Artificial intelligence has too many names. AI, machine learning, deep learning, generative AI, foundation models, and prompt engineering often appear in the same discussion. The problem is simple: these terms describe related systems, but they do not mean the same thing.
This lesson builds a clear map. It explains how the parts fit together, how data becomes a model, and where prompts enter the process.
The AI family tree
Artificial intelligence is the broadest term. It covers software that performs tasks linked with human intelligence. These tasks can include recognizing objects, making decisions, finding patterns, or understanding language.
Machine learning is one way to build AI. Instead of writing every rule by hand, an engineer gives an algorithm data. The algorithm learns useful patterns from that data. The result is a model that can make predictions or decisions about new data.
Deep learning is a part of machine learning. It uses neural networks with many layers. These layers transform input data step by step. For example, a vision model may learn edges first, then shapes, then parts of an object.
The brain analogy is useful at a basic level. Neural networks have connected units, much like a simplified network of neurons. The analogy has limits, though. Modern neural networks are mathematical systems, not digital copies of a human brain.
Generative AI is another part of deep learning. It creates new content from patterns learned during training. That content can include text, images, audio, video, or code. A generative model does not retrieve every answer from a stored document. It produces an output based on its learned representation of data and the input request.
A simple map looks like this:
- Artificial intelligence includes machine learning.
- Machine learning includes deep learning.
- Deep learning supports many generative AI systems.
- Prompt engineering helps people communicate with generative models.
These categories overlap, but they are not interchangeable. Calling every AI system “generative AI” creates confusion. A model that detects a damaged part is using AI. It may use machine learning or deep learning. It is not necessarily generating new content.
From data to a working model
Machine learning starts with data. The data may be a table, an image collection, a set of recordings, or text messages.
Structured data has a fixed layout. A spreadsheet is a familiar example. Each row may describe one item, and each column may hold a known value, such as size, price, or category.
Unstructured data does not follow a simple table layout. Images, video, audio, and free-form messages belong here. This data can still contain strong patterns, but those patterns are harder to describe with ordinary columns and rules.
The basic workflow has four parts:
- Collect suitable data.
- Give the data to a learning algorithm.
- Train the algorithm to detect useful patterns.
- Use the trained model on new input.
Training does not mean the model understands the world as a person does. It means the model adjusts internal values so its output fits the examples or signals used during training.
The quality of the result depends on the task, the data, the labels, the training method, and the evaluation process. More data does not repair every problem. Bad labels can teach a model bad habits at great speed.
Three ways models learn
Supervised learning uses labeled examples. Each input has a known answer. An image may have a label such as “dog” or “phone.” The model learns a connection between the input and the label. Later, it can classify an image it has not seen before.
Unsupervised learning uses data without answer labels. The model looks for repeated patterns or groups. This can help reveal clusters in a large collection. The groups may be useful, but a human still needs to decide what they mean.
Reinforcement learning uses goals and feedback. The system takes actions and receives a score or signal. Good results strengthen some actions. Poor results weaken them. Over time, the system learns actions that improve its score.
These methods solve different problems. A labeled image task fits supervised learning. Finding groups in unknown data fits unsupervised learning. Learning a sequence of actions toward a goal fits reinforcement learning.
The method does not replace problem design. A clear goal still matters. A system can optimize the wrong target with impressive efficiency, which is a polite way of saying that it can fail very quickly.
One small example
Consider a system that identifies phones in images.
With supervised learning, the training set contains images and labels. One image shows a phone, and its label says “phone.” Other images show different objects with their own labels. During training, the model adjusts itself to connect visual patterns with the correct categories.
After training, a new image enters the model. The model examines its learned patterns and produces a prediction. It may identify the object as a phone.
That prediction is not a promise of truth. The image may be blurry. The object may be partly hidden. The model may have seen too few examples of that style or setting. The output needs evaluation against the real task.
Now consider an unsupervised version. The images have no labels. The system may group similar images together. One group may contain phones, but the system does not know that name by itself. It has found a pattern, not a human explanation.
Reinforcement learning changes the task again. Instead of classifying images, an agent might control a system that must reach a destination. It receives feedback about progress and learns actions that improve the result.
Foundation models and prompts
A foundation model is a large, general model trained on broad data. It can support many later tasks. A language model can produce summaries, answer questions, transform text, or write code. An image model can create or alter visual content.
The model is trained before a user writes a prompt. Prompting does not usually retrain the model. A prompt provides instructions or context for the current request.
Prompt engineering is the practice of writing those requests with enough detail for the task. A useful prompt often states the goal, the input, the desired format, and important limits.
For example:
Classify this image as phone, tablet, laptop, or other. Return one label and one short reason.
This prompt defines the task and controls the output shape. It does not guarantee a correct answer. The model still depends on its training, the image quality, and the clarity of the categories.
A vague prompt such as “What is this?” leaves more decisions to the model. That may be fine for exploration. It is weaker for a software system that expects a stable output.
Prompt engineering is therefore part of system design. It sits after model training and before the model produces a response. It can improve clarity and consistency, but it cannot fix missing knowledge, poor input data, or a model that is unsuitable for the task.
What this foundation changes
The main lesson is the relationship between the terms. AI is the broad field. Machine learning learns patterns from data. Deep learning uses layered neural networks. Generative AI creates new content. Prompt engineering shapes the request sent to a capable model.
You can now trace a basic AI system from data to output. You can also tell whether a problem needs labels, pattern discovery, feedback, or a generative model. That distinction prevents a common engineering mistake: choosing a fashionable tool before defining the task.
The Model Log follows this same standard: one practical AI concept, one working example, and one honest look at what actually works.
