What Is Generative AI?

What You Will Learn

By the end of this lesson, you should be able to:

• Explain what generative AI means.
• Understand how generative AI differs from traditional predictive and classification systems.
• Understand why generative AI is not limited to chatbots.
• Identify the major types of generative AI.
• Understand how language models generate text.
• Understand how image-generation systems work at a high level.
• Understand the basic idea behind diffusion models.
• Understand what multimodal generative AI means.
• Understand the difference between a prompt, a model, and generated output.
• Understand why the same model can produce different outputs from similar prompts.
• Understand concepts such as sampling, temperature, and probability at a conceptual level.
• Understand why generated content can be impressive while still being incorrect.
• Understand the role of data, training, and model architecture.
• Understand how generative AI connects to local AI and Loco.

3.1 The Word "Generative" Matters

In the previous lesson, we established that artificial intelligence is a broad field and that modern AI contains many different types of systems.

Now we need to understand one of the most important categories in the current AI landscape:

Generative AI.

The word "generative" comes from the fact that these systems can generate new outputs. That output might be:

• text
• images
• audio
• music
• video
• computer code
• 3D assets
• structured data
• synthetic examples
• combinations of several modalities

A traditional AI system might answer a question such as:

"Is this email spam?"

The system could respond:

Spam

or:

Not spam

A generative system can instead receive:

"Write a professional reply to this customer complaining about a delayed delivery."

and produce an entirely new piece of text. That distinction is fundamental. A classification system selects or predicts among known categories.

A generative system produces an output. The output is not necessarily something that existed as a complete object inside the training data. The model has learned statistical patterns that allow it to generate a new result.

That is the basic idea behind generative AI.

3.2 Generative AI Is Much Bigger Than Chatbots

When most people hear "generative AI," they immediately think of conversational AI. That makes sense. Language models have become one of the most visible forms of generative AI.

But generative AI is much broader. Consider these examples:

Text generation

Prompt: Write a short story about a scientist discovering life on Mars.

Output: A newly generated story.

Image generation

Prompt: A medieval city built into the side of a giant mountain.

Output: A newly generated image.

Code generation

Prompt:
Create a Python function that converts Celsius to Fahrenheit.

Output:
Generated source code.

Audio generation

Prompt: Generate a short cinematic orchestral soundtrack.

Output: Generated audio.

Video generation

Prompt: A drone flying over a futuristic African city at sunset.

Output: Generated video.

3D generation

A system may generate or assist in generating 3D objects, environments, or assets. These applications use different architectures and techniques.

There is no single "generative AI algorithm."

Generative AI is a category of systems that produce new content.

3.3 Generation vs Prediction

Here is an important idea that initially sounds contradictory:

Generative AI systems often generate content by making predictions.

This is especially important for language models. Suppose you type:

The cat sat on the

A language model might calculate probabilities such as:

mat 35% floor 18% chair 7% bed 6% table 4% roof 1% ...

These numbers are illustrative, not actual model probabilities. The model chooses a token according to its generation strategy. It might choose:

mat

Now the sequence becomes:

The cat sat on the mat

The model then predicts what comes next. Perhaps:

because

Then:

it

Then:

was

Then:

warm

And so on. An entire paragraph can emerge through a sequence of predictions. This is one of the most important ideas in modern language AI:

A system can generate complex content by repeatedly predicting what should come next.

That doesn't mean every generative model works this way. Image generators and other generative systems use different mechanisms. But for autoregressive language models, this idea is foundational.

3.4 Why Doesn't Prediction Mean the Model Is Just Copying?

This is an important question. If a language model predicts the next token, does it simply copy text from its training data? Not necessarily.

During training, the model learns statistical relationships represented within its parameters.

The resulting model does not simply function as a conventional database containing a searchable copy of every training example.

Instead, the training process changes the model's numerical parameters. Those parameters encode learned patterns.

As a result, the model can generate combinations of information that may not have appeared as exact passages in the training data.

For example, you might ask a model:

"Write a 300-word explanation of gravity using a pirate metaphor and a South African cooking analogy."

The exact combination may never have appeared in the training data. Yet a sufficiently capable model can generate something that satisfies the request. This is possible because the model has learned patterns that can be recombined.

That does not mean the model has human-like creativity. But it does mean that generative models can produce novel outputs through learned statistical representations.

3.5 Generative AI Starts With a Model

Before a generative AI system can generate useful content, there must be a trained model. The model contains learned parameters. The parameters represent patterns learned during training.

A simplified pipeline looks like:

This distinction is important. The training process creates the model. The generation process uses the model.

These are different stages. For local AI, you are generally working with the second stage. Loco can load an already-trained model and use your local computer to perform inference.

You are not normally training the model from scratch every time you use it.

3.6 The Three Main Pieces

When you interact with a generative AI system, it can be useful to think about three major pieces:

1. The Model

The learned mathematical system.

2. The Input

The information you provide. This could be:

• a text prompt
• an image
• an audio clip
• a document
• a combination of modalities

3. The Generation Process

The computational process that transforms the input into an output using the model. Conceptually:

An application such as Loco provides the interface and surrounding software that connects these pieces.

3.7 What Is a Prompt?

A prompt is the input given to a generative model to guide its output. For a language model, that might simply be:

Explain black holes.

Or it might be extremely detailed:

Explain black holes to a 15-year-old student using a simple analogy involving a trampoline. Avoid equations. Include three examples and finish with five quiz questions.

Both are prompts. The difference is the amount of information and constraints provided to the model. Prompting can therefore be thought of as a form of communication with the model.

But there is an important caveat. A prompt is not programming the model's parameters. When you write a prompt, you are providing input to the already-trained model.

You are not changing the underlying learned weights. This distinction becomes extremely important when we later discuss:

• prompting
• system instructions
• context
• RAG
• fine-tuning
• LoRA

These are different mechanisms.

3.8 What Happens When You Send a Prompt to a Language Model?

Let's follow a simple request. You type:

"Explain how a solar eclipse works."

A simplified process looks like this:

The process repeats until the model reaches an appropriate stopping condition. The actual mathematics inside the model is enormously more complicated than this diagram. But this is the conceptual flow.

Notice something important: The model doesn't generate the entire answer in one magical operation. For an autoregressive language model, generation generally happens one token at a time.

That is why local AI performance is often discussed in terms such as:

tokens per second

We will later explore exactly what that measurement means.

3.9 What Is a Token?

A token is a unit of text processed by a language model. A token might correspond to:

• an entire word
• part of a word
• punctuation
• whitespace patterns
• a number
• a fragment of a word

For example, the sentence:

"Local AI is fascinating."

might be broken into a sequence of tokens roughly resembling:

Local AI is fascinating .

The exact tokenisation depends on the model's tokenizer. The model doesn't directly process ordinary human-readable characters as we experience them. The text is transformed into numerical token IDs.

Conceptually:

Tokens will receive their own dedicated lesson later because understanding them is essential to understanding:

• context windows
• inference speed
• memory usage
• prompt length
• model performance

For now, remember:

Language models operate on tokens, not directly on human concepts or raw sentences.

3.10 Why Does a Language Model Generate One Token at a Time?

This design is one of the most important concepts in generative language AI. Suppose the input is:

The capital of France is

The model processes the sequence and calculates a probability distribution over possible next tokens. Perhaps the model assigns high probability to:

Paris

Once that token is generated, the sequence becomes:

The capital of France is Paris

The model can then generate the next token. For example:

.

The process continues. For a long response, this can happen hundreds or thousands of times. Conceptually:

This is called autoregressive generation. The model uses previously generated information as part of the input for subsequent predictions. This is one reason long responses require more computation than short responses.

3.11 Probability Is at the Heart of Generation

At the end of a forward pass, a language model produces numerical scores representing how likely different tokens are to be selected.

The model is effectively producing a distribution over possible next tokens. Imagine:

"cat" 0.31 "dog" 0.21 "horse" 0.04 "house" 0.02 ...

Again, these are fictional values. The actual vocabulary can contain thousands or hundreds of thousands of possible tokens, depending on the model. The generation system then uses a decoding strategy to select the next token.

That means generation isn't simply:

"The model knows the answer and prints it."

Instead, the system is repeatedly performing mathematical calculations and selecting outputs according to probabilities and generation settings.

This is one reason the same prompt can sometimes produce different responses.

3.12 What Is Temperature?

You may have encountered a setting called temperature. Temperature affects how the probability distribution is used during generation. A simplified conceptual interpretation is:

Lower temperature

The model tends to favour higher-probability outputs. This can make responses more predictable and consistent.

Higher temperature

The generation process can allow lower-probability options to be selected more often. This can increase variation. Imagine the model calculates:

Option A → 70% Option B → 20% Option C → 7% Option D → 3%

At a very conservative generation setting, Option A is strongly favoured. At a more exploratory setting, the lower-probability options may have greater opportunities to appear. This can make output more varied.

However, temperature is not a simple "creativity slider."

Increasing temperature does not automatically make a model more intelligent or creative. It changes the stochasticity of token selection.

3.13 What Is Sampling?

Sampling refers broadly to the process of selecting an output from a probability distribution.

Suppose the model produces:

A → 0.50 B → 0.30 C → 0.15 D → 0.05

A deterministic strategy might always choose A. A sampling strategy can select among the possibilities according to their probabilities. This introduces variation.

That's why you can sometimes ask the same model the same question twice and receive slightly different wording. The underlying model hasn't necessarily changed. The generation process may simply have made different choices.

There are many decoding techniques. Some common concepts include:

• temperature
• top-k
• top-p
• greedy decoding
• repetition penalties

We will cover these in much greater detail later.

3.14 Why Can the Same Prompt Produce Different Answers?

Now you can see the answer. There are several possible reasons.

Sampling

The generation process may select different tokens.

Context

The conversation history may be different.

System instructions

The model may receive additional instructions that the user doesn't see directly.

Model configuration

Different settings can influence generation.

Randomness

Some generation pipelines intentionally introduce stochasticity. This is why a generative AI system isn't necessarily deterministic. A calculator should return the same answer to:

2 + 2

every time. A generative model might answer:

"The capital of France is Paris."

and later:

"Paris is the capital city of France."

Both answers can be correct. Generative AI is not simply a database lookup system.

3.15 Text Generation Is Only One Type of Generative AI

Now let's move beyond language. Another major class of generative models is image generation. Image generation systems can produce images from prompts such as:

"A futuristic Durban skyline at sunset, cinematic photography."

The underlying mechanisms can be very different from those used by language models. One major family of image-generation approaches uses diffusion models. Understanding diffusion at a conceptual level is useful because it demonstrates something important:

Different types of generative AI can use completely different generation mechanisms.

3.16 What Is a Diffusion Model?

The idea behind diffusion can initially sound strange. Imagine taking a clean image. Then gradually adding noise to it.

After enough steps, the image becomes almost entirely noise. Conceptually:

During training, a diffusion model learns how to reverse this process.

Instead of starting with a clean image and adding noise, generation starts from noise and progressively transforms it into a structured image.

Conceptually:

The actual mathematics is much more sophisticated than this description. But this gives you the core idea. A diffusion model learns a process that can transform noise into structured content.

3.17 Why Does Noise Become an Image?

This is where the learned model matters. The model has been trained on enormous numbers of examples. During training, it learns statistical relationships between visual patterns.

It learns representations associated with things such as:

• shapes
• textures
• edges
• colours
• objects
• lighting
• compositions
• visual relationships

The model does not need to store one perfect image of every possible thing. It learns a mathematical representation of patterns.

During generation, the model uses those learned relationships to guide the transformation from noise toward an image matching the conditioning information.

This is why a prompt can produce:

"A red fox standing in snow."

even if the exact image never existed in the training data. The model has learned patterns associated with foxes, snow, colours, composition, and relationships between them.

3.18 Language Models and Diffusion Models Are Different

This is worth emphasising. A language model and an image diffusion model are both generative AI systems. But they don't necessarily generate content in the same way.

A simplified comparison:

This is why "AI model" is such a broad term. Different models solve different problems using different architectures.

3.19 What Is Multimodal Generative AI?

Now we can introduce multimodal AI. "Modalities" are different types of information. Examples include:

• text
• images
• audio
• video

A multimodal model can work across more than one of these. For example:

You might provide a photograph of a broken machine and ask:

"What components can you identify?"

The model processes both the image and your text. Some systems can also generate across modalities. For example:

The exact capabilities depend on the model.

This is another reason modern AI should not be thought of simply as "a chatbot."

It is increasingly becoming a general-purpose interface for different forms of information.

3.20 What Is Code Generation?

Computer code is another form of generated content. A language model can be trained on programming languages and related technical material. You might ask:

"Write a PHP function that validates an email address."

The model can produce code. This is still generative AI. The model is generating a sequence of tokens that form a program.

This is an important example because it shows that language models are not restricted to natural language. Programming languages are also sequences of structured tokens. The model can therefore learn statistical relationships between:

• keywords
• syntax
• functions
• variables
• libraries
• programming patterns
• documentation
• code structure

However, generated code should still be tested. A model can generate syntactically valid code that is logically incorrect. It can use a deprecated API.

It can misunderstand the requirements. It can introduce security vulnerabilities. It can invent functions that do not exist.

Generative AI accelerates programming, but it doesn't eliminate software engineering.

3.21 Generative AI Is Not the Same as "Creative AI"

People often use the word "creative" when discussing generative AI. That can be useful conversationally, but it can also create confusion. Generating something new doesn't necessarily mean the system has human creativity.

A generative model can combine learned patterns into novel outputs. But whether that constitutes creativity in the philosophical or human sense is a much deeper question. For Loco Academy, we don't need to resolve that philosophical question.

The technical fact is simpler:

Generative models can produce outputs that were not explicitly specified as individual outputs in their programming.

They do this through learned mathematical representations and a generation process.

3.22 Why Can Generative AI Hallucinate?

One of the most important limitations of generative AI is hallucination.

A hallucination occurs when a generative model produces information that appears plausible but is inaccurate, unsupported, or fabricated.

This can happen because generation is not identical to factual database retrieval. A language model is trained to model patterns in language and generate likely continuations. That does not guarantee that every generated statement corresponds to a verified fact.

Imagine a model receives:

"Who invented the fictional device called the Quantum Spoon?"

If the model has no reliable information about such a device, it may still generate an answer that sounds plausible. It could produce:

"The Quantum Spoon was invented in 1987 by..."

The answer might be complete nonsense. The model generated a plausible continuation. It didn't necessarily perform a database lookup and discover that the premise was fictional.

This is why users need to understand the distinction between:

plausibility and truth.

They are not the same thing.

3.23 Why Bigger Models Don't Eliminate Hallucinations

It is tempting to think:

"Just make the model bigger and hallucinations disappear."

Reality is more complicated. Larger and better-trained models can become more capable and may reduce certain types of errors. But generative models can still produce incorrect information.

The problem isn't simply a lack of parameter count. It can involve:

• uncertainty
• incomplete knowledge
• ambiguous prompts
• weak grounding
• retrieval failures
• context limitations
• generation dynamics
• training-data limitations

This is why modern AI systems often combine generative models with external information sources. That leads us eventually to RAG, or Retrieval-Augmented Generation. RAG can provide a model with relevant information from documents or databases before generation.

We will dedicate an entire section of the Academy to this.

3.24 Prompting Does Not Magically Reprogram the Model

There is another common misconception. Someone might write:

"You are now an expert nuclear physicist with 50 years of experience."

and believe they have fundamentally changed the model. They haven't. The prompt provides context and instructions.

It can significantly influence the model's behaviour. But the underlying parameters remain the same. This distinction is important:

versus:

Prompting and training are fundamentally different processes. Later, we'll examine fine-tuning and parameter-efficient techniques such as LoRA.

3.25 What Makes a Prompt Good?

A useful prompt often provides enough information for the model to understand:

• the task
• the context
• the desired format
• constraints
• relevant information
• the intended audience

Compare:

Write something about AI.

with:

Explain local AI to a small-business owner who has never used an AI model. Use plain English, explain why privacy can matter, give three practical business examples, and finish with a short checklist for getting started.

The second prompt gives the model considerably more information. This doesn't mean longer prompts are always better. A prompt can become unnecessarily complicated.

The goal is clarity. Good prompting is essentially good communication with a probabilistic language system.

3.26 What Is Context?

Context is the information available to a model when it generates an output. That can include:

• your current prompt
• previous conversation messages
• system instructions
• retrieved documents
• tool outputs
• other information supplied by the application

For example:

This becomes incredibly important for Loco.

If you provide a local model with a collection of company documents, the model can potentially use information from those documents as part of its context.

The model's context capacity is limited. That is why context windows matter. You will get an entire lesson dedicated to context windows later.

3.27 The Difference Between Knowledge and Context

This is a subtle but important distinction. A model's learned parameters contain patterns acquired during training.

That is sometimes casually described as the model's "knowledge."

Context is information supplied to the model during use. For example, imagine you download a model trained months ago. You then give Loco a company policy document created yesterday.

The model's parameters didn't suddenly get retrained on that document. Instead, the document can be supplied as context. Conceptually:

This distinction becomes critical when we study RAG.

3.28 Generative AI vs Search

Generative AI and search engines are not the same thing. A search engine typically retrieves existing information from an index. A generative model produces an output based on its learned representations and current context.

Imagine asking:

"What is the current price of gold?"

A search or live-data system can retrieve current information. A standalone language model may not know the current price. It might generate a plausible-looking answer that is outdated.

This is why AI systems often need tools and external information sources when dealing with current data. Local AI does not automatically mean that a model has access to the internet. A local model can be completely isolated.

If you want current external information, you need an appropriate data path. This distinction becomes important when designing local AI workflows.

3.29 Generative AI and Agents

Another concept we will explore later is the AI agent. A basic generative model might do this:

An agentic system can involve more steps:

The generative model remains an important component. But the surrounding software gives it additional capabilities. For example, a local AI system could potentially be connected to:

• files
• databases
• calculators
• APIs
• business systems
• local applications
• workflows

This is one of the directions in which Loco can become much more than a chat interface.

3.30 Why Generative AI Is Such a Big Deal

The significance of generative AI isn't simply that computers can write paragraphs. The bigger change is that natural language can become an interface to computation. Historically, interacting with computers often required learning a specific interface:

Menu Button Form Command Programming language API

Generative AI allows a user to express an intention in ordinary language. For example:

"Take these five documents, find the pricing differences, put them into a table, and explain which clauses changed."

That is a much more flexible interface. The model can interpret the request and coordinate the generation process. This doesn't mean natural language replaces traditional software.

Instead, it can become another interface layer on top of software. That is one of the most important ideas behind AI agents and AI workflows.

3.31 Why Local Generative AI Matters

Now we arrive at the Local AI angle. Generative AI doesn't have to happen in a remote data centre. A sufficiently capable model can be executed on local hardware.

The general architecture becomes:

This can provide advantages such as:

Privacy

Sensitive information can potentially remain within the local environment.

Control

The user can choose models and configurations.

Offline capability

Some workflows can continue without internet access.

Predictable access

You don't necessarily need a remote AI service available for every inference request.

Local customisation

The system can be integrated with local documents, files, applications, and business processes. However, local AI also introduces challenges:

• hardware requirements
• model compatibility
• memory limitations
• performance
• installation
• configuration
• security
• updates
• storage

Loco's role is to make those complexities more manageable.

3.32 A Practical Example: Generative AI in Loco

Imagine you have a folder containing:

Company Handbook.pdf Refund Policy.pdf Pricing.pdf Product Catalogue.pdf Customer Support Procedures.pdf

You ask:

"What is our refund period for damaged products?"

A basic language model without access to those documents may not know. A Loco-based workflow could potentially provide the relevant information to the model through a local knowledge system. Conceptually:

The language model is generating the answer. The retrieval system is providing relevant information. Loco is orchestrating the overall experience.

This is the beginning of understanding how real AI systems are built.

3.33 Generative AI Is a Stack

By now, you should be starting to see a pattern. Generative AI isn't one thing. It is a stack of technologies.

A simplified language-generation stack might look like:

For an image-generation system, many of the components would be different. For an agent, there would be additional layers. For a RAG system, retrieval and embeddings would appear.

This is why Local AI is such a broad technical field. You aren't simply learning how to download a model. You're learning how the entire stack works.

3.34 The Most Important Mental Model

If you remember one thing from this lesson, remember this:

Generative AI is a class of AI systems that learns patterns from data and uses those learned representations to generate new outputs.

For language models, generation often works by repeatedly predicting and selecting the next token. For image generation, approaches such as diffusion can transform noise into structured visual content. For other modalities, other architectures and generation techniques may be used.

The underlying mechanisms differ. The common idea is generation.

3.35 Key Takeaways

1. Generative AI creates outputs.

It can generate text, images, audio, video, code, and other forms of content.

2. Generative AI is broader than chatbots.

Chatbots are only one application of generative models.

3. Language models often generate text autoregressively.

They repeatedly predict and generate tokens.

4. Tokens are the units processed by language models.

Human text is converted into token representations before model processing.

5. Probability is central to language generation.

The model produces a distribution over possible next tokens.

6. Sampling influences output variation.

Settings such as temperature can affect how outputs are selected.

7. Image generation can use different mechanisms.

Diffusion models are a major example.

8. Multimodal AI works across different types of information.

This can include text, images, audio, and video.

9. A prompt does not retrain a model.

It provides input and context that influence the model's generation.

10. Context is different from learned parameters.

Context is information supplied during use. Model parameters are learned during training.

11. Generative AI can hallucinate.

Plausible output is not automatically factual output.

12. Local generative AI performs inference on local hardware.

Loco provides a platform around local models and the surrounding AI functionality.

3.36 Knowledge Check

Question 1

What does "generative" mean in generative AI?

Answer:
It means the system can generate new content or outputs rather than simply classifying or retrieving predefined information.

Question 2

Does every generative AI model generate text?

Answer:
No. Generative AI can produce text, images, audio, video, code, and other types of content.

Question 3

How does an autoregressive language model generate text?

Answer:
It repeatedly predicts probabilities for possible next tokens and selects tokens according to its generation strategy.

Question 4

What is a token?

Answer:
A token is a unit of data processed by a language model. Depending on the tokenizer, it can represent a word, part of a word, punctuation, or another text fragment.

Question 5

What is temperature?

Answer:
Temperature is a generation parameter that influences the distribution used when selecting tokens, generally affecting how deterministic or varied generation can be.

Question 6

What is a diffusion model?

Answer:
A diffusion model is a class of generative model that can learn to reverse a noise process, allowing generation to proceed from noise toward structured content such as images.

Question 7

Why can generative AI hallucinate?

Answer:
Because generation is not the same as guaranteed factual retrieval. A model can produce statistically plausible language that is inaccurate or unsupported.

Question 8

Does prompting change the model's learned parameters?

Answer:
No. Prompting changes the input and context provided to the model. Training or fine-tuning changes model parameters.

Question 9

What is multimodal AI?

Answer:
AI that can work with multiple information modalities, such as text, images, audio, or video.

Question 10

Why is local generative AI useful?

Answer:
It allows generative inference to be performed on local hardware, potentially providing greater control, privacy, offline capability, and integration with local data and workflows.

3.37 A Preview of the Next Lesson

We have now discussed AI, machine learning, deep learning, generative AI, and the basic mechanics of generation. But there is one question underneath all of this:

What exactly is an AI model?

What is actually inside the file you download? Why can a model be called "7B" or "14B"? What are parameters?

What are weights? Why can a model file be several gigabytes? Why does changing the numerical precision change the memory requirement?

And why can the same underlying model exist in multiple different formats? These questions lead us into one of the most important sections of Local AI 101. In the next lesson, we'll open the black box and start examining the AI model itself.

Sources and Further Reading

• Vaswani et al., Attention Is All You Need: https://arxiv.org/abs/1706.03762
• NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0): https://www.nist.gov/itl/ai-risk-management-framework
• Google for Developers, Machine Learning Glossary: https://developers.google.com/machine-learning/glossary
• Ho et al., Denoising Diffusion Probabilistic Models: https://arxiv.org/abs/2006.11239
• Stanford HAI, 2026 AI Index Report: https://hai.stanford.edu/ai-index/2026-ai-index-report

Loco Academy: Lesson Status

Completed: Lesson 3: What Is Generative AI?

Next: Lesson 4: What Is an AI Model?

Course: Local AI 101