The quality of this tweet and discussion is extremely disappointing. If you read the actual paper, you'll realize that Apple is not claiming to have discovered

The quality of this tweet and discussion is extremely disappointing. If you read the actual paper, you'll realize that Apple is not claiming to have discovered a magical breakthrough allowing their 80M model to somehow be competitive with 1T+ GPT4 on general tasks. This model does reference resolution. Given a query like "call the number" and a list of numbered on-screen elements, it'll tell you which elements the user may be referring to. It doesn't understand anything beyond just outputting number(s) referring to elements as such. You could consider "better than GPT4" to be true in this specific context, but this is the worst kind of truth. Then I've also trained many models that are "better than GPT4", only because GPT4 is not an AGI that is capable of everything in any modality. Windows 95 also shipped with an executable called "calc.exe" that's even better than GPT4