Those hacks are literally how a large language model using a transformer architecture to predict the next token in a sequence works.

They take advantage of how a function choosing a token with maximal probability of appearing works.