Tech IndustryJun 2
Keller Williamskaqb78
Meta AI fell for the strawberry question
Why is it so hard to fix this? There is an easy python code that can fix this, but why aren't they using it. Not just the letter counting, these AI are way off in everything.
Decoder architectures are not suited for math out of the box. They treat numbers like letters and words, and only base on probabilities, frequent questions and answers, etc It would require a lot of additional changes, e.g. special tokenization for numbers, a lot of synthetic math data in pretrain, maybe even another model for math in ensemble
I tested GPT 3.5, 4 and 4o with that question and only GPT 4 passed the test.
Why is this so hard for AI?
It is due to tokenization
Can you elaborate?