In the standard next-token markov model, just have one node for every token and 1 probability for every edge. E.g, "if last letter is "B" most likely next lette

In the standard next-token markov model, just have one node for every token and 1 probability for every edge. E.g, "if last letter is "B" most likely next letter is "C"" To make a big language model in Markov space, you need to very large ngrams. "If last text is "th" next letter is "e"" These get incredibly sparse, very quickly. Most sentences are novel in your dataset. Neural networks (and other models too) get around this by placing tokens in multidimensional space, instead of having individual nodes.