Practice changes the weights

During language-model pretraining, a system predicts tokens from examples. A loss measures prediction error. An optimizer uses gradients to adjust weights across many batches. Later training can adapt behaviour using demonstrations or feedback.

One sentence, two different jobs

A simplified learning loop
Predict→Measure error→Adjust weights

In training, a sentence such as “the sky is blue” supplies a target continuation. During inference, your prompt supplies context; the trained model generates a continuation using its existing weights.

A conversation is not instant retraining

Remembering an earlier message within a conversation does not require changing weights. An application's saved memory is also separate from model training. Whether a provider later uses conversations for training depends on its product and data policies.

KEEP EXPLORINGMake sense of another idea↗