They are the two stages in the life of an AI model. In training, the model adjusts its parameters on a dataset until it learns patterns; it can have several stages and be repeated to update it. In inference, the already trained model answers each request with what it learned, without learning again.
A question from the app. Pick an answer.
A model came out last week. Does it know what happened yesterday?
Answering does not change what it learned. Its training ended at a cutoff date, before it came out.
It learned up to a cutoff date, before it came out. It knows about yesterday only if it searches current sources or you give it the news.
Giving it the date does not add what it never learned. Recent news has to reach it as text: a search or the article you paste in.
Think of a student. Training is studying all year: it takes months and a lot of effort. Inference is taking each exam: answering quickly with what they already know. Machine learning works the same way: pretraining a large LLM from scratch takes a lot of specialized compute; after that, every time you write to it, the model runs an inference in seconds.
Training from scratch is expensive and rarely done. Inference uses compute on every use; many APIs charge for input and output tokens.
It is the set of examples the model learns from. For a language model, that means huge amounts of text: web pages, books, code. During training, the model adjusts its parameters (its weights) to predict that text better. It learns patterns from those examples and applies them to new cases. What happened afterwards, it does not know.
The dataset is collected up to a certain day, the cutoff date, and then come training, testing and the launch. So the release date does not tell you how recent its data is: check the cutoff date the provider states. Using it does not update it: during inference its weights do not change.
For it to know something recent, you have to give it as text in the prompt: a web search, the article you paste or documents fetched by a RAG system. It helps while it stays in the context, but it is not written into its weights.
"The chat learns from what I tell it." Its weights do not change during the conversation; if an app "remembers" things, it stores them as separate text and passes them back in. "Telling it today's date brings it up to date." It knows the date, not what happened. "To teach it something new you have to retrain it." Often giving it the information is enough; fine-tuning is better for style or format than for data that changes.