This has to be some AI bullshit recommendations. I’m browsing spoken.io. Not really looking for anything in particular. I tap a coffee grinder (first image). I scroll down. I see “looking for something similar?” and it’s a bunch of products whose product photo is a black monolith on a light background:
- an indoor trash can
- an outdoor trash can (the size of a barrel)
- a cylindrical light fixture that I could mount on the wall
- a nightstand table that is vaguely similar in shape
It’s the most absurd way to calculate “something similar” It’s hard to get this stupid with natural intelligence. You have to augment human stupidity with computer stupidity to get this stupid.
feld
in reply to Paco Hope wishes ill for JK Rowling • • •that doesn't seem as much like AI as just some weird image similarity vector score
Everyone assumes this kind of thing didn't exist and work poorly before LLMs but it totally exists for a long time
Paco Hope wishes ill for JK Rowling
in reply to feld • • •feld
in reply to Paco Hope wishes ill for JK Rowling • • •Matt C
in reply to feld • • •Paco Hope wishes ill for JK Rowling
in reply to Matt C • • •@mkc Using product image similarity to find “similar products”—without also using some other metadata like category, size, or product name to exclude bad matches? It seems obviously broken.
Someone paid to have this implemented. If so few people click it, why do it at all?
Lots of people have pointed out that there’s no reason this needs to be AI. That’s true. It’s just the first explanation I reached for.
@feld
feld
in reply to Paco Hope wishes ill for JK Rowling • • •> without also using some other metadata like category, size, or product name to exclude bad matches? It seems obviously broken.
Unfortunately this is our future. The cost to accurately attach this metadata is too high. It would require manual human intervention considering that they're just scraping the web to even find these images using, presumably, reverse image search tools.
The days of neatly categorized items in a database are pretty much over unless you're a giant corporation that can afford to do it right. If you're a startup? You're just going to rely on any AL/ML tools you can and hope for the best.