Training
Teaching the model by showing it data and nudging its parameters until the answers get better.
Think of it like
Tuning a piano by ear: play, hear what’s off, adjust the string, repeat thousands of times.
Example
Training a frontier model can take months on thousands of GPUs before it ever answers a single user.
How it actually works
Training runs data through the model, measures error with a loss function, and uses backpropagation and gradient descent to adjust weights. Repeat over enormous data and it slowly learns. It’s the expensive, upfront phase; the result is a frozen set of parameters you then run at inference.
For product teams
It’s the big upfront capital cost; most products consume a model someone else already trained.
For engineers
Iterative parameter optimization via forward pass, loss, backprop, and gradient updates over a dataset.
Related
- Inference — The opposite phase is inference.
- Weights — Adjusts the weights.
- Base model — Produces a base model.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome