First, the key terms.

Weight
A parameter that changes how much an input affects an output.
Loss
A numerical measure of error under a chosen objective.
Learning rate
A setting controlling the size of parameter updates.

The idea, explained

During training, an optimizer changes parameters to reduce a loss function. Gradient descent uses the slope of that function to choose an update direction. The learning rate controls the step size. A model with low training loss can still perform poorly on unseen data, so separate evaluation data matters.

STEP 1

Make a prediction

Use a toy model that predicts delivery time from distance: predicted minutes = weight × kilometres. With weight 2 and distance 3, it predicts 6 minutes. The weight is a parameter; the distance is an input. The model is intentionally simple so we can inspect every part.

STEP 2

Measure the error

If the observed delivery took 9 minutes, the prediction missed by 3. Squared error is (9 − 6)² = 9. The loss turns a mismatch into a quantity that an optimizer can use. A different loss would emphasize errors differently.

STEP 3

Update, then test separately

Training adjusts the weight in a direction intended to reduce loss. A weight of 3 gives 9 minutes for this example, but fitting one example is not enough. Test on other distances and traffic conditions. Inference then uses the selected weights to predict a new delivery without an automatic weight update.

Follow a worked example.

A model predicts 8 when a training example’s target is 10. With squared error, that example contributes (10 − 8)² = 4 to the loss. After training, using the model to predict the next example is inference; it does not automatically mean another parameter update.

Walk it through, one step at a time

  1. Start with weight 2, distance 4 and actual time 12. The prediction is 8 and squared error is 16.
  2. Trying weight 2.5 produces 10 and squared error 4 on this example. This illustrates improvement, not a universal best update rule.
  3. Now use the model on a new distance of 5. With weight 2.5, inference predicts 12.5 minutes. A later true outcome is needed to measure that prediction’s error.

Where people get stuck

More training steps or a lower training loss do not by themselves prove better real-world performance.

Now try it yourself.

Pause before opening each answer. Explain your reasoning in your own words.

1. Check the core idea

Is changing the wording of a prompt the same as retraining model weights?

Reveal the answer & reasoning

No. A prompt changes the input at inference time. Training changes parameters through an optimization process.

2. Apply what you learned

A model gives a better answer after you add detail to a prompt. Does this prove that training occurred?

Reveal the answer & reasoning

No. A more informative input can change an inference result while parameters remain unchanged. Training requires a process that updates parameters; conversation context and permanent model learning are different things.