The only one I know of is language translation. Google/Microsoft Translate are terrible, but DeepL is surprisingly good. And I know of other products in this market that are really good.
Deep learning using GPUs and large amounts of training data has redefined the state of the art in most uses of signal processing (voice, music, images, video, etc) and game play.
The "bombastic news headline" stuff often supports improvements in real applications too. It's just that it ends up getting shown off with party-trick type models that look cool but don't go anywhere, adding to the boom and bust cycle (for example search HN for "generative" AI to see a mini hype bubble in full swing: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu... )
I consider it one of the few projects that use modern models to do something that is directly beneficial to me.
That’s now basic technology that I use regularly at work and in my personal life.
As I am from manufacturing side, object detection and segmentation models changed product quality inspection for good.
Now multiple cameras with trained models can find defects on any surface or part as small as it can get, before it was multiple operators looking at moving parts 8 to 12 hours a day which was very tiring and repetitive job to do.