3 ms·
One example I came across recently: if you search for "Amiga floppy disk capacity" (https://www.google.com/search?q=amiga+floppy+disk+capacity https://www.googl
by rob74 7y ago
One example I came across recently: if you search for "Amiga floppy disk capacity" (https://www.google.com/search?q=amiga+floppy+disk+capacity https://www.google.com/search?q=amiga+floppy+disk+capacity), you get "1.76 MB", which is completely wrong. Of course, the Wikipedia article which Google's algorithm chose to extract this information from doesn't mention the actual capacity of "standard" Amiga floppies (880 KB): "Most Amiga programs were distributed on double-density floppy disks. There are also 3.5-inch high-density floppy disks, which hold up to 1.76 MB of data, but these are uncommon." - so Google picked the first number it found (the "uncommon" one) and ran with it. I'm just wondering which executive thought having this "feature" would be a good idea?!
- devinplatt 7y agoYou can provide feedback on the answer card. I've had a few answer cards fixed after providing feedback. Notably "https port" used to yield the http port. Edit: I think you've actually managed to find a really terrific example of a case where modern ML systems are going to have trouble, because of how that Wikipedia page is worded in relation to the query. * There isn't a single phrase that answers the query directly, so the ML model would need to make very good use of context (attention), both within multiclause sentences and between paragraphs. * There are many different numbers on the page, so the model has to determine the right one. It can't just get lucky by guessing here. * A wrong number (1.7MB) has close proximity to literal keywords in the query (floppy, disk). (The right number [880KB] does too.) * The model has to understand and properly make use of "most" vs "unusual" in its decision.