1. SoftBank to Invest Up to €75 Billion in French Data Centers

SoftBank. SoftBank announced plans to invest up to 75 billion euros to build AI data centers across France, targeting up to 5 gigawatts of additional capacity. The initial phase covers three sites in the Hauts-de-France region (Dunkirk, Bosquel, and Bouchain), expected to deliver 3.1 gigawatts by 2031. The scale signals continued capital concentration in European AI compute infrastructure outside the United States. Source

2. Salesforce Claims Claude Code Cut a 231-Day Migration to 13 Days

Salesforce. Salesforce reported that deploying Claude Code with unlimited tokens across its engineering organization reduced an API migration estimated at 231 days to roughly 13 days. The company said developers merged 79% more pull requests while incidents fell 5%, framing the agentic approach as improving both speed and reliability. The claim is a notable enterprise data point on agentic coding at scale, though it comes from the deploying company itself. Source

3. Studies Suggest AI-Generated Code Raises Maintenance Costs Despite Faster Output

AI coding. Reporting on recent research found that while AI tools speed up code generation, they may not improve overall quality, citing a CodeRabbit analysis that AI produced 1.7x more problems than human-written code and reports of teams spending 44% of tokens fixing AI-generated bugs. The synthesis draws on work from METR, Singapore Management University, and CodeRabbit covering adoption and downstream maintenance burden. The findings matter for engineering teams weighing short-term velocity against long-term technical debt as AI dependence grows. Source

4. Study Finds Helpfulness Training Weakens LLMs' Ability to Simulate Human Behavior

Helmholtz Munich. A study led by researchers at Helmholtz Munich and more than 35 institutions, analyzing 26 million responses, found that the post-training that makes models helpful assistants reduces their capacity to replicate human decision-making and biases. Base models capture human heuristics that fine-tuning pushes toward normatively correct answers, degrading behavioral fidelity with each generation. The finding matters for practitioners using LLMs as proxies for human subjects in research, simulation, or policy analysis. Source

5. Attackers Abuse Shared ChatGPT and Claude Chat Links to Spread Malware

Push Security. Security firm Push Security documented an attack technique it calls LLMShare, where attackers use publicly shared chat links from ChatGPT and Claude, promoted through paid search ads, to distribute malware disguised as error messages. Because the links sit on trusted AI vendor domains, security tools often miss them and users lower their guard. The technique exploits the perceived trustworthiness of AI platforms to bypass traditional threat detection. Source

6. Terence Tao Argues AI Could Bring Division of Labor to Mathematics

Terence Tao. At an OpenAI forum event, mathematician Terence Tao argued that AI combined with formal verification could enable specialization and team-based collaboration in mathematics for the first time, departing from the tradition of solitary researchers handling every step. He cautioned that the amount of AI automation usable before output becomes "slop" is proportionate to how stringent the verification process is. The argument frames verification infrastructure, not raw model capability, as the gating factor for AI-assisted research. Source