Lesson 2 of 6
How can software learn from examples?
Nobody wrote the garden reply in advance. Training is what lets a model respond to inputs it has never seen, and this lesson explains what training actually changes.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
Learning means changing how a calculation behaves
Computers work on numbers. Before a language model can do anything with a sentence, that sentence must be represented as numbers, and the model then performs calculations on those numbers. Inside the model are further numbers that control how strongly each part of the calculation affects the rest. These are called . This course usually calls them learned values, because the point of training is to set them from data rather than by hand. They are not the settings you choose in an application menu.
adjusts the learned values using examples. Learning here means changing a calculation using data; it does not mean gaining experience the way a person does. A common training pattern has three steps that repeat many times. The model makes a prediction. Software compares that prediction with a known target. A training procedure then adjusts the learned values with the aim of reducing the measured error. The size of the gap is measured by a quantity often called loss. Different tasks use different measures, and the chosen measure tells the procedure what counts as improvement. 2
A simplified training cycle
- 1 →PredictUse the current learned values
- 2 →CompareMeasure the difference from the target
- 03AdjustChange learned values, then repeat
A small example of an adjustable rule
Here is a deliberately simplified teaching example with invented records. We want to estimate how long it takes to prepare garden beds. Our records say one bed took 20 minutes and two beds took 40 minutes. Our tiny model has one rule: multiply the number of beds by an adjustable number of minutes per bed. That adjustable number is its only learned value.
Suppose it starts at 10 minutes per bed. For two beds the estimate is 20 minutes; the record says 40. The estimate is too low. Because the estimate grows when the minutes-per-bed number grows, the way to shrink the gap is to raise that number. The training procedure raises it, checks both records again, and keeps adjusting. In this example 20 minutes per bed fits both records exactly: one bed gives 20 and two beds give 40. You can check that arithmetic yourself, which is the point of using such a small example.
Match the pieces to the previous section. The minutes-per-bed number is the learned value. The two records are the training data. The comparison between estimate and record supplies the reason to change the value. Using the adjusted rule to estimate a new job, say three beds, is a separate operation from changing the rule, and it leaves the rule as it is.
Real gardens will not follow this rule. Hard soil, broken tools and experienced volunteers all change the time. If our records only cover easy beds, a rule fitted to them may underestimate difficult ones. This arithmetic example illustrates adjustment; it is not how a language model is built. A language model has vastly more learned values and a far more complex calculation, but the pattern of predicting, comparing and adjusting carries over.
Language models learn much richer patterns
A , or LLM, learns patterns in language from a large collection of text. “Large” describes the scale of the model and its training, including the enormous number of learned values; it does not mean the model only handles big or difficult questions. In an important part of training, the model is shown some text and trained to predict the piece that comes next. The original text supplies the target, so no person has to label each example. Repeating this across a very large amount of text adjusts the learned values toward patterns that help the model continue text. OpenAI’s description of its own models follows this outline: learned values set by training on broad data, with ChatGPT as an application built on the result. 6
Modern LLMs are built as neural networks: layered systems of connected calculations whose numerical values are adjusted during training. The name comes from a loose biological inspiration. It does not mean the software is a brain or works like one. 4
Predicting the next piece of text is not the same as answering a question. Broad text prediction alone does not ensure reliable instruction-following. Further training can favor useful responses to requests, so a model is more likely to write a reply when asked rather than merely extend the note. Anthropic’s model reports, for example, distinguish initial training on broad data from later training meant to produce an effective assistant. This is a developer’s account of its own process. It does not mean every model is trained identically, and a model trained to be helpful can still be wrong. 8
Doing well on old examples is not enough
A useful model needs to handle examples it was not trained on. This ability is called generalization. Developers check it using separate examples that were not used to adjust the model. One way this can go wrong is overfitting: a model fits details of its training examples so closely that it performs poorly on new ones. 3
New situations can also differ from the training examples. Our garden rule was fitted to easy beds; a rocky plot adds a condition those records never covered. Matching the two records does not tell us how well the rule handles that condition. The example shows why checking unfamiliar jobs matters, rather than proving a particular cause of failure.
Broad language learning helps a model produce a polite garden message. The details of this particular session still need to reach it. In our example, they come from the note.
Training data comes from the past
The learned values used for a response reflect earlier training. They do not automatically update when a new notice appears online. Suppose a model finished training months ago and the organizer first writes a cancellation notice this morning. That training could not include the new notice. To use it in a reply, the application needs to supply it as context.
A model description may state a knowledge cutoff to indicate the period its learned information is expected to cover most reliably. This is not a promise that it knows every earlier fact, nor necessarily an exact last date for every item in its training data. Anthropic's reports, for example, sometimes distinguish a training-data cutoff from a reliable-knowledge cutoff. The useful question is whether the system has the particular current information you need. 8
Developers can perform further training or change which model an application uses. Those are separate operations from supplying a new note in a conversation. Next we will see how information supplied now can affect a reply without changing learned values.
Carry this forward
- Training changes learned values, the numbers inside a model, by predicting, comparing with a target and adjusting.
- A training measure rewards what it was built to measure; it is not automatically a measure of truth or usefulness.
- A useful model must handle new examples. Its learned values reflect earlier training; current facts may need to be supplied, and a cutoff does not promise complete knowledge of earlier facts.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
Training sets the learned values. What happens when those fixed values meet a new question?
Sources · Reviewed 2026-09-11
- 2. Google: What training error measures
- 6. OpenAI: How its models are developed
- 4. Google: Neural networks
- 8. Anthropic: Model reports
- 3. Google: Generalization and overfitting
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.