SoftBank plans to build AI data centers with up to 5 gigawatts of capacity in France, the company’s largest AI infrastructure investment in Europe, at up to 75 billion euros. By 2031, facilities worth 45 billion euros are set to go up at three sites in northern France. SoftBank’s mega announcements keep stacking up worldwide, but many projects have yet to materialize.
Nvidia is entering the PC market. According to Axios, the first Windows computers running Nvidia chips as their main processor will be unveiled next week at Computex in Taiwan and Microsoft's Build conference in San Francisco. Both Microsoft's Surface brand and Dell are expected to show devices.
Microsoft is also building new software that lets AI agents handle tasks locally on Windows PCs. The company has been betting on OpenClaw since early this year, setting up a dedicated team under developer Omar Shahine. OpenClaw founder Peter Steinberger, now at OpenAI, is scheduled to hold a session at Build, suggesting Microsoft may use the OpenClaw framework for these PCs.
Microsoft's first AI PC push—the "Copilot+ PC"—tried to use AI as a marketing hook for laptop sales while forcing Copilot into the default stack. It largely flopped. This second attempt seemingly goes deeper, aiming to weave AI agents into actual workflows. Security and reliability concerns around OpenClaw remain, even if everything runs locally.
Mathematician Terence Tao explains how AI could reshape math research by enabling division of labor. Until now, mathematicians had to do everything themselves: framing problems, building strategies, executing them, verifying results, and writing them up. Unlike industry or the natural sciences, specialization was never an option in math, Tao explains.
AI and formal verification could change that by filling skill gaps in collaborations, Tao says. But if AI generates strategies without verifying them, the result is a flood of untested ideas. A new style of math only works when automation advances across several areas at once. Tao sees humans as essential because AI performance is uneven—a principle that likely applies to many other fields, too.
The level of automation and AI power that you can profitably use before it becomes slop is roughly proportionate to how stringent your verification is.
Terence Tao
It seems like the field is moving toward Tao's vision of "industrial mathematics": instead of solo researchers grinding away for years, large AI-supported teams could pursue broader but shallower research. AI crunches billions of data points, while humans make "inspired guesses" from a handful of observations.
Attackers are exploiting shared chat links in ChatGPT and Claude to push malware. Both platforms let users share conversations publicly via URL. Victims stumble onto these chats through paid search ads. Because the links sit on trusted domains, security tools don't flag them, and users are more likely to trust what they see.
| Indicator | Type |
|---|---|
| hxxps://claude[.]ai/share/8e6401b5-4849-46c4-a3cb-29e1c3c49131 | URL |
| hxxps://chatgpt[.]com/s/cb_6a0f1e6bbec88191aa7fede27163f08d | URL |
| openew[.]app | domain |
| de8c50e8ccd240ef9d10ec26c26eeb37a4d1cad7c1e0edf3bb6e5689ec2dde78 | SHA256 |
Security firm Push Security says attackers craft shared chats that mimic official outage notices or install guides. One newer twist uses ChatGPT's code-rendering feature to build a full fake error page right inside a shared chat, then pushes users to download an infected desktop app. On Claude, shared chats pose as Apple support walkthroughs laced with malicious Terminal commands.
Push Security calls the attack technique "LLMShare." BleepingComputer and Kaspersky have both documented similar campaigns.
OpenAI has expanded its Codex app to Windows 11, adding "Computer Use" and mobile access. The AI can now use apps, files, and other resources on a PC, even when the user isn't at the computer. That means testing apps, hunting bugs, or reviewing work. The feature is toggled on in Codex's settings. Commands like @computer or @Paint target specific programs.
Codex is also now available through the ChatGPT app on iPhone and Android for starting or monitoring tasks on a Windows machine remotely. Computer Use launched on macOS in April. Mobile access followed in May.
The rapid Codex expansion is part of OpenAI's plan to build a "super app" for work and daily life. ChatGPT could eventually fold into that app too, though the brand is likely too strong to subsume under the Codex name, which primarily targets developers.
Salesforce says it moved its entire dev org to Anthropic’s Claude Code with no token limits and reports massive productivity gains for April 2026: 79 percent more pull requests per developer, five percent fewer incidents. The numbers can’t be independently verified. The case shows just how divided the coding world is over the agentic shift: real revolution or the biggest build-up of tech debt ever?
OpenAI is making several changes to ChatGPT. The recently released GPT-5.5 Instant is getting an update: responses should feel more natural, easier to read, and better structured, with fewer long bullet-point lists, the company says. OpenAI is also removing the Canvas feature from GPT-5.5 Instant and GPT-5.5 Thinking. Canvas is the side panel that sometimes pops up in chat when you're editing text or previewing code. Instead, writing and coding tasks will be handled directly in the chat through special blocks. Paying users can still access Canvas through older models during the transition.
The company is also retiring the older o3 and GPT-4.5 models, with o3 leaving ChatGPT on August 26, 2026, after a 90-day sunset period, and GPT-4.5 on June 27, 2026, after a 30-day sunset period. Until then, both models remain available to paying users through the model settings. o3 will stick around in the API for now; GPT-4.5 was already pulled from the API a while back.
