But I don't see why you'd autocomplete a sentence that could go anywhere. Just because many people type that isn't a good reason.
It's not a search engine's job to be the thought police.
Keep in mind that you are currently enjoying the benefits of discussing this topic on a highly-moderated discussion forum. How relevant do you think Hacker News would be if such moderation was ended entirely?
Have you learned literally nothing from the utter disaster that was Tay[1]?
But if you suggest completions for queries at a point where it isn't yet obvious what the user wants, you are no longer just helpful. You are steering people in a certain direction, whether that's your intent or not. You are effectively telling people, hey, try this query, it's popular, even if the user had no intent of typing that query.
You want the push to be more local to what you specifically choose to pull, which means greatly reducing the current scope of the push component. I guess that's fine, but a lot of people really like the push component, especially if they don't quite know what they're looking for. I think the whole of the two methods is more valuable than the sum of the parts.