Four stages that sound similar

These words get mixed up constantly in product meetings. Separate them once and a lot of GenAI conversations get clearer.

• Training — the model learns by updating parameters on training data.
• Validation — you measure performance on held-aside data while you are still developing or tuning.
• Testing — you measure performance on fresh held-out data to estimate how well the finished model generalizes.
• Inference — you use the trained model to produce outputs for real inputs.

When you send a prompt to a hosted LLM API, you are almost always doing inference. You are not retraining the foundation model on every request. Your application may add retrieval, tools, or fine-tuning later — but the day-to-day API call is still “run the learned function.”