Would that work for LLMs though? They hypothetically trained on news papers from the second half of the data so they have knowledge of "future" events.