Howdy wizards,
Here’s what’s brewing in AI.
The big thing
Anthropic released the Model Hardware Standard (MHS), which lets AI agents operate lab and factory machines.
Source: anthropic.com
Getting an AI to control a physical machine used to mean writing custom code for that one machine. A lab full of instruments often meant weeks of that work, device by device.
MHS replaces that with a one-time setup.
The machine’s owner writes a plain-language description of the equipment: what it is, what it can do, details like the weight of a robot arm.
MHS then turns that description into a standard driver. From then on, any agent can find the machine and operate it, with commands as simple as read and write.
You’re likely already using MCPs (Model Context Protocol), connectors for agents to apps, databases and tools. It was also launched by Anthropic, and is adopted widely across labs and companies now.
MHS is the same idea for physical equipment, and works with any AI model, not just Claude.
Anthropic’s own early results show MHS at work. QuEra Computing builds quantum computers, which depend on finely tuned lasers that drift and need retuning. With MHS, that retuning is now about 25 times faster, and it works nearly every time instead of about half. Carnegie Mellon University connected its lab instruments to an agent through MHS in eight hours, where it would’ve taken weeks.
A key limit to MHS is that models don’t have the physical intuition to troubleshoot things like we do. Like a bubble forming in a sample. Also, if a machine has no programmable interface, it can’t be connected at all.
The standard is currently in a research preview, and the only way to get access is by applying to Anthropic. After the safety evaluations are done, the plan is to make it open source.
Why it matters
Claude for machines!
Having agents control other software reliably is a problem that’s mostly solved. MCPs paved the way quickly, and became the default way agents connect to apps.
Will MHS become that for hardware? Some say just letting the agent build a tool is the better way. And it might be.
But something the MCP story shows is that a standard is very powerful. Everyone doing the same thing is efficient, and efficiency goes a long way.
IN PARTNERSHIP WITH GRANOLA
You’re two minutes out from a client call. You can’t remember what you discussed last time. You’re scanning old emails and digging through notes, piecing it together as the call starts.
It’s not that you’re disorganised. Prep takes time nobody has between back-to-backs.
Granola just launched Briefs, a meeting prep feature that gathers context for you before you even ask.
Overnight, Briefs pulls together who you’re meeting, what you discussed last time, recent company news and any email threads (if you connect Gmail). By the time you join the call, you glance at two or three lines of prep. Everything’s sourced, so you can dig deeper if you need to. No setup. No prompting.
NEWS NEWS NEWS ❦ NEWS NEWS NEWS
All the small things
New tools & product features
ChatGPT Work can now sign in to your accounts, without your password entering the chat. When it hits a login wall, ChatGPT shows a password box on your side and pipes what you type straight into the browser, bypassing the model’s context. And the session sticks for later jobs, so you don’t have to keep logging into the same platform (clearable in settings). Haven’t tried it, but anything that can make authentication flows for agents easier + more secure sounds like a step forward.
Perplexity and Nvidia released Portable Computer, an agent that runs entirely on your own machine. It’s Computer, Perplexity’s do-it-for-you agent, moved off the cloud. The model runs on your desk and your files never leave the room. On-device work costs nothing per task, and it can still call out to 15+ cloud models. Before each call, it checks the outgoing text for personal details, shows you exactly what would leave the machine, and waits for your OK. It’s off limits to most people still because your own computer needs Linux + Nvidia hardware with at least 24GB of graphics memory, though.
Industry moves
Nvidia has reportedly agreed to buy Hugging Face for $12.9B, while keeping the models open. Earlier this year, Hugging Face turned down an Nvidia investment at a $7B valuation. This is still a rumour.
Salesforce inside Claude is here. Claudeforce uses MCP to give Claude 37 ready-made sales skills, from pipeline reviews to meeting prep. Claude reads live Salesforce data and acts on it, updating deals and running workflows, while Salesforce keeps enforcing who’s allowed to do what underneath.
Models
Ox Alpha, last week’s mystery model, turned out to be GLM-5.3-Flash from China’s Z.ai. And it runs on Chinese-made chips. Anonymous models on OpenRouter tend to be Chinese, and so was this one. The model is now free to download and use (MIT-licensed) on Hugging Face, so the catch I wrote about earlier (the provider keeping your prompts) is gone now that you can run it yourself.
Research
Surgeons in London removed a brain tumour with an AI watching the camera feed and marking what not to cut. It’s a world first. The AI worked from the live video during the operation, on a tumour that sat on the patient’s pituitary gland and was slowly taking his eyesight. Surgeons went in through his nose while the AI watched on its own monitor, flagging hidden arteries and optic nerves. It was trained on hundreds of labelled videos of past operations, and had seen, in the doctor’s words, “more operations than most surgeons see or do in a lifetime.” The best AI models, with the right context, are getting reliable enough to assist real-time brain surgery.
❦
You are a delight.
What's your verdict on today's email?
See which AI use cases are paying off with Context Windows Pro
Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.
I’ve created Context Windows just so you can pick the winners.
🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.