The language model gets tokens as input, and produces tokens as output. Not words. And tokens do not correspond to words one to one.[1] So it cannot see these

The language model gets tokens as input, and produces tokens as output. Not words. And tokens do not correspond to words one to one.[1] So it cannot see these "words" you are talking about. The fact that it can still get close when asked to count words is actually very impressive. [1] https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them https://help.openai.com/en/articles/4936856-what-are-tokens-...