Training is the step at which a model stops being an empty architecture and starts doing something. The basic loop described by Chollet (2021) has four steps: draw a batch of examples, compute the model's predictions (the forward pass), measure with the loss function how wrong it was, and update all the weights so that the error goes down a little (2021, sec. 2.4). Repeated millions of times, with backpropagation and gradient descent, the loop takes the parameters from random values to useful ones. Each full pass over the data is an epoch.
It is distinct from inference, the use of the already trained model: training a large model takes weeks and enormous amounts of compute; querying it takes fractions of a second. For language models, training is split into stages: pretraining on massive text, followed by fine-tuning for specific tasks or behaviors (Alammar & Grootendorst, 2024, p. 26).
The word is not innocent. Witten et al. (2017) point out that in everyday language "training" denotes a mindless kind of learning: we train animals and plants (a vine is trained to climb a trellis), and the difference is that in learning the purpose is the learner's, whereas in training it is the teacher's (2017, pp. 8–9). The field's jargon, curiously, uses "train" and "learn" as synonyms.
Santelli (2026) stresses the practical flip side: these programs are not programmed but trained; they are grown rather than built, and they acquire capacities their designers did not aim for. If the purpose belongs to the trainer, whose is the capacity that nobody set out to train? (see emergent abilities).