Training
The process of building an AI model by showing it vast amounts of data and adjusting its weights to reduce its errors.
Also: model training
Training sets the values of a model's parameters. The model makes predictions on batches of data, the error is measured, and an algorithm called back-propagation adjusts every parameter slightly in the right direction; this repeats millions of times. Training a large language model can occupy tens of thousands of GPUs for weeks or months, with the GPUs constantly exchanging results over the network. It is the most demanding workload in AI compute and the reason for the largest clusters, though inference is expected to use more compute overall as AI use grows.