Anthropic Just Made AI Talk to Machines. Talking Was the Easy Part.

So what did they actually build? On the twenty-seventh of August Anthropic, the company behind the Claude models, released its Model Hardware Standard: one shared way for an AI agent to read from, write to and run a physical machine — a robot arm, a microscope, a pump. The claim that matters to you is time. Carnegie Mellon University connected three instruments and had an experiment running in eight hours; a vendor, they say, would have taken weeks. So is the weeks-to-hours story true? For the talking part, probably. But stay with me — one of Anthropic's own pilots shows you exactly where it stops. In a chip factory a new machine already speaks a shared language — agreed thirty years ago. Yet a new machine still takes three to six months to reach production. Were those months spent teaching it to talk? No. They were spent proving it makes the same product as the machine beside it. That proving is what the AI cannot do for you — here is the pilot. At Genentech, the drug company, the agent chose the same pump speed for plain water and for a thick protein solution. That made bubbles, the bubbles ruined the reading, and the agent tried again in the same well — worse. A human had to explain that the error was physical. The driver was perfect. The judgement was not. So where does your own factory's gap sit? Rockwell Automation, the automation supplier, found in July that ninety-three percent of manufacturers run a factory execution system — the software tracking every job on your floor — but only twenty-three percent have it fully connected. Why? Every connection has to be proven, owned and approved, and that is your people's time, not code. To be fair, Anthropic published those failures itself, and its standard puts hard limits in the driver and a human approval before anything risky. That is the right design — and how your factory has worked for thirty years. Now, your FabSpeak Tip of the Week — ten minutes, none of your company's data. Since the twenty-fifth of August, Claude keeps one memory across its chat app and its Cowork app — on by default on the consumer plans. It learns about you while you type. So read what it has learned about you. Open Claude, go to Settings, then Memory. You will see a list of topics. Open each one, correct what is wrong, delete what you never meant to keep, and decide whether to allow sensitive subjects — they stay out until you say so. Treat that page the way a plant treats a standing instruction: nothing runs from a sheet nobody has read and approved. The machine can talk now. Proving it is right is still your job. Same time next week.