I'm pretty dubious - how would the model not get absolutely swamped by the vast amount of potential context if it's not learning to ignore long range signals fo

I'm pretty dubious - how would the model not get absolutely swamped by the vast amount of potential context if it's not learning to ignore long range signals for the most part?