The training data doesn't need to be strictly accurate. If it was, you'd just be programming a deterministic robot. The whole point is the feed it actual human

The training data doesn't need to be strictly accurate. If it was, you'd just be programming a deterministic robot. The whole point is the feed it actual human language. Giving it shitposts and sarcasm is literally what makes it good. Think of it like 100 people guessing the number of marbles in a jar. Average their guesses and it will be very close. The training data is the guesses, the inference is the average.