels. This matters because it’s not just about “usage,” but about early access to a tool that can change working methods, speed, and productivity. If one group uses these tools more often, new inequalities quickly emerge in output, learning curves, and visibility. Also interesting: according to the study, the gap is larger for coding agents than for general AI use. That suggests specialized developer tools do not automatically spread widely, but need to be introduced and supported deliberately. Source: The Decoder

🛠️ Tool tip of the day

If you want to experiment today with Computer Use or other AI agents for desktop automation, plan a clean test setup right away: a sandboxed Windows VM, clear target tasks, logging, and as few distractions as possible. That’s exactly where the most exciting workflows happen — and the most embarrassing misclicks. For teams that want to use this productively, it’s worth looking into suitable developer tools around # and automation #. If you want to secure the right setup right away, also take a look at #.

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