LLMs are timeless, when you are sampling from it, words are `materialized`, `instantiated`, become time bounded and you now have newly minted words. the probability of choosing a word, given a context...
The cut-off sentence likely points to temperature and seed. Determining the next token isn't just about static weights; it's a stochastic process where the sampler breaks the tie.
1
