I keep seeing an irritating misconception in this space, which is that the alternatives chosen by these algorithms are supposed to mean the same things as what they're displacing. That's not true, and not how LLM generation works. Complaints that two different choices don't mean the same thing miss the entire point.