Training a model versus using it

What is pre-training?

The model learns general patterns from a huge number of examples. For language models, a common task is predicting the next token.

What is fine-tuning?

A model can be trained further for a particular task, language or style of answering.

Does it learn from every answer?

Usually not, in the sense of permanently changing its parameters. But the app may store data, use a memory feature or allow later training.

What is inference?

The process in which an already-trained model produces output from the current input.

Remember: “I’ll remember that” in a chat does not necessarily mean the model’s parameters have changed.