Neural networks learn by training: feed data in, predict an output, compare to the correct answer (that’s the loss), then adjust the weights so the loss goes down. The math that adjusts the weights is backpropagation. One pass forward (prediction) + one pass backward (gradients) + an update step is repeated millions of times.
Training loop (simplified)
Data
Text (or pairs) the model learns from
Tokenize
Turn text into token IDs
Forward
Model predicts next token
Loss
Compare prediction to correct answer
Backward
Compute gradients
Update
Adjust billions of parameters
Repeat over huge datasets for many steps. "Parameters" are the numbers being updated; more parameters = more capacity to memorize patterns.