The AI Landscape
Before we build anything, let’s draw the map. “AI” is a huge field with decades of history, and the chat assistants everyone’s excited about right now are one bright corner of it — not the whole thing. This lesson gives you the big picture: what artificial intelligence actually means, the nested circles from AI down to the language models you’ll use, and exactly where this course lives on that map. No math, no jargon left unexplained — just the lay of the land so nothing later feels like it came from nowhere.
Think of AI like the word “sports.” Sports isn’t one game — it’s soccer, chess, swimming, all of it. When people say “AI,” they mean a giant family of clever computer tricks. The chatbots you’ve heard about are just one game in that family, like soccer inside “sports.” This lesson is a map of the whole family, with a big X marking the one corner we’re going to explore together.
What we mean by “AI”
☺ Like you’re 10: “Artificial intelligence” just means “a computer doing something that usually needs a thinking person” — like recognizing your face, suggesting the next word you’ll type, or beating you at chess. It is not a shiny metal robot that wakes up, gets bored, and decides to take over. That part is only in movies.
Artificial intelligence (AI) is a broad field of computer science: the study of getting machines to do things that normally require human intelligence — recognizing images, understanding language, making decisions, playing games, spotting patterns. It’s not a single product or a single technology. It’s an umbrella term stretched over dozens of very different techniques, some of them older than the personal computer. The spam filter on your email, the app that recognizes a song, the route your maps app picks, the system that flags a suspicious credit-card charge — all of that is AI, and none of it is new or dramatic.
The single most useful distinction to learn on day one is between two kinds of AI:
- Narrow AI (also “task AI” or “weak AI”) — a system built to do one kind of task, or a narrow band of related tasks. It can be superhuman at that task and completely useless outside it. A model that plays world-champion chess can’t drive a car; one that transcribes speech can’t diagnose an X-ray. Every AI system that exists today, including every chatbot you’ve used, is narrow AI. Even a chat assistant that seems to “do everything” is really doing one thing — predicting helpful text — very flexibly.
- Artificial general intelligence (AGI) — a hypothetical system with human-level competence across the board: it could learn any intellectual task a person can, transfer skills from one domain to another, and set its own goals. AGI does not exist. It’s a research aspiration and a subject of serious debate about whether, when, and how it might arrive — not something you can download or buy today.
This matters because the word “AI” carries a lot of science-fiction baggage. The robots that scheme and rebel — HAL, Skynet, the Terminator — are AGI-and-beyond, and they’re fiction. What’s real today is enormously capable narrow AI: tools that are astonishing at specific jobs and have no inner life, no goals of their own, and no awareness. Keeping “what exists” (narrow AI) separate from “what’s hypothetical” (AGI) is the first defense against both hype and fear.
All the AI you can use today is narrow AI — brilliant at particular tasks, blank everywhere else. AGI — a machine that matches humans at everything — is hypothetical. When a headline blurs the two (“the AI decided…”, “the AI wants…”), that’s your cue to slow down and check what’s actually being claimed.
The nested circles
☺ Like you’re 10: Picture a set of measuring cups that fit inside each other. The biggest cup is “AI.” Inside it sits a smaller cup, “machine learning.” Inside that, a smaller one, “deep learning.” Then “generative AI.” And the tiniest cup, right in the middle, is the language models we’ll spend this course on. Every small cup is fully inside the bigger ones — so a language model is all of those things at once.
The clearest way to picture the field is as a set of nested circles, each one a subset of the one around it. Working from the outside in:
- Artificial intelligence — the whole field (everything above). Includes plenty of approaches that involve no learning at all, like hand-written rules and search algorithms.
- Machine learning (ML) — the subset of AI where the system learns patterns from data instead of being explicitly programmed with rules. You don’t write “if the email says ‘free money’, mark spam”; you show it thousands of labelled emails and it figures out the pattern.
- Deep learning — the subset of ML that uses neural networks with many layers (that’s the “deep”). This is what unlocked modern breakthroughs in vision, speech, and language, because deep networks can learn very rich patterns from very large data.
- Generative AI — the subset of deep learning whose job is to create new content — text, images, audio, code — rather than just classify or predict a label. This is the wave you’ve been hearing about.
- Large language models (LLMs) — generative models specialized for language. This is the exact spot on the map where the assistants in this course live.
Read the circles as “is a kind of.” An LLM is a kind of generative model, which is a kind of deep-learning model, which is a kind of machine-learning model, which is a kind of AI. So when someone calls ChatGPT, Claude, or Gemini “an AI,” they’re right — but they’re describing it at the widest, vaguest circle. The precise label is “a large language model,” the innermost circle. Notice too that the outer ring holds more than the inner ones: there’s AI that isn’t machine learning (rule-based systems, classic search), and machine learning that isn’t deep learning. The exciting recent wave lives in the two smallest circles. This diagram is, quite literally, the map of the whole lesson — keep it in your head and every later topic will have a place to hang.
How machines learn: three paradigms
☺ Like you’re 10: There are three ways to teach. You can show flashcards with the answers on the back (“this is a cat, this is a dog”). You can dump a pile of photos and say “sort these into groups yourself, I won’t tell you the labels.” Or you can let someone try, then give a thumbs-up or thumbs-down so they learn what earns the treat. Machines learn in those same three ways.
The middle circle — machine learning — is the idea of learning patterns from data rather than following hand-written rules. But “learning from data” happens in three broad styles, and knowing them makes almost everything else click into place.
| Paradigm | The one-line intuition | Everyday example |
|---|---|---|
| Supervised learning | Learn from labelled examples — inputs paired with the right answers — then predict answers for new inputs. | Thousands of emails tagged spam / not-spam → a spam filter; labelled X-rays → a tumor detector. |
| Unsupervised learning | Find structure in data that has no labels — group similar things, spot patterns, compress. | Cluster customers into segments no one defined in advance; flag an unusual transaction as an outlier. |
| Reinforcement learning | Learn by trial and reward — take actions, get feedback (good/bad), and adjust to earn more reward over time. | A program that masters a game by playing millions of rounds; a robot learning to walk without falling. |
These aren’t rival teams — real systems mix them. The LLMs at the center of our map are a great example: their core skill (predict the next word) is learned in a self-supervised way, which is a flavor of supervised learning where the “labels” come free from the text itself (the next word is the answer). Then they’re polished with a dose of reinforcement learning from human feedback — people rate answers, and the model is nudged toward the ones people prefer. You don’t need the machinery yet; just hold onto the three shapes of learning, because How models work will show exactly which ones build a chat assistant.
Three ways to learn: supervised (labelled examples → predictions), unsupervised (no labels → find structure), reinforcement (actions + reward → better behavior). Modern language models are trained with a blend of all three.
From neurons to networks
☺ Like you’re 10: Imagine a huge crowd of tiny helpers standing in rows. Each helper listens to a few helpers in front, does one super-simple thing, and passes a number to the next row. On its own, one helper is almost silly. But millions of them, wired together and practiced on lots of examples, can do something amazing — like knowing a cat from a dog. Practice adjusts how loudly each helper listens to the others, and that’s the whole “learning.”
Step into the “deep learning” circle and you meet the neural network — the engine underneath almost every AI system you’ll use. The name borrows loosely from the brain, but you don’t need any biology. Here’s the honest, math-free picture:
- A network is built from lots of tiny units (loosely, “neurons”), arranged in layers. Data enters the first layer, flows through the middle layers, and a result comes out the last.
- Each unit does something dead simple: take some numbers in, multiply each by a weight, add them up, and pass the total on. The magic isn’t in any one unit — it’s in how millions of them combine.
- The weights are the knobs. Training means showing the network examples and nudging every weight a hair at a time so its outputs get closer to the right answers. “Learning from data” is exactly this: the data sets the weights. Nobody hand-programs them.
- “Deep” just means many layers. More layers let the network build up richer ideas — early layers might catch edges in an image, later layers whole faces — which is why deep networks cracked problems shallow ones couldn’t.
There’s a second split worth carrying forward, because it separates most of “classic” AI from the wave this course is about — what the network is for:
| Predictive / discriminative models | Generative models | |
|---|---|---|
| The job | Look at an input and output a label or number — classify or predict. | Create new content that resembles the training data. |
| Question it answers | “Which is it?” / “How much?” | “Make me one.” |
| Examples | Spam or not-spam · cat or dog · tomorrow’s temperature · fraud risk score | Write a paragraph · draw a logo · compose a melody · generate code |
For decades, most deployed AI was predictive — sorting things into buckets and forecasting numbers. That work is enormously valuable and still everywhere. The recent excitement comes from generative models crossing a quality threshold: they got good enough to write, draw, and code usefully. That leap is what put AI on the front page — and it’s the leap that lands us squarely in the two innermost circles of our map.
The transformer & attention (lightly)
☺ Like you’re 10: When you read “she poured it into the cup because it was empty,” you instantly know “it” means the cup, not the water — because you glance back at the words that matter. The transformer gave computers that same knack: for every word, quietly look around at the other words that matter most and let them shape the meaning. That “pay attention to the important words” trick is what made today’s chatbots possible.
Not every kind of neural network is equally good at language. The breakthrough that made modern LLMs possible is a specific network design called the transformer. You’ll go deeper in How models work; here we just want the flavor, because “transformer” is a word you’ll hear constantly.
The transformer’s key ingredient is a mechanism called attention. Language is full of long-distance connections — a pronoun refers back to a noun from ten words ago; the end of a sentence depends on how it began. Earlier network designs processed words strictly in order and tended to lose track of those distant links. Attention fixed that by letting the model, for every word, look across the whole passage at once and decide which other words matter most for understanding this one — then weight them accordingly. Loosely: attention ≈ “for each word, look at the words that matter.”
Two things made this a turning point. First, attention captures meaning and context far better than what came before, because nothing important is out of reach. Second — and just as important in practice — the design runs efficiently on modern hardware and scales: feed a transformer more data and make it bigger, and it keeps getting better. That scalability is what took language models from a research curiosity to Claude, GPT, and Gemini. The letters in “GPT” even spell it out: Generative Pre-trained Transformer. You don’t need the equations to build with these tools — just the intuition that a transformer reads a whole passage at once and uses attention to focus on what’s relevant.
The transformer is the network architecture behind modern LLMs, and attention is its core trick: for each word, look across the whole passage at the words that matter most. It captures context beautifully and scales with size and data — which is why the current wave happened when it did.
Foxy: Okay, real question — is AI just robots that wake up and take over the world? Should I be scared to even start?
Master Panda: Breathe. No. What you’ll use is narrow AI — very clever at one job, with no wishes and no plans of its own. The scheming robots are movies. You’re safe to be curious. Take it slow; we’ll build the picture one circle at a time.
Professor Owl: Let me draw it. (sketches nested rings) Big ring, all of AI. Inside it, machine learning. Inside that, deep learning. Then generative AI. And this little ring in the middle — large language models — is where today’s assistants live. Everything you’ll learn hangs somewhere on this map.
Timmy the Turtle: And a fact-check for the road: “AGI,” a machine as broadly smart as a person, does not exist yet. So when a headline says an AI “decided” or “wants” something, slow down and verify — that’s hype language, not what today’s narrow tools actually do.
Master Panda: Well said, Timmy. Curiosity in, panic out. Now — shall we walk into the middle ring together?
Where generative & agentic AI fit
☺ Like you’re 10: The map is huge, but we’re only visiting one neighborhood: the part where AI writes and creates (that’s “generative”) and the part where it can use tools and take steps to get something done (that’s “agentic”). We’ll wave hello to the other neighborhoods — like AI that only sorts photos, or robots that walk — but we won’t move in there.
Now we can plant the flag. This course lives in the two innermost circles: generative AI, and its close cousin agentic AI. Generative AI is the LLM assistant answering your question or drafting your email. Agentic AI is the next step up — an LLM that doesn’t just reply but can act: use tools, look things up, take several steps toward a goal, and check its own work. A model that answers you is generative; a model that reads your calendar, drafts the reply, and books the room is agentic. The whole arc of this course walks from one to the other — from prompting a single reply, to reasoning through hard problems, to full agentic systems with memory, retrieval, tools, and teams of agents.
A good map shows what you’re not studying, too, so you can place the news you hear and not feel like you’re missing the point of the course. These are real, important branches of AI — they’re just outside our scope:
| Field | What it does | In this course? |
|---|---|---|
| Generative & agentic AI | LLM assistants that write, reason, use tools, and act toward goals | ✅ Yes — this is our whole focus |
| Classic / predictive ML | Fraud scoring, recommendations, forecasting, spam filtering, tabular prediction | △ Context only — you’ll know where it sits |
| Computer vision (as a field) | Detecting objects, reading medical scans, industrial inspection | △ We touch vision only where an LLM sees images |
| Robotics & control | Machines that sense and move in the physical world — arms, drones, self-driving | ❌ No — a separate discipline with its own hardware |
None of these are lesser; they’re just different addresses on the same map. Drawing the boundary now means that when you read about a self-driving breakthrough or a new cancer-screening model, you can place it instantly — “ah, robotics,” “ah, predictive vision” — and know it isn’t a gap in what you’re learning here. Our neighborhood is language models that generate and act, and that’s a big, rich neighborhood on its own.
“AI” is not synonymous with “chatbot.” Generative LLMs are a thrilling slice of AI, not its entirety — and plenty of the most valuable AI in the world (fraud detection, logistics, medical imaging) isn’t generative at all. Holding the full map in mind keeps you from over-applying a language model to a job an old-fashioned predictive model would do better and cheaper.
Your path from here
☺ Like you’re 10: You’ve got the map. Now we start walking. The very next stop opens up that little middle circle and shows how a language model actually turns your words into an answer — one small step, then the next. No leaps, no cliffs.
That’s the landscape. You can now situate any AI story you hear inside the nested circles, tell narrow AI (real, everywhere) from AGI (hypothetical, not here), name the three ways machines learn, picture a neural network as layers of weighted units, and explain in one breath why the transformer’s attention made modern assistants possible. That’s the vocabulary the rest of the course builds on.
The natural next stop is How models work, which zooms all the way into the innermost circle and shows, step by step, how an LLM turns your prompt into a reply — the “next-token prediction” idea that everything downstream depends on. From there the path opens up: prompting to steer a model well, reasoning for harder problems, then agentic AI, tools and MCP, memory, retrieval & RAG, multimodal, and on to actually building agents, running them in production, and doing it responsibly and securely.
One last reassurance from Master Panda: this course is written to serve the whole spectrum, from someone who has never opened a chatbot to a professional shipping AI systems for a living. If a section feels too basic, skim ahead; if one feels too fast, the AI Academy replays these ideas as a story. Either way, there’s no wrong pace — only the next circle inward.
Pull up three recent AI headlines from any news site. For each one, place it on today’s map: is it narrow AI or a claim about AGI? Is it generative (writing, drawing, coding) or predictive (scoring, forecasting, classifying)? Which circle does it live in — plain ML, deep learning, generative, or an LLM? You’ll be surprised how quickly the fog of hype clears once every story has a spot on the map.
(1) What’s the difference between narrow AI and AGI, and which one exists today? (2) Put these in order from biggest to smallest: deep learning, LLMs, AI, machine learning, generative AI. (3) Name the three ways machines learn and give a one-line intuition for each. (4) In a sentence each: what is a transformer, and what does “attention” do — and why did that combination make modern chatbots possible?
Check your answers
- Narrow AI vs. AGI: Narrow AI is very clever at one job with no wishes or plans of its own, and it’s what exists and is everywhere today. AGI — a machine as broadly smart as a person — is hypothetical and does not exist yet, so headlines saying an AI “decided” or “wants” something are hype, not reality.
- Biggest to smallest: AI → machine learning → deep learning → generative AI → LLMs. Each is nested inside the one before it, with large language models as the little ring in the middle of the map where today’s assistants live.
- Three ways machines learn: Supervised — learn from labelled examples (inputs paired with the right answers) to predict answers for new inputs. Unsupervised — find structure in unlabelled data, grouping similar things or spotting patterns. Reinforcement — learn by trial and reward, taking actions and adjusting from good/bad feedback to earn more reward over time.
- Transformer & attention: A transformer is the specific neural-network architecture behind modern LLMs. Attention is its core trick: for every word it looks across the whole passage and decides which other words matter most for understanding this one. That combination captured context far better than earlier designs and scaled with more data and size, which is why modern chatbots became possible.