A neuron, a network of them, and a 2017 design called the transformer — built from scratch.
First, the words
"AI" is a stack of nested ideas.
People use AI, machine learning, and LLM as if they're the same thing. They're actually rings inside rings — each one a smaller, more specific idea living inside the last.
Rings inside rings
Four labels, one nesting.
Artificial Intelligence
The whole goal — machines doing things that seem intelligent.
Machine Learning
The winning approach inside it: don't hand-code rules — learn patterns from examples.
Deep Learning
Machine learning done with many-layered neural networks.
Large Language Models
One very big deep-learning network trained on text. Agents live here.
The atom
It starts with one neuron.
Takes a few numbers in.
Multiplies each by a weight — how much that input matters.
Adds them up with a bias, then squashes the result through an activation.
Out comes one number. Learning = finding good weights.
Multiply → add → squash
One neuron, one decision.
(0.9·0.8) + (0.1·−0.5) + (0.7·0.6) + bias → activation → one number out.
The one-liner
A neuron just multiplies, adds, and squashes.
Intelligence doesn't live in any single neuron — it emerges from millions of them, and the weights between them.
Scale it up
Wire neurons into layers.
Stack neurons into layers; connect each to the next.
Numbers enter on the left, flow rightward, each layer transforming them a little.
The last layer produces an answer — this sweep is a forward pass.
A modern network has billions of these weighted connections.
Training
Guess, measure, nudge — a billion times.
Guess
The network runs a forward pass and produces an answer.
Loss
We measure how wrong the guess was — a single "wrongness" number.
Backprop
Backpropagation nudges every weight a hair toward being less wrong.
Repeat on billions of examples — that's a trained model.
2017 · the turning point
The transformer, and one big idea: attention.
Older networks read one word at a time and forgot the start of a long sentence by the end. In 2017 a paper with a bold title — "Attention Is All You Need" — introduced the transformer, the architecture behind today's LLMs.
Its trick: let every word look at every other word at once and decide which ones matter. That mechanism is self-attention.
Self-attention
What does "it" refer to?
Why it mattered
Parallelism unlocked the giants.
Runs in parallel
Attention looks at all words at once, so transformers train on enormous text far faster than step-by-step models.
Made LLMs possible
That speed is exactly what made billion-parameter models — GPT, Claude, Gemini — practical to train.
The stack
Dozens of attention layers, stacked.
Early layers
Catch the low-level stuff — grammar, word order, which words go together.
Later layers
Catch meaning and intent — what the sentence is actually trying to say.
Each layer refines meaning a little more than the last.
Chapter 1 in one breath
Multiply, add, squash — a billion times.
A neuron weighs its inputs. A network stacks neurons into layers and sweeps signals through them. Training tunes the weights. The transformer adds attention so every word can consider every other word — and that's the engine every AI agent runs on.
Up next · Chapter 2
How LLMs actually work.
We've built the engine. Next we turn the key — tokens, context, and what really happens when you hit send.