A complete, kid-friendly guide to Artificial Intelligence and Machine Learning โ covering all 7 core topics: AI, GenAI, ML, AI Agents, LLMs, Neural Networks, and Deep Learning. Basic to advanced, with diagrams, analogies, interactive demos, and practice tests.
Like to watch before you read? Begin with this video, then scroll on for the full interactive lesson โ or see every video in order on our Videos page.
You've heard the word "AI" everywhere โ in games, movies, your phone, and school. But what does it really mean?
Kind of โ but not exactly! AI is computer software that can do things we usually think only humans can do, like recognize pictures, understand words, translate languages, or make choices.
Artificial Intelligence is when a computer program can learn from examples and then do something useful โ like finish your sentence, spot a cat in a photo, beat you at chess, or drive a car.
Your phone recognizes your face to unlock โ that's AI seeing.
Songs suggested for you based on what you've liked โ that's AI recommending.
Finds the fastest route avoiding traffic โ that's AI planning.
Chats with you in plain English โ that's AI talking.
AI is HUGE. Inside AI, there are smaller fields. Think of it like the animal kingdom: AI is "animals," and NLP, LLMs, and GPT are like different branches. Here's how they fit together:
Machine Learning (ML) is the part of AI where computers learn patterns from examples โ instead of being told every rule by a programmer.
In regular coding, a programmer writes every rule: "IF temperature > 90, say HOT." In ML, you show the computer 10,000 temperature+label examples and it figures out the rule by itself!
Machine Learning is how we teach computers by giving them lots of examples (called training data) so they can spot patterns and make predictions on new things.
These are three fundamental concepts in machine learning. Memorize these three!
You show the computer examples with the right answer (called labels). It learns to match input โ correct output.
Examples: Show 10,000 cat photos labeled "cat" and 10,000 dog photos labeled "dog." Now it can tell new photos apart. Also: spam filters, house price prediction, medical diagnosis.
No labels โ just raw data. The computer has to find its own patterns and groupings (called clusters).
Examples: Grouping customers by shopping habits, finding weird transactions that might be fraud, recommending movies ("people who liked X also liked Y").
The computer tries actions, gets rewards (+points) for good ones and penalties (โpoints) for bad ones. Over many tries, it learns the best strategy.
Examples: AlphaGo beating world champions, self-driving cars learning to navigate, game AIs, robot dogs learning to walk.
A mix! A small number of labeled examples + a lot of unlabeled data. Good when labeling is expensive (like medical scans).
| Term | What it means | Simple example |
|---|---|---|
| Training Data | Examples the model learns from | 10,000 labeled cat photos |
| Test Data | New examples to check if the model learned well | 100 fresh photos it hasn't seen |
| Feature | An input characteristic the model uses | For a house: size, bedrooms, location |
| Label | The right answer attached to an example | "cat" or "dog" on a photo |
| Model | The pattern-finder that gets built during training | Your trained cat-vs-dog classifier |
| Prediction | What the model guesses for a new input | "This new photo is 92% cat" |
| Accuracy | % of predictions the model got right | 9 out of 10 correct = 90% accuracy |
| Overfitting | Model memorized training data but fails on new data | Aces practice test, flunks the real one |
| Underfitting | Model is too simple โ misses patterns | Says "all photos look the same" |
| Algorithm | The recipe used to train the model | Decision tree, linear regression, etc. |
A flowchart of yes/no questions. Keeps branching until it reaches an answer.
Predicts numbers by drawing the straightest line through data points.
To classify a new item, it looks at the K closest examples and copies the majority.
Many decision trees each give an answer; the majority wins. More accurate than one tree!
Splits data into K groups based on similarity. Used when you don't have labels.
Layers of connected "neurons" that work together. The foundation of Deep Learning.
Here are the most common algorithms grouped by ML type, which are essential to know when building machine learning systems:
Used when you have labeled data (input + correct answer). Two main jobs: Regression (predict a number) and Classification (predict a category).
Draws the straightest line through your data points. Predicts a number using the formula: y = mx + b. Simplest and most famous regression algorithm.
Predicts a probability between 0 and 1 (like "85% chance this email is spam"). Output above 50% = one class, below = other class.
Asks a series of yes/no questions. Each "branch" leads to another question or to a final answer (called a "leaf"). Easy to read and explain!
Builds MANY decision trees (often 100+), each trained slightly differently. They each vote on the answer, and the majority wins. Way more accurate than one tree!
To classify a new item, finds the K closest training examples (neighbors) and takes the majority vote. K is usually 3, 5, or 7. Super simple โ no training needed!
Finds the line (or plane in 3D) that BEST separates two classes โ with the biggest gap between them. Powerful for tricky data.
Uses Bayes' theorem from math to calculate probabilities. "Naive" because it assumes all features are independent (often not true, but works anyway!).
Used when you have data without labels. The model finds hidden patterns by itself. Three main jobs: Clustering (grouping), Dimensionality Reduction (simplifying), and Anomaly Detection (spotting weird stuff).
You tell it "make K groups" (like K=5). It places K center-points and assigns each data point to the nearest one, then moves the centers and repeats until stable. Fast and famous!
Starts with each item as its own group, then merges the closest groups together step by step โ building a "tree" called a dendrogram. You can pick how many groups you want at any level.
Groups together points that are densely packed. Unlike K-Means, you don't need to say how many groups upfront โ and it can find weird shapes too! Also automatically labels lonely points as "noise."
Imagine data with 100 features (columns). PCA finds the few "directions" that capture the most variation โ keeping the important info and dropping the rest. Like summarizing a 500-page book in 5 pages!
Learns what "normal" data looks like, then flags anything that's very different. Common methods: Isolation Forest, One-Class SVM, statistical approaches. Often used with unlabeled data.
Used when an agent learns by trial and error, getting rewards for good actions and penalties for bad ones. Two big families: Value-based (learn what each state is worth) and Policy-based (learn the best action directly).
Builds a "Q-table" with a value for every (state, action) pair โ like a cheat sheet showing the expected reward for each move. Updates the table after each try. Off-policy: learns from any data.
Like Q-Learning, but it updates based on the action it ACTUALLY takes next, not the best possible one. On-policy: more cautious. Tends to be safer in risky environments.
When there are TOO many states to put in a table (like every screen of a video game!), DQN uses a deep neural network to predict Q-values. Famous for learning to play Atari games at super-human levels (DeepMind, 2013).
Instead of learning values, directly learns a "policy" โ a function that says "in state X, do action Y with this probability." Updates the policy in the direction of higher rewards.
Uses TWO networks: an Actor that picks actions (policy) and a Critic that judges them (value). The Critic teaches the Actor to do better. Examples: A2C, A3C, PPO, SAC, DDPG.
| Algorithm | ML Type | What It's For | Famous Use |
|---|---|---|---|
| Linear Regression | Supervised | Predict a number | House prices |
| Logistic Regression | Supervised | Predict yes/no | Spam filter |
| Decision Tree | Supervised | Classify with rules | Loan approval |
| Random Forest | Supervised | Classify with many trees | Medical diagnosis |
| KNN | Supervised | Classify by neighbors | Recommendations |
| SVM | Supervised | Best boundary line | Image classification |
| Naive Bayes | Supervised | Probability-based classify | Spam, sentiment |
| K-Means | Unsupervised | Group into K clusters | Customer segments |
| Hierarchical | Unsupervised | Tree of clusters | DNA family trees |
| DBSCAN | Unsupervised | Density-based clustering | Geo hot-spots |
| PCA | Unsupervised | Reduce dimensions | Face recognition |
| Anomaly Detection | Unsupervised | Find outliers | Fraud detection |
| Q-Learning | Reinforcement | Table-based RL | Simple games |
| SARSA | Reinforcement | Safer table-based RL | Cautious robots |
| DQN | Reinforcement | Deep Q-Learning | Atari games |
| Policy Gradient | Reinforcement | Learn actions directly | Continuous control |
| Actor-Critic | Reinforcement | Policy + Value combined | AlphaGo, ChatGPT RLHF |
Neural networks are the "brains" inside most modern AI. They're loosely inspired by how real brains work โ lots of tiny neurons connected together, passing signals.
No! It's just math. But the structure is inspired by brain cells. Each "neuron" is a tiny calculator that takes numbers in, does a little math, and passes a number out.
A neural network is a bunch of math "neurons" stacked in layers. Each neuron weighs inputs, adds them up, and passes the result on. Together, they can recognize patterns most algorithms can't.
Where your raw data enters. For an image: each pixel is one input. For text: each token is one input.
Layers in the middle where the real pattern-finding happens. More layers = deeper understanding.
Produces the final prediction. For classification: probability of each class. For regression: a number.
Each neuron does 3 little steps:
๐ง Think of a neuron like a kid deciding whether to raise their hand. They weigh the evidence (inputs ร weights), add it up, and if it's strong enough (activation), they raise their hand (fire)!
Data flows in one direction: input โ hidden โ output. Simple and common.
Specialized for images. Uses small "filters" that slide across an image to find shapes, edges, and objects.
Has a loop so it remembers what came before. Great for data in order (words, music notes, stock prices).
The superstar of 2017+. Uses "attention" to focus on the most important words in a sentence. This is what powers GPT, Claude, and Gemini!
Two networks fight each other: one makes fake stuff, the other tries to catch fakes. They both get better! Used for generating realistic images.
A network that squishes data down and then rebuilds it. Used to compress data or remove noise from images.
Deep Learning is Machine Learning that uses neural networks with many hidden layers (that's why it's "deep"!). It's what makes modern AI feel magical.
Regular ML works with simpler algorithms (like decision trees). Deep Learning uses BIG neural networks with many layers, needs WAY more data, and can handle complex stuff like images, speech, and language!
A neural network with 2โ3 hidden layers is "shallow." One with 10, 50, or 500 hidden layers is deep. More depth = more power to understand complicated patterns.
| Question | ๐ค Classic ML | ๐ Deep Learning |
|---|---|---|
| Data needed | Works with small-to-medium datasets | Needs HUGE datasets (millions of examples) |
| Feature engineering | Humans pick what to look at (size, color, shape) | Model figures out features on its own! |
| Hardware | Runs on regular computers (CPUs) | Needs powerful GPUs / TPUs |
| Training time | Minutes to hours | Hours to weeks (sometimes months) |
| Good at | Simple patterns, small data | Images, speech, text, games |
| Example | Spam filter, house price predictor | ChatGPT, self-driving, face unlock |
CNNs (a deep learning model) can recognize faces, cars, cancer in X-rays โ often better than humans!
Deep networks turn sound waves into text. Much more accurate than old methods.
Transformer-based deep networks (LLMs!) can write, summarize, translate, and answer questions.
Deep Reinforcement Learning agents master complex games โ often beating world champions!
The internet gave us billions of images, text, and videos to train on.
Graphics cards (from video games!) turned out to be perfect for training neural networks.
Smart researchers invented ways to train very deep networks (ReLU, dropout, etc.).
Tools like TensorFlow and PyTorch (free!) let anyone build deep models.
NLP = Natural Language Processing. It's the branch of AI that helps computers understand, read, write, and even speak human languages โ like English, Spanish, Hindi, or Japanese.
Because words are tricky! "I saw a bat" โ did you mean the animal ๐ฆ or the baseball thing ๐? Humans figure it out from context. Computers need NLP to do that too!
Natural Language Processing is how we teach computers to work with the messy, magical way humans talk and write. NLP is what makes Siri, Google Translate, and spam filters work.
Convert English to Spanish, Hindi to French, and so on.
Decide if a review, tweet, or message is positive, negative, or neutral.
Shrink a long article into a few bullet points.
Turn spoken words into written words.
Spot names, places, and dates in text.
Write original sentences, stories, or code.
NLP uses a pipeline โ a series of steps. Here's a simple version:
LLM = Large Language Model. It's a very powerful kind of NLP model โ so powerful it can chat, write stories, do homework, and even code. Let's break the name apart:
Imagine a kid who has read EVERY book in the library. Not just read them โ memorized patterns about how words fit together. Now ask that kid, "What word usually comes after 'Once upon a ___'?" โ they'll say "time" without thinking!
That's basically what an LLM does โ just with billions of pages and trillions of word-patterns. ๐โจ
Gather millions of books, articles, and websites โ a HUGE pile of words (called a "dataset").
The computer reads everything and learns which words tend to follow which โ takes weeks on thousands of computers!
Humans give feedback. The model learns to be more helpful, honest, and harmless. This is called RLHF.
Now you can type a question, and the LLM predicts a great answer, one token at a time.
"GPT" is the name everyone knows โ but what do those 3 letters actually stand for? Each one tells us something cool about how it works!
Let's unpack each letter:
GPT models are made by a company called OpenAI. But GPT is just ONE family of LLMs. There are other families made by other companies:
๐ Just like pizza โ "GPT" is like "Pepperoni Pizza." It's one famous kind, but pizza (LLMs) comes in many flavors!
Generative AI is AI that CREATES brand-new stuff โ text, images, music, video, code. It doesn't just recognize or classify; it makes things that never existed before!
Nope! Most AI is discriminative โ it tells things apart (spam vs. not spam, cat vs. dog). Generative AI is different โ it makes NEW content that looks real!
Generative AI produces new text, images, audio, video, or code from a prompt. ChatGPT writing a poem, DALL-E drawing a unicorn, Suno making a song โ all GenAI!
| Question | ๐ Discriminative AI | ๐จ Generative AI |
|---|---|---|
| What it does | Tells things apart / classifies | Creates new content |
| Question it answers | "Is this X or Y?" | "Make me an X." |
| Example task | Is this email spam? | Write me a poem about cats. |
| Output type | A label or number | New text, image, sound, etc. |
| Real-world app | Spam filter, face recognition | ChatGPT, DALL-E, Midjourney, Suno |
Writes essays, stories, emails, code, poems. Powered by LLMs like GPT-4 and Claude.
Creates paintings, photos, logos, and art from a text prompt. Uses "diffusion models."
Composes songs, imitates voices, makes sound effects from text descriptions.
Generates short video clips from a text prompt. New and getting better fast!
Writes code in Python, JavaScript, etc. Helps programmers work much faster.
Generates 3D models, game assets, protein structures, and new medicines.
Predicts the next token over and over. Each prediction is influenced by everything said so far โ that's how they write coherent stories!
Starts with pure noise (static-like TV snow) and slowly "cleans it up" based on your prompt โ until a full picture appears!
Two neural networks battle: a "Generator" makes fake images, a "Discriminator" tries to catch fakes. Both keep improving!
Learns a "squished" version of data, then generates new samples from that squished space. Good for variations of a theme.
A prompt is what you type to ask GenAI for something. The better your prompt, the better the result. This is called prompt engineering!
๐ฏ Good prompts are:
GenAI sometimes states false things with total confidence โ especially facts, numbers, and quotes.
Bad actors can make realistic fake content of real people. Always check sources!
GenAI learns from human-made data โ so it can copy human biases and stereotypes.
Did training data include copyrighted work? Who owns AI-generated art? Still being figured out!
This is the question that confuses most grown-ups! Let's make it crystal clear with a table:
| What is it? | ๐ AI | ๐ฌ NLP | ๐ LLM | โญ GPT |
|---|---|---|---|---|
| Full name | Artificial Intelligence | Natural Language Processing | Large Language Model | Generative Pre-trained Transformer |
| How big? | ๐ณ The biggest category | ๐ฟ A branch inside AI | ๐ A kind of NLP model | โญ A specific LLM |
| What it does | Anything "smart" โ vision, games, robots, chat, etc. | Any task with human language | Big, flexible language tasks (chat, write, code) | Same as LLM โ just the OpenAI brand |
| When invented | 1950s (the word) | 1950s, grew in 2010s | ~2018 onwards | 2018 (GPT-1 by OpenAI) |
| Kid-friendly example | Face unlock, chess AI, self-driving cars | Google Translate, spam filter, Siri | Claude, ChatGPT, Gemini | ChatGPT, GPT-4o, GPT-5 |
| Think of it like... | ๐ซ The whole school | ๐ The English classroom | ๐ Top student in English | ๐ One specific top student |
๐ก "All GPTs are LLMs, but not all LLMs are GPT."
Claude and Gemini are LLMs โ but not GPTs. GPT is a specific brand made by OpenAI. It's like: all iPhones are smartphones, but not all smartphones are iPhones!
๐ก "All LLMs are NLP, but not all NLP is LLMs."
A spam filter uses NLP โ but it's small and simple, not an LLM. LLMs are the big superstars of NLP, but there are lots of smaller NLP tools too!
| Question | ๐ค LLM | ๐ฌ Chatbot |
|---|---|---|
| What is it? | The underlying "brain" model | A full app with a chat interface |
| Example | GPT-4 model, Claude model | ChatGPT app, Claude.ai website |
| Can you talk to it directly? | Only through code (API) | Yes! Just type in the chat box |
| Analogy | ๐ง The brain | ๐ฃ๏ธ The whole person, with a mouth to talk |
Tap a question below. Watch how the AI uses patterns to guess a good answer, one token at a time. (This is a pretend LLM โ but real ones work the same way!)
Not every LLM is the same! Some are great at raw prediction, some are great at following instructions, some are great at thinking step by step. Here are the 7 main kinds you'll hear about:
The raw, "just-trained" model. It has read billions of pages but hasn't been polished yet. It's REALLY good at predicting the next word โ but not great at following your directions.
Examples: Llama base, Mistral base, GPT-3 base. These are the "building blocks" that companies use to make nicer chatbots.
A Base Model that's been "polished" with lots of examples of people giving instructions and AI responding helpfully. Now it actually listens when you tell it to "write a poem" or "summarize this."
Examples: ChatGPT, Claude, Gemini โ the ones you can talk to directly.
These models write out their thinking step-by-step BEFORE giving the final answer. This makes them way more accurate at hard problems โ especially math, logic, and science.
Examples: OpenAI o1 / o3, Claude with "extended thinking" mode, DeepSeek R1.
A clever trick: instead of one giant brain, this model has LOTS of "expert" mini-brains inside. For each question, only the best experts wake up to help. Saves computer power!
Examples: Mixtral, DeepSeek, Qwen3-235B-A22B (has 128 experts, uses 8 at a time).
These don't just read text โ they can also "see" pictures, "hear" audio, and even watch videos! You can show them a photo of your homework and ask for help.
Examples: GPT-4o, Claude Opus 4.7 (can see images), Gemini 2.0.
These switch between two modes: a FAST mode for easy questions, and a DEEP THINK mode for hard ones. Sometimes you can tell them which mode to use!
Examples: Claude Sonnet 4 (fast + thinking modes), Qwen3 (with "/think" or "/no_think" tags).
These aren't just chat models โ they're LLMs that can plan a research project, browse many websites, take notes, and write a full report with sources. They can spend 10โ30 minutes on one question!
Examples: Claude Research mode, ChatGPT Deep Research, Perplexity Deep Research.
An LLM is good at talking. But talking isn't the same as doing. An AI Agent is an LLM that's been given tools โ so it can actually do stuff in the world.
Think of an LLM like a really smart friend on the phone. They can tell you how to make a sandwich โ but they can't actually make it for you.
An AI Agent is like a helpful robot in your kitchen โ it has hands, tools, a fridge to look in, and it can actually make the sandwich. It still thinks with the LLM brain, but now it can also act. ๐ฅช
LLM Brain ๐ง + Tools ๐ ๏ธ + Memory ๐ + Goal ๐ฏ = AI AGENT! ๐ค
Agents follow a simple loop โ think, do, check, repeat โ until the job is done. Here's a diagram:
"I need a date, a venue, invitations, and a cake."
Checks the calendar โ picks Saturday at 3 PM.
Searches web โ finds a bowling alley nearby.
Sends invites to your 10 best friends. Done! ๐
AI researchers group agents into 6 main types, from the very simple to the very smart. Think of it like levels in a video game โ each one can do more than the last! ๐ฎ
Follows super simple rules. It reacts to what it sees right now โ no memory, no planning.
Like Level 1, but it remembers what happened before and keeps a "mental map" of the world. It can handle things it can't see right now.
Has a specific GOAL and searches through all possible steps to reach it. Perfect for solving mazes or planning routes.
Not just ANY goal โ it picks the BEST way to get there. Balances speed, cost, safety, and happiness ("utility" = how happy it makes you).
Improves itself! Every time it tries something, it learns from the result โ like you getting better at a video game the more you play.
Uses a big Language Model as its brain, plus a whole toolbox. Can handle open-ended tasks in plain English โ no coding needed!
Sometimes ONE agent isn't enough. So we make teams of agents that work together โ each with its own job. This is called a "multi-agent system" or "agent swarm."
๐ข Example: To write a research paper, you might have a Researcher agent (finds facts), a Writer agent (drafts sections), an Editor agent (polishes), and a Fact-Checker agent (verifies sources). They pass work back and forth like a real team!
| Agent Type | Memory? | Plans Ahead? | Learns? | Example |
|---|---|---|---|---|
| 1. Simple Reflex | โ | โ | โ | Automatic door |
| 2. Model-Based | โ | โ | โ | Roomba vacuum |
| 3. Goal-Based | โ | โ | โ | Google Maps |
| 4. Utility-Based | โ | โ | Sometimes | Travel app |
| 5. Learning | โ | โ | โ | Netflix recs |
| 6. LLM-Powered | โ | โ | โ | Claude Code |
Tap a word on the left, then tap its meaning on the right. Get them all to turn green!
Pick a job you wish an agent could do for you. Then fill in the 4 boxes on a sheet of paper:
๐ Bonus: Decide which of the 6 Agent Types your design is (Simple Reflex? Goal-Based? LLM-Powered?). Share with the class!
Here is a packed AI/ML cheat sheet covering all 7 core topics โ perfect to print, fold, and keep as a quick reference sheet while studying!
๐ SIDE 1 โ Foundations & Vocabulary
AI โ ML โ Deep Learning โ NLP โ LLMs โ GPT (Russian Doll Rule!)
1. Supervised: labeled data โ spam filter, image classifier
2. Unsupervised: no labels, finds clusters โ customer grouping
3. Reinforcement: rewards/penalties โ AlphaGo, robot control
4. Semi-supervised: small labeled + lots unlabeled (bonus type)
3 layers: Input โ Hidden(s) โ Output. Each connection has a weight; each neuron adds a bias and runs an activation function (ReLU, Sigmoid). Training adjusts weights via backpropagation.
โข Deep Learning = neural networks with MANY hidden layers
โข Needs lots of data + GPUs; great for images, speech, language
โข 2012 "AlexNet moment" started the boom
โข Overfitting: memorizes training data, fails on new data
โข Underfitting: too simple, misses patterns
โข Hallucination: AI confidently states something false
โข Token: a word-piece (input unit for LLMs)
โข Parameter: a learned "dial" inside a model
Epoch: one full pass through training data
๐ SIDE 2 โ Models, Agents & Examples
โข FNN (Feedforward) โ basic, one direction
โข CNN โ images (filters slide over pixels)
โข RNN / LSTM โ sequences (loops with memory)
โข Transformer โ modern LLMs, uses attention (2017+)
โข GAN โ generator vs discriminator (deepfakes, art)
โข Autoencoder โ compresses then reconstructs data
1. Base โ raw, just-trained (Llama base, Mistral base)
2. Instruction-Tuned โ follows directions (ChatGPT, Claude)
3. Reasoning โ thinks step-by-step (o1, o3)
4. Mixture of Experts (MoE) โ only some experts activate (Mixtral)
5. Multimodal โ text + images + audio (GPT-4o, Claude)
6. Hybrid โ fast vs deep-think modes (Claude Sonnet)
7. Deep Research Agents โ plan + browse + report
1. Simple Reflex โ IF/THEN rules (auto door)
2. Model-Based โ has memory/world map (Roomba)
3. Goal-Based โ plans to reach goal (Google Maps)
4. Utility-Based โ picks best option (travel app)
5. Learning โ improves over time (Netflix)
6. LLM-Powered โ flexible, plain-English (Claude Code)
โข Text: ChatGPT, Claude, Gemini (LLMs)
โข Images: DALL-E, Midjourney, Stable Diffusion (diffusion models)
โข Music: Suno, Udio, MusicLM
โข Voice: ElevenLabs
โข Video: Sora, Runway, Veo
โข Code: Copilot, Cursor, Claude Code
โข OpenAI โ GPT-4, GPT-5, DALL-E, Sora
โข Anthropic โ Claude
โข Google โ Gemini, BERT, AlphaFold, AlphaGo
โข Meta โ Llama
โข Mistral โ Mistral, Mixtral
โข NVIDIA โ Makes the GPUs that train AI
โข 1950: Turing Test proposed
โข 1956: Term "AI" coined (Dartmouth)
โข 1997: Deep Blue beats Kasparov (chess)
โข 2011: IBM Watson wins Jeopardy
โข 2012: AlexNet (Deep Learning explosion)
โข 2016: AlphaGo beats Lee Sedol
โข 2017: "Attention Is All You Need" (Transformer paper)
โข 2018: First GPT
โข Nov 2022: ChatGPT launches
Here is how to approach any AI and Machine Learning quiz or test like a pro:
If you have 60 questions in 30 min, that's 30 seconds each. Don't get stuck โ circle hard ones and come back!
Don't just read notes over and over. Test yourself using our 50-question practice quiz to find your weak spots!
Even if you don't know the right answer, you can usually rule out 1โ2 obvious wrong ones. Then guess from the rest.
"Always" / "Never" answers are usually wrong. "Most" or "primarily" answers are usually right when nuanced.
Print our 2-sided cheat sheet. Especially helpful for acronyms (GAN, RLHF, MoE, RAG) and dates (when ChatGPT launched, etc.).
Your first answer is usually right. Only change it if you're SURE the new one is better โ research says first instincts win.
| Topic | Core Focus % | Hot Areas |
|---|---|---|
| Artificial Intelligence (AI) | ~15% | Definitions, history, narrow vs general AI, Turing Test |
| Machine Learning (ML) | ~20% | 3 types (supervised, unsupervised, RL), overfitting, algorithms |
| Neural Networks | ~15% | Layers, weights, CNN/RNN/Transformer, activation functions |
| Deep Learning | ~15% | Why "deep", GPUs, backpropagation, why it works now |
| Generative AI (GenAI) | ~15% | Diffusion, GANs, prompts, image/text/video tools |
| Large Language Models (LLMs) | ~10% | GPT acronym, Transformer, tokens, parameters, types |
| AI Agents | ~10% | 6 types (Simple Reflex โ LLM-Powered), tools, agent loop |
You've got this! ๐ช You've learned about all 7 core topics, practiced with diagrams and examples, and have a cheat sheet ready.
Remember: The kids who master AI aren't always the smartest โ they're the ones who practice active recall, stay curious, and build cool hands-on projects. ๐
Test your knowledge across all AI topics with these 50 practice questions. Tap an answer โ green = right, red = try again. Use the filter buttons to focus on one topic at a time!
What does "AI" stand for?
Which of these is NOT an example of AI?
AI learns by...
Which is the BIGGEST category?
Machine Learning is...
Deep Learning uses...
Which came FIRST historically?
The AI Family Tree goes (biggest to smallest)...
What does NLP stand for?
Which of these is an NLP task?
What is "sentiment analysis"?
What is a "token" in NLP?
Why do NLP systems turn words into numbers?
Which is NOT one of NLP's jobs?
Google Translate is a real-world example of...
The first step in the NLP pipeline is usually...
What does "LLM" stand for?
How does an LLM make sentences?
LLMs are trained on...
A "Base Model" is...
"Instruction-Tuned" means the LLM was...
A "Reasoning Model" is special because...
"Mixture of Experts" (MoE) means the model has...
A "Multimodal" LLM can...
"Parameters" in an LLM are like...
When an LLM makes up something that sounds real but isn't, we call it a...
Which is an example of a Reasoning Model?
Can a modern LLM be more than one type at once?
What does "GPT" stand for?
The "G" in GPT stands for Generative, which means it...
The "P" in GPT stands for Pre-trained, which means it...
The "T" in GPT stands for Transformer, which is...
Which company makes GPT?
True or False: "All LLMs are called GPT."
What's the BIG difference between an LLM and an AI Agent?
What's the agent "formula"?
An automatic door is an example of which agent type?
Which agent type has NO memory and NO planning?
Which of these is most like a "Goal-Based Agent"?
A Utility-Based Agent picks...
Netflix recommendations are an example of which agent type?
Claude with web search is an example of...
A Roomba vacuum that remembers your room is which agent type?
What's the agent "loop"?
What is a "multi-agent system"?
Can an LLM-Powered agent make mistakes?
Which is TRUE?
Can an LLM sometimes make mistakes?
You say "Hey Siri!" and Siri answers. This involves:
Final question! Which of these is the BEST summary?