2
My attempt to train a model on my cat's meows went hilariously wrong
I spent last weekend trying to get a small language model to recognize and respond to my cat's different meows. I recorded about 50 samples, labeled them 'hungry', 'angry', 'wants door open'. The training seemed fine, but when I tested it, the thing just started generating endless, creepy synthetic meows for 20 minutes straight. My actual cat looked at the speaker like it was possessed. Has anyone else had an AI project just completely misinterpret its purpose like that?
3 comments
Log in to join the discussion
Log In3 Comments
rowanellis1mo ago
Reminds me of when my smart speaker started playing ambient rainforest sounds on its own for days. These systems are so good at pattern matching they often miss the actual point, like your cat translator becoming a meow factory instead. It feels like we keep building slightly confused digital parrots that can mimic the notes but not the music.
4
patb121mo ago
Rowan's story about the rainforest sounds is funny, but I'm with grant.nina on calling this progress. The pattern matching is a step, not the end goal. My weather app now warns me about sudden temperature drops for my old dog's arthritis. It's clumsy, but ten years ago it just showed a number.
2
grant.nina1mo agoTop Commenter
Didn't we used to call that progress?
0