From a Prompt to an Output
In the previous lesson, we learned what Generative AI is.
Now we will look at what happens when you actually ask a Generative AI system to create something.
A simplified workflow looks like this:
Prompt → Model Processing → Generation → Output
The exact process depends on the type of model, but this basic idea applies across many generative AI systems.
Step 1: The User Provides an Input
Everything begins with an input from the user.
This might be a question, instruction, description, image, audio recording, piece of code, or another type of information.
For example:
"Write a short explanation of machine learning for a beginner."
This instruction is commonly called a prompt.
What Makes a Good Prompt?
A prompt should provide enough information for the model to understand the task.
Compare:
"Write something about AI."
with:
"Explain artificial intelligence to a beginner in 150 words using one simple everyday example."
The second prompt gives the model more information about the subject, audience, length, and desired style.
Step 2: The Input Is Prepared
For a language model, the input text is generally divided into tokens.
The tokens are then converted into numerical representations that the neural network can process.
The model may also receive additional context, such as previous conversation messages or instructions supplied by the application.
Step 3: The Model Processes the Input
The trained neural network processes the available information.
Modern language models use architectures based on ideas such as attention to process relationships between tokens.
The model uses its learned parameters to determine what kinds of outputs are likely to fit the current input and context.
Step 4: The Model Generates an Output
For text generation, the model predicts a likely next token.
That token becomes part of the growing output.
The model then predicts another token.
This process continues repeatedly.
Eventually, the sequence of generated tokens forms a complete response.
A Simple Example
Suppose the prompt is:
"The capital of France is"
A language model may assign a very high probability to a token representing "Paris".
After generating that token, the system may continue generating additional tokens depending on the task.
This example is deliberately simple. Real language generation involves much more complex calculations across many layers of a neural network.
Generation Is More Than Selecting One Word
When a model generates a long response, it performs this process repeatedly.
Each newly generated token becomes part of the context used to generate subsequent tokens.
This means that the response develops step by step.
A paragraph may therefore be produced through many individual generation steps.
Why Can the Output Be Different Each Time?
Generative AI does not always have to produce exactly the same output for the same prompt.
Depending on the model and generation settings, the system can select between multiple possible tokens.
This allows variation in generated content.
For creative tasks, variation can be useful because it allows the system to explore different possibilities.
Probability and Choice
Imagine that a model assigns probabilities to several possible next tokens.
One token might have a probability of 60 percent, another 25 percent, and another 10 percent.
The generation system can use these probabilities when deciding which token to produce.
The exact selection process depends on the model and generation settings.
Temperature
One commonly discussed generation setting is temperature.
Temperature can influence how strongly the generation process favours higher-probability choices.
Lower settings generally make output more predictable, while higher settings can allow more variation.
The exact effect depends on the implementation and other generation settings.
Why Does Temperature Matter?
Imagine asking an AI system to generate ten possible names for a new company.
You may want variety.
A more varied generation setting can produce more diverse suggestions.
For a task where consistency is more important, a more conservative generation approach may be preferable.
Stopping Generation
The model does not necessarily generate forever.
The application can define conditions that tell the generation process when to stop.
For example, generation may stop when the model produces a special end-of-sequence token or when a maximum output length is reached.
Maximum Output Length
Applications can place limits on how much content a model generates in one response.
This can help control computing costs, response length, and user experience.
A short request might require only a small number of generated tokens, while a long document may require many more.
What About Images?
Image generation works differently from text generation, but the overall workflow has a similar structure.
The user provides an input such as a text description.
The image-generation model processes that input and generates visual content based on patterns learned during training.
The technical process can involve specialized architectures and mathematical representations that are different from language generation.
What About Audio?
Generative AI can also create or transform audio.
A system might generate speech from text, create music, or transform one type of audio into another.
Again, the underlying model and generation process depend on the specific task.
What About Code?
Code generation can be treated as another form of content generation.
A coding model can receive an instruction such as:
"Create a PHP function that calculates the total price including tax."
The model can generate a possible implementation.
However, generated code should be tested and reviewed because it can contain bugs, incorrect assumptions, or security weaknesses.
Prompt, Model and Output
It is useful to keep three concepts separate:
- Prompt: The input or instruction provided to the system.
- Model: The trained system that processes the input.
- Output: The content generated by the model.
Changing the prompt can change the output.
Changing the model can change the output.
Changing generation settings can also change the output.
The Application Around the Model
A Generative AI product usually contains more than the model itself.
The application may handle user accounts, prompts, conversation history, safety controls, file uploads, external tools, databases, and the user interface.
The model is therefore often one component inside a larger system.
External Information
A generative AI application can sometimes retrieve information from an external source before asking the model to generate an answer.
For example, an application could retrieve information from a company database and provide that information to the model.
The model can then generate a response using the retrieved information as part of its context.
Generation Does Not Guarantee Truth
The fact that a model can generate a fluent response does not mean the response is automatically correct.
The model may produce an answer based on patterns that seem appropriate but are not factually accurate.
This is one reason verification and reliable external information sources can be important.
Generation Does Not Mean Understanding in the Human Sense
People sometimes describe an AI model as understanding a prompt in the same way a human understands language.
This can be useful as a simplified description, but the underlying system performs mathematical operations over learned representations.
It does not mean that the model has human consciousness or human experience.
A Complete Simplified Workflow
For a text-based Generative AI system, we can describe the process as:
User Prompt → Tokenization → Numerical Representation → Neural Network Processing → Next-Token Prediction → Repeated Generation → Final Output
This is simplified, but it provides a useful mental model for understanding text generation.
Why This Is Powerful
The ability to generate content from instructions makes Generative AI flexible.
The same underlying model can potentially perform many different tasks depending on the input.
For example, a user can ask it to explain a concept, summarize text, brainstorm ideas, translate language, write code, or create a structured document.
Why This Is Also Challenging
Because the model can generate many types of outputs, users may assume that it is equally reliable at every task.
That is not necessarily true.
The quality of the result depends on the model, the prompt, the available context, the data, the task, and other factors.
The Big Picture
Generative AI transforms an input into generated content using a trained model.
For language systems, this involves tokens, context, neural network processing, probability, and repeated token generation.
For other media such as images and audio, the technical mechanisms differ, but the central idea remains the same: a trained model generates an output based on learned patterns and current input.
What You Should Remember
- A prompt provides the input or instruction.
- The model processes the prompt using learned parameters.
- Language models work with tokens and numerical representations.
- Text is generally generated one token at a time.
- Generation settings can influence how predictable or varied the output is.
- Generative AI can produce many types of content.
- The model is only one component of a complete AI application.
- Generated output should not automatically be assumed to be correct.
What Comes Next?
We now understand the basic process used to generate content.
But Generative AI can produce very different kinds of content depending on the technology being used.
In the next lesson, we will explore text, image, audio and video generation and understand how these different forms of Generative AI are used.