SoftBank Plans €75 Billion AI Data Center Buildout in France
SoftBank has announced plans to build AI data centers across France with up to 5 gigawatts of capacity. The investment could reach as much as €75 billion (roughly 12 trillion yen), with facilities worth €45 billion slated for completion at three sites in northern France by 2031. It is the company's largest AI infrastructure investment in Europe to date.
That said, SoftBank has announced a string of mega-projects around the world, and observers note that many have yet to materialize. How far this plan actually progresses remains to be seen.
The Reality of AI Search Agents: "Verification," Not "Search"
A research team at Harbin Institute of Technology examined how leading AI search agents such as GPT-5.4 and Kimi K2.6 actually behave. They found that rather than looking up new information on the web, these agents use the web to confirm knowledge they already acquired during training.
The study introduced a new time-based benchmark called LiveBrowseComp. By asking only about events from the past 90 days — creating a situation where models cannot rely on memorized knowledge — performance dropped sharply and existing rankings were upended. It's a notable piece of research suggesting that the "search ability" of AI search may be more fragile than it appears.
Netflix Engineer Open-Sources AI Cost-Reduction Tool
A technical lead at Netflix has developed and open-sourced an application that dramatically cuts AI usage costs by optimizing and pruning the prompts given to AI agents.
As agentic AI adoption expands, simply trimming redundant parts of a prompt can substantially curb token consumption. As a practical cost-optimization approach, it's an initiative worth referencing for enterprises adopting AI.
UC Berkeley Law to Ban AI Use Entirely Starting Summer 2026
UC Berkeley School of Law has announced a policy to comprehensively prohibit the use of AI starting with the summer 2026 term.
As AI tools become ubiquitous, responses at educational institutions remain divisive. Berkeley Law's blanket ban draws attention as one case study indicating the direction of AI policy in higher education.
The Maintenance Problem with AI-Generated Code: "AI Slop Is Hard to Fork"
The quality of AI-generated code is stirring up debate. One developer published an article titled "AI slop is hard to fork," pointing out the low maintainability and reusability of AI-generated code.
While AI tools make it possible to generate code quickly, the output often ends up structured in ways that humans find difficult to understand and modify. The discussion once again throws both the benefits and the challenges of AI-assisted development into sharp relief.
"mlx-Chronos": A New LLM Benchmarking Tool for Apple Silicon
A computer science student has developed mlx-Chronos, an open-source benchmarking tool for fairly comparing MLX inference engines running on Apple Silicon.
Supported engines include oMLX, Rapid-MLX, mlx-lm, and Ollama (MLX backend). It measures cold/cached TTFT (time to first token), throughput, memory usage, thermal state, and more, and results can be submitted to a community-driven leaderboard.
---