LEVEL 01 · LESSON 10
Inference vs training
Training changes the model; inference applies it to the information supplied now.
What it is
Training changes a model’s learned parameters using data. Inference runs an existing model on new input to produce an output.
Why it matters
This distinction helps explain where customization happens and what a system actually learns from an interaction.
A business example
An assistant uses a company handbook included in its request. It can answer about the handbook during that interaction without the company training a new model. Updating the handbook in retrieval changes the evidence used at inference time.
How it actually works
Training repeatedly adjusts parameters according to an objective. Fine-tuning is additional training on a selected dataset. During ordinary inference, the model uses its existing parameters together with the request context. An application can store preferences or conversation summaries and supply them later; this is application memory, not necessarily parameter learning. Whether a service later uses interactions for training depends on its policies and configuration.
A common misconception
A model using something you just told it does not mean its weights changed or that it permanently learned the fact.
ONE SENTENCE TO REMEMBER
Training changes the model; inference applies it to the information supplied now.
CHECK YOUR UNDERSTANDING
Early in a chat, you tell an assistant that your fiscal year starts in April, and it uses that fact later in the same conversation. What happened?
Show the answer
Answer: B.
During inference, the model combines its existing parameters with the context it receives. Using something you just said does not mean its weights changed. An application can store facts and supply them again later, but that is application memory, not training.
Take it into your next meeting
Would your requirement be better met by fresh context or a change in learned behavior?
One more concept, understood.
Mark this lesson complete to keep your place.