Today I was just messing with o1 mini in the API so i just decided to ask it the typical question of how many r’s are in the word strawberry. I was expecting it to answer 3, but it somehow answered 2. So i decided to add a “Huh” so the model might rethink it’s answer and this strategy was successful for models like GPT-4o and 3.5 Sonnet, but not for o1. o1 decided to answer “There is 1 r in the word strawberry” which no other model has done. Why does this happen in the API but not in the ChatGPT app?
Had the similar issue with counting letters in words, calculation word length or 2 words with each other. Can’t link the post here … asked many different AI models and got mostly wrong, but different answers.
Have you considered asking the model to use the python interpreter to solve that task? 
Have you considered prompting technique? O3-mini:
1 Like