Training
Also known as · pretraining · model training
The compute-intensive process of teaching a model by adjusting its parameters.
Training is how a model learns. It's shown enormous amounts of data and repeatedly tries to predict the next token; each time it's wrong, an algorithm adjusts its parameters slightly to do better next time. Done across trillions of tokens, this builds the model's capabilities.
Training usually has stages: pretraining on broad text to build general competence, then fine-tuning and alignment steps (often using human feedback) to make the model helpful, follow instructions, and avoid harmful output. Pretraining is the expensive part — it can require thousands of specialized chips running for weeks.
The key contrast is with inference, which is using the finished model. Training happens once (or periodically); inference happens every time you send a prompt.