AI Zero to Hero — Start Here

Most AI content online is either too technical ("here's a 400-line PyTorch training loop") or too hand-wavy ("AI is like a brain, but digital!"). Neither helps if you just want to understand what's actually going on.

This course is the thing in between.

It's a structured 33-post journey from "what even is AI?" to "I can reason about agents, alignment, and where this is all headed." It's written for anyone curious — not just developers. No math beyond arithmetic. No coding required. The focus is on mental models, not tools.

If you can read a news headline and wish you actually understood it, you're in the right place.

---

What this course is

• Concept-first. Every post starts with an analogy and a diagram before it introduces a technical term.
• Plain English. Every ML term is defined the first time it appears, in ten words or less.
• Visually heavy. Hand-drawn Mermaid diagrams on every post, using a consistent palette.
• Short reads. 1500-2500 words per post. One coffee, one post.
• A throughline. 33 posts across 6 modules, each building on the last. Skim the map below.

What this course is not

• Not a coding tutorial. If you want to build a transformer from scratch in NumPy, this isn't it. (Come back after — you'll get way more out of Karpathy's zero-to-hero series.)
• Not an academic reference. If it doesn't change how you think, it doesn't make the cut.
• Not a video course. Every post is a self-contained read.
• Not gatekept. If you're curious, you're ready.

---

The map

Six modules. Thirty-three posts. One story.

---

Module 1 — What AI Actually Is (live now)

The fog-clearing module. Start here even if you think you know.

1. What AI actually is (and isn't) — The four nested circles (AI, ML, DL, LLM) and one honest definition you can use every day.
2. A short, honest history of AI — Seventy years of rollercoaster: Dartmouth, two winters, AlexNet, ChatGPT.
3. The three tribes of AI — Symbolic, connectionist, Bayesian — and why every real system is a collaboration.
4. Why data changed everything — The real reason AI finally worked in 2012 (spoiler: not better math).
5. How to spot AI hype — A five-question field guide for every AI product pitch you'll hear this year.

Time investment: ~40 minutes total. Five coffees.

---

Module 2 — How Machines Learn (live now)

The mental model for "learning." No calculus, no code — just the one move behind every spam filter, every chatbot, and every game-playing AI.

1. What "learning" actually means for a machine — Machines don't learn, they fit. The one metaphor behind everything else in this module.
2. Supervised learning, in one story — A toddler in a park is doing exactly what a spam filter does. Told as one story.
3. Unsupervised learning, or why patterns appear on their own — Clustering, similarity, and embeddings — built from a party you walked into.
4. Reinforcement learning — the dog-treat model of learning — Why every game-playing AI, and the final polish on every chatbot, come from the same puppy trick.
5. Overfitting, underfitting, and why your model lies to you — The single most important failure mode in ML, told through a student who memorised the practice test.
6. Features, labels, and the data diet — Why "garbage in, garbage out" isn't a slogan — it's the whole story.

Time investment: ~55 minutes total. Six coffees.

Module 3 — Neural Networks, Demystified (live now)

What a neural network actually is, told without ever asking you to compute a derivative. No calculus. No code. Six metaphors stacked on top of each other.

1. The neuron is a lie (and that's fine) — An artificial neuron is not a brain cell. It's a tiny voting committee. That swap unlocks everything else.
2. Why stacking layers works — Depth is a ladder of abstractions. Strokes become letters, letters become words, words become meaning. The single most important idea in deep learning.
3. Training a neural network — rolling downhill, blindfolded — Gradients and backprop, told as a blindfolded walk down a crumpled hillside. One tiny step at a time.
4. CNNs — looking at images with a flashlight — The architecture that cracked computer vision open in 2012. One little flashlight that slides across the image, looking for the same pattern everywhere.
5. RNNs — reading one word at a time — The old-school way of handling sequences. A neural network with a notebook. Why it worked, why it couldn't last.
6. Embeddings — turning everything into a point in space — The single most important idea in modern AI that isn't "neural network." Words, pictures, songs, users — all points on one invisible map.

Time investment: ~65 minutes total. Six coffees.

Module 4 — The Transformer Revolution (live now)

The architecture behind every LLM you've heard of. Attention, the block, the scroll, the three-act training story, and the bitter lesson that drove the whole industry to "just make it bigger."

1. Attention, explained without equations — Every word in a sentence walks into a library, glances at every other word, and copies the relevant content into its own notebook. That's the whole revolution.
2. The transformer, piece by piece — Inside one transformer block: attention, feed-forward, residual connections. Stack them and you get ChatGPT.
3. Tokens, context windows, and why length matters — LLMs don't read words, they read chunks. The scroll they can see at once is finite, and it explains half the weird behaviour you've noticed.
4. Pretraining, finetuning, RLHF — the three-act training story — How a raw language model becomes a polite assistant. Three acts, three teachers, one shipped chatbot.
5. Scaling laws and the bitter lesson — Why bigger keeps beating cleverer. The most uncomfortable truth in AI research, and the force behind every billion-dollar training run.

Time investment: ~55 minutes total. Five coffees.

Module 5 — Living with LLMs (live now)

The practical module. How LLMs actually generate, how to prompt them well, when to ground them with retrieval, the fine-tune vs prompt vs retrieve decision, how to evaluate, and the four characteristic failure modes.

1. What an LLM is actually doing — token by token — The five-step loop every chatbot runs. Why it hallucinates, why it can't plan, why it pauses.
2. Prompting as a skill, not a trick — Eight principles grouped into three families. The patterns that keep working as models change.
3. RAG — giving the model a library card — Retrieval-augmented generation as open-book exam. When it helps, when it fails, what to watch for.
4. Finetune vs prompt vs retrieve — when to use what — The decision framework nobody writes down. Prompt fixes instructions, retrieval fixes facts, fine-tuning fixes behaviour.
5. Evaluation — how do you know an LLM is any good? — Benchmarks, golden sets, human rating, LLM-as-judge, and vibes. An honest look at the least-solved problem in LLM work.
6. The failure modes of LLMs — Hallucination, sycophancy, jailbreaks, bias. What causes each one, what you can do about each one, and what you can't.

Time investment: ~65 minutes total. Six coffees.

Module 6 — AI in the Wild (live now)

The "so what now?" module. The field zoomed out: what AI is going out into the world as, what to worry about, and how to keep learning after this course ends.

1. Agents — LLMs that can do things — A chatbot with tools and a loop. Why the intelligence is the easy part and reliability is the hard part.
2. Multimodal AI — models that see, hear, and act — One encoder per modality, one shared embedding space, one set of operations. How a sentence, an image, and a song all become points on the same map.
3. The alignment problem, honestly — No doom, no dismissal. What serious researchers actually worry about and what gets overstated.
4. AI and your job — the honest answer — Your job is a bundle of tasks. AI is reshaping which parts are cheap and which are valuable. Reshape or get left behind.
5. Where this all goes next — and how to keep learning — The final post of the course. Frontier research, realistic forecasts, and an opinionated guide to staying curious.

Time investment: ~50 minutes total. Five coffees.

---

How to read this course

• In order, if you can. Each module builds on the last.
• One post at a time. They're designed to be read over coffee, not binged.
• Out loud sometimes. The analogies stick better that way.
• Skeptically. If a post doesn't change how you see something, leave a comment — I want to know.

You don't need to take notes. You don't need to do anything between posts. The goal isn't to memorise terms; it's to build a mental map that sticks.

When you're ready, start here:

👉 M1.1 — What AI actually is (and isn't)

---

A note on what you'll find elsewhere in this blog

If you poke around, you'll find some older posts that go deeper into code — things like neural networks from scratch, training loops, PyTorch examples. Those are a separate, more developer-focused track. They're still valuable if you want that — but the main AI Zero to Hero experience is the track you just saw above. No code required.

---

Cover photo via Unsplash.