On this page
Training means adjusting a model using examplesAn example is something the model can practise onA simple learning loopWhy new test examples matterStarting training and continuing training are differentTry it: separate learning from usingSourcesTraining uses examples and a learning procedure to adjust a model’s values. Using a downloaded model for chat usually does not repeat that training process.
3Blue1Brown: Gradient descent, how neural networks learn
The opening examples show predictions being checked and learned values being adjusted. The later mathematics is optional. Gradient descent is the name of one method for choosing adjustments; this lesson explains the basic idea without formulas.
Watch on YouTube or turn on captions ↗Here is what matters, why it matters and how to check it on your own setup.
Training means adjusting a model using examples
Think about practising spelling. You try an answer, compare it with the correct spelling, and use the result to improve your next attempt.
Training a model also uses a procedure to change its adjustable values. In a common training method, the system makes a prediction, measures an error, and updates those values. It repeats the process over many examples.
The spelling comparison is an analogy. A computer follows a mathematical procedure; it does not practise in the same way a child does.
An example is something the model can practise on
For a model that recognises handwritten numbers, an example might be a picture labelled “3”. The system can compare its guess with that label.
A language model can learn from text by predicting a missing or next text piece. The text itself supplies the piece to compare with the prediction. People do not have to write a question and answer for every training example.
Different training methods use different goals. This lesson describes common examples, not a claim that every AI system learns in exactly the same way.
A simple learning loop
Imagine a toy model guessing a number. It guesses 7, while the example says 5. A training procedure measures that mismatch and chooses an adjustment.
Real systems use a numerical error measure, often called loss. A smaller training error can be a useful sign, but it is not proof that the model will perform well on new work.
The numbers 7 and 5 are made up for this explanation. They are not results from testing a real language model.
A simplified training loop. The choice of adjustments depends on the training method; improvement is checked rather than assumed.
Why new test examples matter
A model can do well on examples used for practice and poorly on new ones. Researchers call this overfitting. In plain words, it learned the practice set too closely instead of finding useful general patterns.
A fair check therefore uses examples kept separate from training. If the model saw the exact test answers during practice, a high test score may be misleading.
For your own AI project, keep some ordinary tasks aside to check later. Do not claim success from repeating one example the model has already been shown.
Starting training and continuing training are different
Training from the beginning builds a model’s learned values through a large learning process. Fine-tuning means taking an existing trained model and doing more training for a narrower purpose.
Writing a better prompt is different again. You change the input, not the model’s learned values. Giving a document to a chat app is also not, by itself, proof that the model was retrained.
You do not need to train a model just to use local AI. A downloaded trained model is normally the starting point. Whether further training is worthwhile depends on the task and evidence.
Try it: separate learning from using
Sort these actions into two groups: “changes learned values” and “uses existing learned values”.
Action A: a training program updates values after checking example predictions. Action B: you ask a downloaded model to rewrite a harmless note. Action C: you make the request clearer and ask again.
A belongs in the first group. B and C belong in the second during ordinary chat use. If an app specifically offers a training feature, read its documentation before assuming what it does.
Quick check: open the recap
Does asking 100 questions automatically train your downloaded model? No. Ordinary chat and training are separate processes.
Finished this lesson?
Mark it complete when you have read the lesson and tried the exercise. This saves progress on this browser. It is your own assessment, not a test score.
Sources for this lesson
Official documentation supports product and model facts. Community discussions show real setups, failures, and questions. A community result is supporting evidence, not a promise that another computer will perform the same way.
Watch a real test. Check what fits your computer.
See practical local AI tests from DeviceTerra, then use MamiLens to build a hardware-aware shortlist for your own setup.
This guide is educational. Model software, licenses, and hardware support can change. Check official sources before an important deployment.