Why Your First ML Model Can’t Talk Like ChatGPT

You train your first machine-learning model. It accepts study hours and sleep hours, then predicts whether a student will pass or fail.

Then a reasonable question appears:

“If this is AI, why can’t it answer questions like ChatGPT? Would more training data help?”

The answer reveals an important distinction: belonging to AI does not mean having every AI capability.

Let’s use this small student predictor to untangle AI, machine learning, deep learning, and ChatGPT—and understand what we actually build when we create an AI application.

Start with the task

Our student predictor has a specific job:

Inputs: study hours and sleep hours
                    ↓
          Trained decision tree
                    ↓
Output: a predicted pass-or-fail label

It is a classifier: a model that predicts which category an input belongs to.

During training, a decision-tree procedure used example records containing the two inputs and their recorded outcomes. It learned questions that help separate the categories.

We chose the method and supplied the examples. We did not manually write every final decision in the trained tree.

That makes this an example of machine learning.

The tiny dataset is useful for demonstrating the process. It does not establish that the model can reliably predict outcomes for real students.

AI, ML, and deep learning describe related categories

Here is the relationship:

Artificial intelligence
└── Machine learning
    └── Deep learning

These are nested categories. Studying deep learning means studying one part of machine learning, which is itself part of AI.

Artificial intelligence: the broad field

AI is the broad field of building systems for tasks associated with intelligence, including planning, reasoning, learning, and working with language.

Not every AI approach learns from examples.

For instance, an expert system could compare crop observations against knowledge and rules supplied by agricultural experts. Search algorithms can explore possible moves in a game.

Also, automatic operation alone does not establish AI. A kettle switching off at a fixed temperature is ordinary automation.

The useful questions are: What does the system do, and how does it make its decisions?

Machine learning: learning patterns from examples

In machine learning, a learning procedure finds patterns in data. The resulting model uses what it learned on new inputs.

Different ML models can perform different tasks:

Task	Example output
Classification	A pass-or-fail label
Numerical estimation	An estimated number of hotel bookings
Content generation	A generated written reply

Our student decision tree belongs here because it learned from example records.

Deep learning: a kind of ML

Deep learning uses neural networks with many layers.

These layers contain connected mathematical units. Each layer combines numbers and passes results onward. Training adjusts internal numerical values using examples.

For handwritten-digit recognition, an example might be an image of a handwritten 7 paired with the correct label. The network learns calculations that help predict the digit.

Our decision tree did not use a neural network. It was therefore:

AI → ML → decision tree, but not deep learning.

Deep learning is not automatically the best choice for every problem. The task, available data, computing resources, and measured results should guide the choice.

ChatGPT is a product

ChatGPT belongs in a different kind of description: it is a conversational AI product.

Deep-learning models provide its language capabilities. An interface and supporting software make those capabilities available to users.

That is why drawing ChatGPT as another smaller scientific category inside deep learning can be misleading.

A clearer picture is:

AI
└── Machine learning
    └── Deep learning
            │
            └── provides models used by products such as ChatGPT

A model provides a particular capability.

A product brings capabilities and supporting software together for users.

An LLM, or large language model, is a type of model trained to work with patterns across large amounts of text. The model and the product through which someone uses it are different things.

Why more student records won’t turn the predictor into a chatbot

Suppose we collect many more student records.

Those records might help us investigate or improve the same prediction task, depending on their quality and relevance. They do not automatically change the classifier’s inputs, outputs, or learning task.

The system still receives two numbers and predicts a category.

It has not been trained to interpret an English question and generate a written explanation.

More data alone does not give a model a different kind of capability.

Generating language requires a model suited to that task. Adding a chat box to the predictor’s interface would not change this.

Training, model, and application are separate responsibilities

This distinction becomes especially useful when building software.

Training is the learning process

For our example:

Student examples + decision-tree training procedure
                         ↓
                 Trained decision tree

The trained model is the result

We can save the trained tree, load it later, and use it to make predictions.

In the beginner lab, a prediction script loads the saved model and supplies study hours and sleep hours.

The application makes the capability usable

We could build a simple web application around it:

User enters two numbers
           ↓
Application validates the inputs
           ↓
Model predicts a category
           ↓
Application displays the result

The model performs the prediction. The application handles interaction and connects the components.

You can also build an AI application using an existing trained model. Training one yourself is not a requirement for every project.

You still need to check whether its capabilities and results suit your task.

A quick notebook exercise

No installation is needed.

1. Draw a large circle labelled AI.
2. Inside it, draw Machine Learning.
3. Inside ML, draw Deep Learning.
4. Place “Our trained student decision tree” inside ML but outside deep learning.
5. Beside the circles, write “ChatGPT: conversational product.”
6. Draw an arrow from deep learning to ChatGPT, labelled “provides models used by.”

Now explain two things in your own words:

• Why is the decision tree outside deep learning?
• Why does ChatGPT have an arrow rather than another nested circle?

If you can explain those relationships, you understand more than the definitions alone.

One final challenge

A developer says:

“I built an AI application, so I trained a deep-learning model.”

Does that conclusion follow?

Think about it before reading on.

No—and there are two separate assumptions to question.

First, the application might use an existing trained model.

Second, even if the developer trained a model, it could use an ML method other than deep learning.

Building an application, training a model, and choosing a deep-learning method are different decisions.

What to remember

• AI is the broad field.
• ML learns patterns from data and belongs within AI.
• Deep learning uses neural networks with many layers and belongs within ML.
• ChatGPT is a product using AI models and supporting software.
• A model’s task matters more than whether we call it “AI.”
• Building an AI application does not necessarily mean training a model.

This article accompanies Chapter 2 of my free AI for Beginners course. The video above explores these ideas through a conversation between Mira and Subra AI.

You can also explore the course on MetaWise.

When you built your first model, which distinction was hardest to understand: AI versus ML, ML versus deep learning, or model versus application?