Module 1: Introduction to AI, ML, Deep Learning, GenAI, and LLMs

Module Goal

Build a rock-solid mental model of the AI landscape so that every later module makes sense. By the end of this module you will be able to draw the AI to ML to DL to GenAI to LLM hierarchy from memory and explain when each one is the right tool.

Estimated Duration

3 to 5 hours.

Skills Learned

Distinguishing AI, machine learning, deep learning, generative AI, and LLMs
Reading AI news without getting confused by buzzwords
Identifying which problems suit which technique
Spotting marketing dressed up as "AI"

Real-world Importance

In 2026 every product team is "doing AI." Most of them confuse machine learning with LLMs and waste months. The first skill of an AI engineer is calling the right tool by its right name and choosing it for the right reason.

Lessons in this module

What is AI, really? (and what it is not)
The AI to ML to DL to GenAI to LLM hierarchy
Discriminative versus Generative AI in 60 seconds
The 2026 LLM landscape and how we got here
Where LLMs shine, and where they should not be used

Lesson 1.1: What is AI, really? (and what it is not)

Hook / Why This Matters

If you cannot define AI in one sentence, every later concept will feel slippery. AI is also the single most misused word in technology marketing today. Getting this right is your shortcut to seeing through hype.

Beginner Analogy

Think of a thermostat. It senses temperature and decides whether to turn the heater on. That is a tiny piece of intelligence in a machine. AI is the same idea, scaled up: a machine that senses, decides, and acts in a way that we would call intelligent if a human did it.

Concept Explanation

Artificial Intelligence is the broad goal of building machines that perform tasks which normally require human intelligence: understanding language, recognizing images, making decisions, planning, and learning from experience.

AI is the goal. Everything else you will hear about (machine learning, deep learning, LLMs) is a technique used to reach that goal.

Technical Breakdown

AI splits into two families based on how the intelligence is built:

Symbolic AI (rule-based): Humans write explicit rules. If X then Y. The chess engine Deep Blue (1997) was largely symbolic. Strengths: predictable, explainable. Weaknesses: brittle, scales badly.
Statistical / learned AI: The machine learns patterns from data instead of being told rules. Everything modern (ChatGPT, image recognition, recommendation systems) is in this family. Strengths: handles messy real-world data. Weaknesses: less interpretable, needs lots of data.

In 2026, "AI" almost always means the second category.

Visual Learning Suggestion

Insert a diagram here: a large circle labeled "AI", with two child circles inside labeled "Symbolic AI" and "Learned AI". Inside "Learned AI", a smaller circle "Machine Learning", and inside that "Deep Learning", and inside that "Generative AI", and inside that "LLMs". This nested-circles diagram is the most important visual of Module 1.

Interactive Element

Open ChatGPT or Gemini. Ask: "Give me three tasks where AI is genuinely useful and three where humans calling something AI is mostly marketing." Compare its answer to your own gut. Notice where you disagreed. That gap is where your intuition needs reps.

Hands-on Lab

No code yet. Do this writing exercise (15 minutes):

Open a document. List 10 apps or features you used today.
For each, write whether it uses AI (Y/N) and why you think so.
For the Y items, guess whether it is symbolic or learned.
Save this. We will revisit it at the end of Module 2 and you will be amazed how much your mental model sharpens.

Mini Exercise

Classify these as AI or not AI: (a) the recommendation row on Netflix, (b) a CAPTCHA, (c) Excel's SUM function, (d) Google Translate, (e) Siri's wake-word detection. Answers at the end of the lesson.

Common Mistakes

Calling every automation "AI". A SQL query that updates a column is automation, not AI.
Believing AI is conscious. It is not. It is statistics at scale.
Thinking AI is new. The term was coined in 1956. The current boom is one of many waves.

Debugging Tips

When you hear "this product uses AI", ask three questions:

What data did it learn from?
What is the prediction or output?
What happens when it is wrong?

If a salesperson cannot answer all three, it is probably not AI in any useful sense.

Knowledge Check Questions

What is the difference between symbolic and learned AI?
Why do most 2026 products use the learned approach?
Give one task where symbolic AI is still the right choice.

Quiz Questions

AI is best defined as: a) Any computer program that automates a task b) Machines performing tasks that normally require human intelligence c) Robots d) ChatGPT Answer: b
Deep Blue (1997 chess engine) is mostly an example of: a) Generative AI b) Statistical AI c) Symbolic AI d) An LLM Answer: c

Challenge Task

Write a one-page "AI vs not-AI" explainer for a non-technical friend. Use only analogies, no jargon. Publish it as a GeekHub post and tag it #ai-beginners.

Real-world Use Cases

Email spam filters (learned AI, classification)
Voice typing on your phone (learned AI, speech recognition)
Self-driving lane assist (learned AI, computer vision)
A traffic light timer (not AI, just a rule)

Industry Insight

In hiring, "AI experience" without context is meaningless on a resume. Specify which family and which task: "trained a classification model on customer churn data" or "built a RAG app with the OpenAI API". This is the language that gets interviews.

Interview Questions

Define AI in one sentence.
What is the difference between rule-based and learned AI? Give an example of each.
Why is the current AI boom different from the 1980s boom?

Summary

AI is the goal of making machines act intelligently. In 2026 it is dominated by learned approaches, where machines find patterns in data rather than being told explicit rules. Everything else in this course is a technique inside this umbrella.

Mini Exercise answers: a) AI (recommendation model), b) AI (image classification model behind the scenes), c) Not AI (it is a formula), d) AI (neural machine translation), e) AI (small on-device model).

Lesson 1.2: The AI to ML to DL to GenAI to LLM hierarchy

Hook / Why This Matters

This is the single most useful diagram in the entire course. If you internalize it, every news headline and job description suddenly slots into the right bucket.

Beginner Analogy

Think of Russian dolls. The biggest doll is AI. Open it: inside is Machine Learning. Open that: inside is Deep Learning. Open that: Generative AI. Open that: Large Language Models. Each is a special case of the one outside it.

Concept Explanation

AI: anything that makes a machine act intelligently.
Machine Learning (ML): AI that learns from data instead of being programmed with rules.
Deep Learning (DL): ML that uses neural networks with many layers.
Generative AI (GenAI): DL models that create new content (text, images, audio, video).
Large Language Models (LLMs): GenAI for text, trained on huge text corpora.

ChatGPT is an LLM. An LLM is a GenAI model. A GenAI model is a deep learning model. Deep learning is machine learning. Machine learning is AI.

Technical Breakdown

Each layer adds a constraint and a power:

Layer	Adds	Trades off
ML over AI	Learns from data	Needs lots of clean data
DL over ML	Handles unstructured data (text, images, audio)	Needs lots of compute
GenAI over DL	Produces new content, not just predictions	Hard to evaluate "good"
LLM over GenAI	Specifically for language	Hallucinates, costly

Visual Learning Suggestion

Insert the nested circles diagram from Lesson 1.1, this time labeled with one real example in each ring: "Symbolic AI: Deep Blue. ML: Spam filter. DL: Image classifier. GenAI: DALL-E. LLM: ChatGPT."

Interactive Element

Read three AI news headlines from this week (any tech site). For each, write which layer it is talking about. Be strict: if it says "AI" but means "an LLM call", note that. This pattern recognition compounds fast.

Hands-on Lab

Open a blank doc. From memory, draw the nested-circles diagram. Place these in the right ring:

Netflix recommendations
ChatGPT
Self-driving Tesla
Midjourney
Gmail's spam filter
DeepFake video
A bank's credit-scoring model

Then check yourself: most are ML, some are DL, two are GenAI, one is an LLM.

Mini Exercise

True or false: Every LLM is a deep learning model, but not every deep learning model is an LLM. Justify.

Common Mistakes

Treating "AI" and "LLM" as synonyms. LLMs are a tiny sliver.
Calling image generators LLMs. They are GenAI but not LLMs.
Believing ML always uses neural networks. It often does not (decision trees, gradient boosting, linear regression).

Debugging Tips

When reading a paper or product spec, ask: which ring are we in? It changes what evaluation methods, costs, and risks apply.

Knowledge Check Questions

Is every machine learning model a deep learning model? Why or why not?
Why is generative AI a subset of deep learning specifically?
Where would you place a stock price prediction model in the hierarchy?

Quiz Questions

ChatGPT is best described as: a) Machine learning that is not deep learning b) A deep learning generative language model c) A symbolic AI system d) An image generator Answer: b
Decision trees are a form of: a) Deep learning b) Machine learning that is not deep learning c) Generative AI d) An LLM Answer: b

Challenge Task

Pick any AI startup from a YC batch in 2025 or 2026. Read its landing page. Write a 200-word note placing its tech in the hierarchy and identifying the layer where the actual differentiation lives.

Real-world Use Cases

ML (not DL): credit scoring, fraud detection on tabular data
DL (not GenAI): face recognition, voice typing
GenAI (not LLM): Midjourney, Suno music, Sora video
LLM: ChatGPT, Claude, Gemini, Llama

Industry Insight

A common 2026 product mistake: teams use an LLM where a tiny ML classifier would be 100x cheaper and more accurate. Knowing the hierarchy saves money.

Interview Questions

Draw the AI to ML to DL to GenAI to LLM hierarchy on a whiteboard with one example each.
Give an example where ML is the right choice and an LLM would be overkill.
Is DALL-E an LLM? Justify your answer.

Summary

AI is the umbrella. ML learns from data. DL is ML with neural networks. GenAI generates content. LLMs are GenAI for text. Internalize this and the rest of the course unfolds cleanly.

Lesson 1.3: Discriminative versus Generative AI in 60 seconds

Hook / Why This Matters

Every AI model in the world is doing one of two things: drawing a line between things, or creating new things. Once you see this split, model choice becomes obvious.

Beginner Analogy

A bouncer at a club is discriminative: looks at you and decides "in" or "out". An artist drawing a portrait is generative: produces something new that did not exist before.

Concept Explanation

Discriminative models answer: "What category does this input belong to?" or "What is the predicted value?". Examples: spam or not spam, house price, image of a cat or dog.
Generative models answer: "Given this prompt, produce a plausible new sample." Examples: write an essay, draw a logo, compose music.

Technical Breakdown

Mathematically, discriminative models learn P(label | input). Generative models learn P(input) and then sample from it. You do not need to compute this, but knowing the framing helps you read papers.

Visual Learning Suggestion

A two-column diagram. Left column "Discriminative" with arrows: image -> label, email -> spam or not, transaction -> fraud or not. Right column "Generative" with arrows: prompt -> essay, prompt -> image, prompt -> music.

Interactive Element

In ChatGPT, run two prompts. First: "Classify this email as spam or not: [paste]". Second: "Write a sample spam email about a fake lottery win." First is discriminative use of an LLM, second is generative.

Hands-on Lab

Pick three real apps you use. For each, identify whether the AI inside is discriminative, generative, or both. Examples to start: TikTok, Photoshop, Gmail.

Mini Exercise

Is Google Translate discriminative or generative? Defend either answer in two sentences.

Common Mistakes

Assuming LLMs are always generative. They can be used discriminatively (classify, score, route).
Confusing "creative" with generative. A creative-looking output can still be a classification underneath.

Debugging Tips

If your AI app is unreliable, ask whether you are using a generative model for what is really a discriminative job. Often, a tiny classifier solves it.

Knowledge Check Questions

Define discriminative vs generative in one sentence each.
Can the same LLM be used both ways? Give an example.
Why are generative models harder to evaluate?

Quiz Questions

A model that decides if a tweet is positive or negative is: a) Generative b) Discriminative c) Both d) Neither Answer: b

Challenge Task

Take your favorite SaaS app. Sketch two AI features: one discriminative, one generative. Specify inputs, outputs, and which would be more valuable to ship first.

Real-world Use Cases

Discriminative: fraud detection, content moderation, search ranking
Generative: writing assistants, image generation, code completion
Mixed: AI agents that classify intent then generate a response

Industry Insight

Most 2026 production AI features are 80% discriminative under the hood, with a generative layer on top to make outputs natural. Knowing this lets you architect for cost and reliability.

Interview Questions

Difference between discriminative and generative models. Give an example each.
Why might you use a discriminative classifier in front of an LLM call?

Summary

Every AI task is either picking a label or producing new content. LLMs can do both, but treat them as expensive generative tools and reach for classifiers when classification is what you need.

Lesson 1.4: The 2026 LLM landscape and how we got here

Hook / Why This Matters

You will hear a hundred model names this year. Without context, you will pick wrong. This lesson gives you the map.

Beginner Analogy

If LLMs were cars, GPT is the Toyota that proved the model, Claude is the Lexus famous for safety, Gemini is the Tesla integrated with the Google ecosystem, Llama is the open-source Honda you can mod yourself.

Concept Explanation

The current LLM era began with the 2017 paper "Attention Is All You Need" (Transformer architecture). Key milestones:

2018: GPT-1 and BERT. Proved Transformers scale.
2020: GPT-3. Showed few-shot prompting works.
2022: ChatGPT. The consumer breakthrough.
2023: GPT-4, Claude, Llama 2 (open source).
2024: Gemini, Llama 3, multimodal mainstream.
2025: Reasoning models (o1-style), agents go mainstream.
2026: On-device LLMs, longer contexts, cheaper inference, better tool use.

Technical Breakdown

The 2026 LLM market has four tiers:

Frontier closed models: OpenAI GPT family, Anthropic Claude family, Google Gemini family. State of the art quality, paid API, easy to use.
Frontier open weights: Meta Llama, Mistral, DeepSeek. Strong quality, downloadable, can be self-hosted.
Small efficient models: Phi, Gemma, Qwen small. Run on a laptop, surprisingly capable for narrow tasks.
Specialized models: code (DeepSeek Coder), vision-language (PaliGemma), embeddings (text-embedding-3, voyage-3).

Visual Learning Suggestion

A timeline horizontal chart from 2017 to 2026 with key models on it. Below the timeline, a 2x2 quadrant: x-axis "closed to open", y-axis "frontier to small", with the 4 tiers placed.

Interactive Element

Open openai.com/api, anthropic.com, ai.google.dev, and llama.com side by side. Note the latest model and price per million tokens for each. This snapshot will be your baseline for the rest of the course.

Hands-on Lab

Create a "LLM scorecard" spreadsheet with columns: Provider, Model, Input price per 1M tokens, Output price per 1M tokens, Context window, Strengths, Weaknesses. Fill in 6 current 2026 models. You will reuse this in Module 5.

Mini Exercise

In 2026 dollars per million output tokens, sort the four current frontier models from cheapest to most expensive. Note which has the largest context window.

Common Mistakes

Picking a model because it is famous, not because it fits the task.
Choosing the largest model when a small one would do.
Forgetting that prices and capabilities change every quarter. Re-check before any production decision.

Debugging Tips

If your app is too slow or expensive, the fix is almost never a fancier prompt. It is a different (smaller) model.

Knowledge Check Questions

What was the architectural breakthrough behind modern LLMs?
Name one frontier closed model and one frontier open-weights model.
When would you pick a small model over a frontier model?

Quiz Questions

The Transformer architecture was introduced in: a) 2012 b) 2017 c) 2020 d) 2022 Answer: b
Llama is best described as: a) A closed frontier model b) An open-weights frontier model c) A symbolic AI system d) A small on-device model Answer: b

Challenge Task

Pick a real use case (for example: summarizing customer support tickets). Argue in 200 words which tier and which specific model you would use, including cost reasoning.

Real-world Use Cases

ChatGPT for general writing and brainstorming
Claude for long-context document analysis
Gemini for Google Workspace integration
Llama for self-hosted, privacy-sensitive deployments
Phi or Gemma for on-device features in a mobile app

Industry Insight

The "best model" question is the wrong question. The right one is "which model is best for this specific task within this specific budget and latency target". Engineers who internalize this ship reliable products. Those who do not chase the leaderboard every week.

Interview Questions

Walk me through the major LLM milestones from 2017 to today.
Compare frontier closed vs open-weights models. When would you pick each?
What is the largest context window you have worked with, and what did it unlock?

Summary

The 2026 LLM landscape has four tiers across closed-versus-open and frontier-versus-small. The best engineers pick the smallest, cheapest model that does the job, not the largest brand-name one.

Lesson 1.5: Where LLMs shine, and where they should not be used

Hook / Why This Matters

Half of failed AI projects in 2026 used an LLM where it had no business being used. This lesson saves you that mistake.

Beginner Analogy

A Swiss army knife is great. But you would not use it to chop down a tree. LLMs are the same: powerful, general, but wrong for some jobs.

Concept Explanation

LLMs shine when:

Input is natural language and output is natural language
Tasks need flexibility and creativity
There is no clean labeled dataset
"Pretty good and fast" beats "perfect and slow"

LLMs are wrong when:

Outputs must be exactly correct (medical dosages, legal citations) without checks
Strong privacy or compliance requires no data leaving your machine
The task is deterministic and small (a regex would do)
Latency must be under 50 ms
You need explainable, auditable decisions

Technical Breakdown

Reach for an LLM when the task is fuzzy: summarization, rewriting, classification with subtle context, drafting. Reach for traditional tools when the task is sharp: arithmetic, exact lookups, deterministic rules.

In practice, the best 2026 systems combine both. An LLM routes the request, a deterministic system does the precise work, an LLM formats the response.

Visual Learning Suggestion

A two-column "Reach for an LLM" vs "Do not reach for an LLM" table, with 5 to 8 example tasks in each.

Interactive Element

Ask ChatGPT to multiply two 6-digit numbers. Then ask it to summarize the plot of any movie. Notice which one it nails and which one it can flub. Math without tools is its classic weakness.

Hands-on Lab

In a spreadsheet, list 15 features your favorite app could add. For each, decide: LLM, classical ML, traditional code, or hybrid. Defend each pick in one sentence.

Mini Exercise

Pick three: (a) password validation, (b) summarizing a legal contract, (c) translating English to Hindi, (d) detecting credit card fraud in real time. Which suit an LLM? Why?

Common Mistakes

Calling an LLM for tasks a regex solves
Using an LLM for arithmetic without giving it a calculator tool
Sending sensitive data to a public LLM API without checking your provider's data retention policy
Promising users a "smart AI assistant" that turns out to be a $0.10-per-query glorified chatbot

Debugging Tips

If your LLM app behaves inconsistently, ask whether the task should have been deterministic. Often the fix is to remove the LLM from the critical path.

Knowledge Check Questions

Name three task types where LLMs are the right tool.
Name three task types where LLMs are the wrong tool.
Why are LLMs poor at exact arithmetic by default?

Quiz Questions

Validating that a string is a valid email address: a) Is a great use of an LLM b) Should use a regex, not an LLM c) Requires GPT-4 d) Needs RAG Answer: b

Challenge Task

Audit any AI feature in a product you use. Decide whether the team made the right tool choice. Write a 300-word memo.

Real-world Use Cases

LLM-suited: drafting emails, classifying support tickets, generating product descriptions, code completion, summarization
Not LLM-suited: calculating taxes, validating IDs, real-time price computation, exact database lookups

Industry Insight

The phrase you want on your resume is: "I used an LLM where it was the right tool, and a classical method where it was not." This single sentence sets you apart from 90% of bootcamp grads.

Interview Questions

Give an example of a task where you replaced an LLM with a simpler approach. Why?
How do you handle math inside LLM workflows?
What is the cost-quality tradeoff when picking a model for production?

Summary

LLMs are excellent at fuzzy language tasks and terrible at exact, deterministic ones. Real engineers compose LLMs with classical code and use each tool for what it is best at.

Module 1 Recap

You can now place any AI buzzword in its correct ring, distinguish discriminative from generative, name the 2026 LLM landscape, and decide when an LLM is and is not the right tool.

SEO Notes for this Module

Primary keyword: "introduction to AI and LLMs"
Suggested URL: /courses/ai-llms-beginners/introduction-to-ai-and-llms
Internal links: link to Module 2 (transformers), Module 5 (APIs), the course overview, and the AI tools page.
FAQ schema: include FAQs from each lesson under one FAQPage schema block at the module level.
Featured snippet target: the "AI vs ML vs DL vs GenAI vs LLM" table in Lesson 1.2.
AI-search-friendly summary: TL;DR at the top, definition blocks, comparison tables, step-by-step lists.

Next Module

Module 2: How ChatGPT and Transformers Work