Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Interesting, it was able to guess it's Rot 17, but it translated it wrong, although "HAY" and some other words were correct.

I've tried only in English so far though.

It told me it's 17, and "HAY GENTE MU DIFERENTE LECTURA, A LO MUY GENTE DE TODO ES QUE VER COSAS COMO SON EN VEZ DE COMO DEBEN SER"

although it really should be "HAY LOCURA EN DEMASIADO SANIDAD, Y LO MAS LOCURA DE TODO ES QUE VER COSAS COMO SON EN VEZ DE COMO DEBEN SER"



Claude made similar mistakes of generating decryption that was similar to plaintext but with stuff mixed in. I suspect my version of the quote (Miguel de Cervantes) is an apocryphal translation, and there's some utility well on both models to pull it towards the real one. With that said, I did not see o1-preview get as close as you did.


For testing I think it's better to use uncommon sentences and also start with English first, if it can solve that, then try other languages.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: