AMD Just Bought an AI Chip That Can Never Be Updated

AMD just bought a chip that can only ever run one AI model. Why would anyone want that? Taalas etches the weights into the silicon itself — one transistor holds a four-bit weight and does the multiplication where the data already sits. Seventeen thousand tokens a second per user, on a rack pulling twelve to fifteen kilowatts instead of a hundred and twenty to six hundred. So what happens when the model needs updating? If the weights are in the mask, a model update is a mask revision. Not a download — a trip through the fab. Roughly two metal layers change, not a full respin. Does that make it fine? No. You cannot patch it. A bug, a jailbreak, a safety fix — that die runs what it was born with until somebody cuts a new mask set. So is this an AI story, or a manufacturing one? Your model version is now a reticle revision, and your release notes are change control. Now, your FabSpeak Tip of the Week. Keep a prompt changelog — one page, three columns: the date, what you changed, and what the output did. Because when an answer gets worse, which one moved: your prompt, or the model underneath it? Without a log you cannot tell, and you will rewrite a prompt that was never the problem. Pin the model version at the top of that page too, because providers ship quietly and it used to work is not a diagnosis. That is exactly how a fab handles a process change. Forty years of the same rule — you never change two things at once, because afterwards nobody can say which one did it. That's FabSpeak. Back next week — check what changed before you change anything.