The ability to systematically create file embeddings and use them for semantic search. This tech is now 9 years old, it’s not wildly energy inefficient, it only sucks that when you change the model you need to recreate the entire index.
A good accessibility interface that AI tools incidentally benefit from
Accessibility features for humans, like how Apple lets you select text on any image. Or I can send an image to Gemma 4 and the transcriptions are actually quite good running locally (though the model is large)
All I really want related to AI in my OS is: