What is training (in AI)?

Training is how a model gets built. It is shown enormous amounts of material, usually text but often images and code as well, and its internal settings are adjusted until its predictions stop improving. Further stages then shape how it responds rather than what it knows. The settings are frozen for release only after all of that, and nothing you type in a conversation changes them.

A model begins as millions of numbers set at random, producing nonsense. Training hides the next word in a passage, compares the model's guess against the real answer, and nudges the numbers slightly toward being less wrong, a method called stochastic gradient descent. Repeat that a few trillion times and the numbers settle into something that predicts language well. Why not program the rules of language directly instead? Because nobody can state them: explaining why "strong tea" sounds right and "powerful tea" does not would take a lifetime, and that is one phrase out of millions. Learning from examples sidesteps the need to write any rule down. Three stages follow one another: pretraining on the bulk of the material, fine-tuning on a narrower set of examples, and reinforcement learning from human feedback, where people rank answers. All three adjust the same numbers. Pretraining is the expensive stage and finishes long before release, which is why every model has a knowledge cutoff. Fine-tuning is cheaper and can happen later, and with open-weight models anyone can do it.

Learning to catch a ball. Nobody hands you the equations for a parabola, and reading them would not help much anyway. You put your hands out, miss, and something adjusts very slightly. A few thousand catches later you are good at it, and you still could not write down the rule you learned.