4 ms·
AI can't even be bothered to do half-reliable text translations, seriously doubt that 40 years from now it will work half as well as we imagine it to be. After
by ekianjo 5y ago
AI can't even be bothered to do half-reliable text translations, seriously doubt that 40 years from now it will work half as well as we imagine it to be. After all, we all imagined back in in the 80s that we would have flying cars and robots around us by 2020's - we should learn that we are constantly bad predictors and completely off the mark.
- teekert 5y agoNo way. The Photos section of the iPhone Photos app is already almost there. It makes the nicest videos from live photos (.heic), with music and it chooses the nice moments by itself. Maybe this not-super-critical-and-necessarily-100%-specific-part is where AI already shines. The "For You" content has complicated titles like "$kid'sName spring 2020" or "A day in the park, summer 2021" or "Happiness". 40 years is a lot.
- huntertwo 5y agoWhile I agree that the expectations for AI are too high, I think the use case described is almost achievable with today's tech. It's just language processing to process the request (User wants to see dog around this time period in this location) and finding relevant images (find pictures with dog with exif data that matches time period and location).
- gwern 5y agoYou could do most of that with some creative use of CLIP. 'Show more of the mountain' is trivial out of the box, and the 'kids aged 5-10' just needs a bit of logic: take a few photos of each kid at various ages 5-10 (perhaps already specified by the user?), embed, average along with the text prompt 'kids aged 5-10', and use those as the targets. 'Around that time' and 'in and around that house' can be done by metadata dates + GPS + image embedding of the 'house' (or just the text embedding of the word 'house'!) to filter photos by date+location, and then similarity to 'house'. And so on. The question there is how to be able to create those queries automatically in response to transcribed user voice commands. But hey, given Codex/Copilot, who can doubt that programming such queries is out of reach of a contemporary NN either?