Fix Prevents AI From Giving Wrong Answers on Nvidia's New GPUs

The team behind PyTorch, the open-source foundation of countless AI programs, just shipped a safety fix. They are disabling a buggy chunk of Nvidia's cuDNN library on Nvidia's newest Blackwell GPUs. cuDNN is Nvidia's toolbox for making AI math run fast on its graphics chips. The bug they're sidestepping could cause AI models to miscalculate during the very moment they're composing an answer.

You may never see PyTorch or cuDNN. But when you ask ChatGPT or another assistant a question, those tools are working underneath. During "decode" — the step where the model predicts one word at a time — a tiny math error can snowball into a bizarre or factually wrong response. Worse, the model doesn't always show signs of failure. It gives a confident but wrong answer. So this is a quietly important fix: it protects the trust that makes people rely on AI.

The workaround is not a permanent cure. The team is simply adding a "guard" — a check that tells the system: skip this particular routine on Blackwell chips with certain cuDNN versions, specifically 9.19 and up. They're waiting on Nvidia to confirm a fixed build that mathematically corrects the bug. Think of it as removing a corrosive ingredient from a recipe instead of throwing out the whole cake.

Because this is a release candidate, it's aimed at developers and organizations testing new PyTorch 2.14. It shows how much behind-the-scenes care goes into reliable AI. When you get a correct answer from a chatbot running on a hopped-up Blackwell server, you can thank this little guardrail.