A model has two completely separate phases in its life. Almost every misconception about AI comes from blurring them.
Phase one: training
Training is where the model is built.
It starts with a network whose weights are effectively random — it produces nonsense. Then, repeatedly:
- Show it an example from the training data.
- Compare its output to what the output should have been.
- Calculate how each weight contributed to the error.
- Nudge every weight slightly in the direction that reduces the error.
Repeat this billions of times, across an enormous dataset, and the weights settle into values that produce useful output.
This is extraordinarily expensive. It requires large clusters of specialised hardware running for extended periods. It happens once (plus later refinements), before you ever touch the model.
Phase two: inference
Inference is where the model is used.
The weights are now fixed. Your input goes in, flows through the network once, and an output comes out. Nothing is adjusted. Nothing is remembered.
Inference is cheap by comparison — that is why millions of people can use a model simultaneously.
The consequence people miss
The model does not learn from talking to you.
If you correct a model mid-conversation and it apologises and gets it right, nothing was learned in the sense that matters. Your correction became part of the text it can see — the context — and it influenced the rest of that conversation. The weights never moved.
Open a new conversation and the model is exactly as it was.
This has practical consequences:
- Repeating context is normal. A new conversation genuinely does not know what happened in the last one, unless the product deliberately stores and re-supplies it.
- "It learned my preferences" usually means storage, not training. Some products save notes about you and paste them into future context. That is a feature built around the model, not the model changing.
- Your corrections do not improve the product for others. Unless the provider separately collects and uses that data — which is a policy question about your data, not a property of the model.
Where fine-tuning fits
There is a middle case. Fine-tuning takes an already-trained model and continues training it on a smaller, specific dataset — adjusting the weights again.
That genuinely changes the model. But it is a deliberate, offline engineering process with a curated dataset. It is not something that happens because you told a chatbot it was wrong.
Checking your understanding
If a colleague says "I've been training the AI all week by giving it feedback", you now know what they most likely mean: they have been iterating on their prompts and building up useful context. That is a real and valuable skill — it is the subject of the prompt engineering course — but it is not training.