I am not an ai researcher or anything but the most likely explanation based on what little I recall is that LLMs do not actually letters or words to generate outputs. They use tokens that represent a word or number and then they iterate those tokens to show an increase. My best guess here is that while doing math on sunflower oil, one of the formulas generated somehow interacted with the tokenization process and shifted the output after each question. Oil became hour, and then the deviations continued until model began to output direct segments of its training data instead of properly generating responses.

Again this is absolutely speculation on my part. I don't have much of a direct understanding of the tech involved