Supervised learning

For instance, a model might be shown thousands of emails marked as “spam” or “not spam.” By analyzing those examples, it learns what spam typically looks like. Then, when it sees new emails, it can make its own classifications based on what it learned.

Most modern language models, like ChatGPT, use a variation of this called self-supervised learning. Instead of needing humans to label data, the model generates its own labels by hiding part of a sentence, usually the last word or token, and trying to predict it. This method lets the model learn from massive amounts of text without the need for manual labeling, making it much more scalable and efficient.