One training step: loss, backprop, optimizer
A · 10-second intuition Run the batch forward, measure how surprised the model was by the true next tokens, compute how each weight contributed, and nudge every weight a little.
Predict before you read
During a chat with a deployed model, are its weights updated?