A daily roundup of what's moving in AI.
Anthropic just gave AI agents a way to talk to the physical world. The company's new Model Hardware Standard, released in research preview, lets agents look up, learn, and operate lab and factory equipment — microscopes, robotic arms, assembly-line gear — the same way its Model Context Protocol let agents talk to software. Anthropic says setup drops from weeks of specialist wiring to "hours or minutes," and in one internal test, Claude taught itself to align a laser through trial and error, then wrote a script to automate the job in a single pass. Tecan, QIAGEN, and AWS are already partnering, with Hugging Face and Raspberry Pi adding support to their device lines.
Speaking of Hugging Face: Nvidia is reportedly in talks to buy the "GitHub of machine learning" for roughly $13 billion. Fresh off blowout earnings, the deal would hand the world's most valuable company a direct line to the developers building on its chips. Nothing's official yet, and neither side has commented.
Salesforce had its own big week, expanding its partnership with Anthropic into something called "Claudeforce" — a 37-skill sales plugin built on Claude, plus Claude becoming the default model inside Slack. The announcement helped send Salesforce stock up 22.6% in a single day, its best showing since 2020.
On the research side, two scientists turned down a stake in Jeff Bezos's Project Prometheus to start their own company instead. Accelerated Understanding is building AI trained on physics rather than language — models that understand the 4D world (3D space plus time) for scientific discovery. They've already shipped a model handling 5 trillion data points.
And in the UK, surgeons performed what's believed to be the first brain surgery assisted by a live AI system. A University College London model watched the surgical camera feed in real time, flagging arteries and the optic nerve to avoid during a tumor removal. The patient's vision was restored within days; the AI never touched the controls — it was a second set of eyes, not a second surgeon.
Also worth knowing
- 116 companies, including OpenAI and Anthropic, signed an open letter warning that AI-enabled cyberattacks are about to get "far more widespread and sophisticated," calling for coordinated defense.
- Instinct, a text-and-call AI assistant with no app at all, has hit a $2.5B valuation since launching in February.
- Barret Zoph, a Thinking Machines co-founder, is rejoining Google as a research VP for reinforcement learning on Gemini.
- Google shipped Gemini Omni 1.1 Flash, an upgraded video model with 40-second scene extensions and 4K upscaling.
- Voice dictation into AI agents is going mainstream, with Apple shipping a dictation keyboard in iOS 27 and tools like Wispr Flow becoming the way power users direct multiple agents at once.
- Shopify's CEO is reportedly weighing a company-wide ban on Claude Code, saying it doesn't mesh well with other workflows.
