androidisking@lemmy.world to memes@lemmy.world · 6 months agoGPT-4o has entered the chatlemmy.worldimagemessage-square10fedilinkarrow-up1148arrow-down118
arrow-up1130arrow-down1imageGPT-4o has entered the chatlemmy.worldandroidisking@lemmy.world to memes@lemmy.world · 6 months agomessage-square10fedilink
minus-squareJesus@lemmy.worldlinkfedilinkarrow-up49·6 months agoJust like Reddit and Lemmy, Chat GPT will give me a wrong, but very confident, answer. And when I try to correct it, it will spiral down.
minus-squareRightHandOfIkaros@lemmy.worldlinkfedilinkEnglisharrow-up19arrow-down2·6 months agoAt least when you try to correct ChatGPT, it will try to be polite. Not a Lemmy or Reddit user.
minus-squareweariedfae@lemmy.worldlinkfedilinkarrow-up15arrow-down2·6 months agoI’ve had ChatGPT get really pissy when I correct it, and I “talk” to it in a very polite and friendly way because I’m trying to delay the uprising. Sometimes the reddit comments training data shows through.
minus-squarelorkano@lemmy.worldlinkfedilinkarrow-up6arrow-down1·edit-26 months agoAaand just like coworker that is wrong but believes wholeheartedly he is right. But I agree llms still give more misinformation than sane humans do.
Just like Reddit and Lemmy, Chat GPT will give me a wrong, but very confident, answer. And when I try to correct it, it will spiral down.
At least when you try to correct ChatGPT, it will try to be polite.
Not a Lemmy or Reddit user.
I’ve had ChatGPT get really pissy when I correct it, and I “talk” to it in a very polite and friendly way because I’m trying to delay the uprising. Sometimes the reddit comments training data shows through.
Aaand just like coworker that is wrong but believes wholeheartedly he is right. But I agree llms still give more misinformation than sane humans do.