LLMs predict next tokens following their training data. Following "so i wrote this little program in", "Haskell" is not a likely prediction.

At this stage, people who parrot “next token prediction” appear the most predictable of all!