AI Fluency Module 3: The Four Properties of AI | 4 Free Video & Text Lessons

AI FLUENCY · PHASE 1: AWARENESS

Module 3
The Four Properties of AI

Understand How Language Models Generate, Use, Remember, and Respond to Information

Why does AI predict the next word? What does it actually know? How does it use earlier messages? And why does it follow instructions?

Explore these questions through four short visual lessons and their detailed written explanations.

4 Video Lessons 4 Text Lessons 54 Seconds Each Free Learning

Welcome to Module 3

In Module 1, we explored the meaning of Artificial Intelligence and its presence in everyday life.

In Module 2, we discovered how AI learns from data, recognizes patterns, develops models, and generates language.

Now we examine four important properties of modern Large Language Models (LLMs).

Drawing on the educational framework introduced by Claude Academy, this module explores:

• Next-Token Prediction
• Knowledge
• Working Memory
• Steerability

Together, these properties help us better understand how generative language models behave and why their responses have both capabilities and limitations.

How to Learn

Each lesson combines a short, silent visual video with a detailed written explanation.

Watch: Begin with the 54-second video to understand the central idea.

Read: Open the published text lesson for explanations and examples.

Understand: Connect the concept with familiar everyday situations.

Reflect: Answer the questions to check your understanding.

The Four Properties at a Glance

1. Next-Token Prediction

A language model generates text by predicting successive tokens using learned patterns and available context.

2. Knowledge

Training develops learned representations of information, but generated answers may be incomplete, outdated, or wrong.

3. Working Memory

The context window contains information available to a model while generating its current response.

4. Steerability

Instructions, context, and feedback can influence how an AI model responds.

Watch and Read the Four Lessons

Follow the lessons sequentially to build a connected understanding of these four properties.

LESSON 009 · 54 SECONDS

Next-Token Prediction: How AI Generates Answers

How does an AI assistant construct a response to your question?

Many modern language models generate text by predicting one token at a time.

Each newly generated token becomes part of the context used to predict the next.

Example: "The capital of France is..." may be followed by "Paris."

Key Learning: Language models build responses through successive token predictions, but plausible text is not necessarily factually correct.

Read Full Lesson 009 → Watch on YouTube →

LESSON 010 · 54 SECONDS

Knowledge: What Does AI Actually Know?

Where does an AI assistant's information come from?

During training, language models learn statistical relationships, facts, and associations.

However, this does not mean that every generated claim is accurate or up to date.

Example: An LLM may correctly identify Rabindranath Tagore as the author of India's national anthem, but may lack information about a competition that ended minutes ago.

Key Learning: Learned knowledge differs from current retrieved information, and important factual claims require verification.

Read Full Lesson 010 → Watch on YouTube →

LESSON 011 · 54 SECONDS

Working Memory: What Is the Context Window?

How can an AI assistant use instructions you provided earlier in a conversation?

The context window contains information available for processing during the current response.

Its capacity is limited, and older details may not always remain available.

Example: If you ask AI to use simple language for 10-year-old students, that earlier instruction can influence a later explanation while it remains in context.

Key Learning: The context window acts as a form of working memory, but it is limited and distinct from training knowledge or separately saved memory.

Read Full Lesson 011 → Watch on YouTube →

LESSON 012 · 54 SECONDS

Steerability: Why Does AI Follow Instructions?

Why can AI explain the same topic differently depending on your instructions?

Steerability allows users to guide AI responses through instructions and context.

Audience, tone, format, length, and purpose can all influence the generated result.

Example: "Explain photosynthesis" may produce a general explanation. Adding "for a 10-year-old student" can result in simpler language.

Key Learning: Instructions can guide AI behavior, but instruction-following is not always perfect, and a well-formatted response can still contain errors.

Read Full Lesson 012 → Watch on YouTube →

How Do the Four Properties Connect?

The four properties are not isolated ideas. They help explain different aspects of how language models operate.

Knowledge

Training shapes the information and relationships represented within the model.

↓

Working Memory

The current context provides information and instructions for the response.

↓

Steerability

Instructions influence how the response should be generated.

↓

Next-Token Prediction

The model generates its response through successive token predictions.

This diagram is a conceptual summary, not a literal step-by-step representation of a model's internal architecture. These properties interact during language generation.

Check Your Understanding

After watching and reading the four lessons, try answering these questions.

1. How does a language model generate longer responses through next-token prediction?
2. Why might an LLM lack information about a very recent event?
3. How does a context window help an AI assistant follow earlier instructions?
4. How can changing an instruction influence the style and usefulness of an AI response?

If you can explain these ideas clearly in your own words, you have developed an important part of foundational AI awareness.

Module 3 Complete!

You have now completed the four learning lessons covering important properties of generative language models.

You understand how language models generate tokens, use training knowledge, process context, and respond to instructions.

12 AI Fluency Lessons Completed

Your journey through Phase 1: Awareness continues.

COMING NEXT

Module 4 — The AI Landscape

Next, we move beyond the fundamental properties of language models to explore the wider AI ecosystem.

We will investigate multimodal AI, chatbots, AI assistants, agents, search engines, and databases.

Lesson 013: What Is Multimodal AI?

Further Learning

This module draws inspiration from Claude Academy's Four Properties of AI framework.

Explore Claude Academy →

Explore OpenAI Academy →

The AI Fluency Journey

Our program progresses through four phases.

AWARENESS → APPRECIATION → APPLICATION → ACCOMPLISHMENT

You have completed three modules in the Awareness phase.

Explore the Complete AI Fluency Course Roadmap →

AIYING

Accomplishing with AI

Everything AI. For Everyone.

AI Fluency — From Awareness to Accomplishment

Created by Ravindra Dastikop

https://dastikop.blogspot.com

10. What Does AI Actually Know? Understanding AI Knowledge | AI Fluency Lesson 010

AI FLUENCY · PHASE 1: AWARENESS

Module 3 — The Four Properties of AI | Lesson 010

Knowledge: What Does AI Actually Know?

Explore how AI acquires information, what it can recall, and why its answers still need verification.

Watch this 54-second silent visual lesson. Then explore the explanation and examples below.

Watch on YouTube →

Introduction

When you ask ChatGPT, Claude, or Gemini a question, it may provide an answer almost instantly.

Sometimes that answer includes historical facts, scientific concepts, mathematical explanations, or information about the world.

This raises an important question: What does AI actually know?

To answer it, we need to understand how a Large Language Model (LLM) learns information, how it uses that information, and why its responses cannot always be trusted without verification.

Knowledge is the second property explored in Module 3 of AI Fluency.

1. How Does AI Acquire Knowledge?

During training, language models process large amounts of text and other data.

Through this process, they learn statistical patterns, factual relationships, concepts, and associations.

For example, training data may contain information about countries, historical events, scientific discoveries, languages, and literature.

The training process adjusts numerical values called model parameters. These parameters influence how the model generates responses.

In Simple Words

An LLM learns relationships from training data. Those learned relationships help it generate answers to new questions.

2. How Is Knowledge Represented in an LLM?

A language model is not simply a conventional database containing a separate document for every answer.

Much of what it learns during training is represented across its numerical parameters.

These learned representations help the model generate explanations, connect ideas, and respond to unfamiliar questions.

How Training Builds Model Knowledge

Training Data

Books, articles, examples, and other training material

↓

Model Training

Numerical parameters are adjusted to learn useful relationships.

↓

Trained Language Model

Learned representations influence the model's generated answers.

Simplified conceptual illustration. Actual LLM training and representations are more complex.

This is why an LLM can often respond to questions without performing a new web search.

However, generating a response from learned representations is not the same as checking an authoritative source.

3. Does AI Know Things Like Humans?

Humans develop knowledge through learning, observation, experience, reasoning, and interaction with the physical and social world.

An LLM processes information through learned computational representations.

These capabilities may support sophisticated answers and reasoning, but they do not establish that a model has human experiences or the same kind of understanding as a person.

Human Learning

Humans may learn through direct experience, observation, education, and interaction with the world.

LLM Training

Models learn computational representations from data and use them to process and generate information.

Understanding this distinction helps us avoid assuming that human-like language necessarily implies human-like knowledge.

4. An Everyday Example

Imagine asking an AI assistant:

Your Question

"Who wrote India's national anthem?"

The model may correctly answer:

AI Response

"Rabindranath Tagore."

This answer may reflect information learned during training.

However, a correct response does not necessarily mean the model consulted a reliable historical source at the moment you asked.

It may have generated the answer using relationships already represented in its parameters.

Important Point

An accurate AI-generated answer and a source-verified answer are not necessarily the same thing.

5. What About Recent Information?

Now imagine asking:

"Who won a competition that ended five minutes ago?"

A language model without access to current information may not have the result.

Its training data may not include events occurring after its knowledge cutoff.

Some AI assistants can use tools, including web search, databases, and connected information systems, to retrieve more recent information.

Learned Knowledge

Information represented through the model's training.

May be incomplete or outdated.

Retrieved Information

Information obtained from an external source through a tool or connection.

May be more current, but still requires source evaluation.

The important distinction is between information learned during training and information retrieved when answering a question.

Even access to live sources does not automatically guarantee that every response is accurate.

6. Can AI Knowledge Be Wrong?

Yes.

Language models may be trained on data containing errors, contradictions, biases, or outdated information.

They may also generate incorrect statements even when their training includes relevant and accurate material.

Sometimes AI produces fabricated facts or references. Such errors are commonly called AI hallucinations.

A fluent and confident answer does not establish factual accuracy.

Remember

AI-generated knowledge claims should be checked when accuracy matters.

7. How Can We Verify AI Answers?

An AI-fluent learner does not need to reject every AI answer. Instead, the learner develops good verification habits.

1. Identify the claim: What exactly is the AI saying?
2. Check reliable sources: Consult official publications, credible references, or primary evidence.
3. Check the date: Is the information current enough for your purpose?
4. Compare the evidence: Do the sources actually support the claim?
5. Use human judgment: Consider whether the information is suitable for the decision.

For important health, legal, financial, academic, or professional decisions, additional expert review may be appropriate.

Key Takeaway

Large Language Models learn information and relationships during training, but their generated answers may be incomplete, outdated, or incorrect.

Understanding the difference between learned knowledge, retrieved evidence, and verified facts is essential for AI Fluency.

Think About It

Imagine asking an AI assistant about a new scientific discovery.

The assistant gives a confident and detailed explanation.

How would you determine whether the answer is based on older training information, a current source, or an unsupported statement?

What evidence would help you trust the answer?

What Have You Learned?

After completing Lesson 010, you should be able to:

• Explain how training helps an LLM acquire information.
• Describe how learned knowledge differs from a conventional database.
• Distinguish learned knowledge from retrieved information.
• Recognize why a model may lack recent or specialized facts.
• Explain why factual AI claims sometimes require verification.

CONTINUE LEARNING

Next: Lesson 011

Working Memory: What Is the Context Window?

Now that you understand how language models acquire and use information from training, the next lesson explores their working memory.

You will discover what a context window is, how it influences a conversation, and why AI may lose track of information outside the available context.

Further Learning Resources

Claude Academy — AI Learning Resources

OpenAI Academy

Google Machine Learning Crash Course

Your AI Fluency Journey

AI Fluency develops through four connected phases.

AWARENESS → APPRECIATION → APPLICATION → ACCOMPLISHMENT

AI Fluency begins with understanding and progresses toward the ability to accomplish meaningful things with AI.

Explore the Complete AI Fluency Roadmap →

AIYING

Accomplishing with AI

Everything AI. For Everyone.

AI Fluency — From Awareness to Accomplishment

Created by Ravindra Dastikop

https://dastikop.blogspot.com