By skyforbes Hey everyone,
Just wanted to share a project we’ve been experimenting with. We built a tool called ClarityBubble after getting annoyed at how often AI text still sounds weirdly stiff or synthetic. Instead of trying to prompt our way out of it, we collected a big dataset (about 2.5M+ examples) of “AI-ish text → human rewrite” pairs and finetuned a smaller model ourselves.
Just wanted to share a project we’ve been experimenting with. We built a tool called ClarityBubble after getting annoyed at how often AI text still sounds weirdly stiff or synthetic. Instead of trying to prompt our way out of it, we collected a big dataset (about 2.5M+ examples) of “AI-ish text → human rewrite” pairs and finetuned a smaller model ourselves.
It’s been interesting seeing how it behaves. It’s not perfect, sometimes it swings too casual, sometimes it misunderstands tone but overall it does a decent job at smoothing out that artificial vibe. Surprisingly, it’s also getting past most AI-detectors people like to test, which we didn’t even plan for initially. Also got quite a good amount of users which again didn't expected 🙂
If anyone’s curious or wants to poke at it, we put it online at claritybubble.com. Not trying to pitch anything, just sharing what we’ve learned so far.