I was in a long term relationship with someone with a drinking problem. When a drunk person says something to you they fucking mean it. Their filter is weaker and, as I label it “they have the courage to say something that they were afraid to say sober.” When you figure this out everything is easier to understand.
We should are least refer to inference LLMs as LLMs. The fact that if you asked it something like who is the current CS2 top team, it would give you the top team at the time it was trained is enough proof that the models effectively know nothing.