Does a model learn from my conversation while I use it?

A common worry, and a common misunderstanding: does the model learn from what I tell it? In the usual sense, no. While you are using a model, its weights are frozen. Your conversation does not rewrite them. Anything the model “remembers” later comes from a separate training run on collected data, not from your chat updating it live.

There is one twist that looks like learning but is not. If you put a few examples directly in your prompt, the model can pick up the pattern and follow it for that answer. The GPT-3 authors showed this clearly: the model was “applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model.”^[1] The examples steer the answer, but no weights change, and the effect vanishes once the prompt is gone. Researchers call this specific phenomenon in-context learning — not because any learning in the weight-changing sense is happening, but because the model’s behavior visibly adapts within a single context window, then resets completely the moment that window closes.

This wasn’t a minor trick on a small model. GPT-3 had 175 billion parameters, “10x more than any previous non-sparse language model,”^[1] and the paper’s headline finding was that scaling up alone let this prompt-only approach get competitive with methods that actually fine-tuned their weights on the task.^[1] The model never learned your specific task. It got good enough at the general pattern of “here are some examples, continue it” that a live weight update stopped being necessary.

So it helps to keep two ideas apart. Real learning means changing the weights, which happens during training. Adapting to your prompt happens during inference and lasts only for that one request. The model that answers your next question is exactly the same model that answered your last one.