this isn’t really even related to GenAI at all

Besides the ocr there appears to be all sorts of image-to-text metadata recorded, the nadella demo had the journalist supposedly doing a search and getting results with terms that were neither typed at the time nor appearing in the stored screenshots.

Also, I thought they might be doing something image-to-text-to-image-again related (which - I read somewhere - was what bing copilot did when you asked it to edit an image) to save space, instead of storing eleventy billion multimonitor screenshots forever.

edit - in the demo the results included screens.