AI explained

How machines think, in diagrams

From the perceptron to language models: a simple, jargon-free visual guide to the ideas that make artificial intelligence work.

01The building block

The perceptron: the artificial neuron

It all starts here. A perceptron learns a 'template' of what it's looking for and, when something arrives, measures how well it fits. Let's see it with examples.

Example 1 · Recognizing a handwritten digit

The drawing

Fits the template: high match.

The 7 template (learned)

azul suma · rojo resta

Is it a 7?

Yes, it's a 7
match with the template94%

How it groups what it learns

Each drawing is a point. Similar ones land close and form groups; a new one is recognized by the group it lands near.

the onesthe sevensthe eightsthis drawing

Example 2 · Cat or dog? (more pixels)

With a higher-pixel 'photo' it works the same: the template learns that a cat has pointy ears on top (blue) and a dog has floppy ears on the sides (red).

The photo

Pointy ears: fits the template.

The cat template (learned)

azul suma · rojo resta

Is it a cat?

Yes, it's a cat
match with the template95%

And it groups them too

Each photo is a point: cats land on one side, dogs on the other.

the catsthe dogsthis photo

Under the hood · Add, subtract and a number

How does it measure the match? It multiplies each pixel by its weight: some add (blue), some subtract (red). All of it collapses into a single number.

pixelweightadds
0.9+0.80+0.72
0.1-0.60-0.06
0.8+0.70+0.56
0.2-0.90-0.18
0.7+0.50+0.35
Σ adds it all+1.39
−0+

activation

0.80

…and with many, a vector

0.800.200.55

That number (and the vector it forms with others) is the coordinate that places the point on the map of groups.

It works for anything · Audio, video, documents

To the machine there's no difference between a photo, a song or a text: everything becomes a list of numbers. That's why the same mechanism reads any digital medium.

Audio

wave → numbers

0.20.80.10.60.9

Image

pixels → numbers

0.70.30.50.20.8

Video

frames → numbers

0.40.90.60.10.5

Document

letters → numbers

0.60.20.70.80.3

All: one list of numbers

[ 0.2 0.8 … ]

the same network

Change the template (the weights) and the same mechanism recognizes a 7, a cat or anything else. Adjusting that template on its own, from mistakes, is what diagram 2 shows.

02How it learns

Backpropagation: learning from mistakes

A single perceptron is limited; the magic appears when you connect many in layers. The network makes a prediction, compares it to the right answer and measures how far off it was. That error travels backward, tuning each weight to get a little closer. Repeated millions of times, the network learns on its own.

error gradientinputhidden layerŷoutputpredicted0.72target1.00error: 0.28Forward pass →

Forward computes an answer; backward spreads the blame for the error across the weights and corrects them. That back-and-forth, over and over, is 'training' a model.

03Language models

What an LLM is (like ChatGPT)

A Large Language Model is a huge network trained with this same idea, but on almost all the text on the internet. It reads your sentence as a sequence of pieces (tokens), uses 'attention' to decide which words matter most and predicts the most likely next word. By chaining those predictions, it writes.

Input text

Thecatclimbedontothe___
Attention (Transformer)N× layers

Likely next word

roof88%
tree7%
wall3%
car2%

It doesn't 'understand' like a person: it computes probabilities over patterns learned from millions of texts. That's why it can sound confident even when it's wrong.

04What it's for

Where we already use it every day

The same basic idea, trained on different data, solves very concrete problems across every sector. You probably interact with several of these without noticing.

Neural networks

deep learning

Computer vision

Cameras that 'see': recognize objects, faces or defects in a photo.

Healthcare

Spot tumors or injuries in scans, sometimes before the human eye.

Natural language

Assistants that understand, answer, translate and summarize text.

Fraud detection

Stop a suspicious payment the instant it happens.

Recommendation

Suggest the next show, song or product you'll love.

Autonomous vehicles

Cars that read the road and decide how to drive.

In every case it's the same recipe: lots of examples, a network that tunes its weights, and a prediction at the end.

The scale jump

From dozens to trillions of parameters

The 1958 perceptron tuned a handful of numbers. Today's models tune over a trillion. And at that scale, something unexpected happens.

~1,000Perceptron195860KLeNet-5199860 millionAlexNet2012340 millionBERT2018175 billionGPT-32020~1 trillionGPT-42023

Log scale: each step is ~1000× bigger.

Model size →Capability →thresholdtranslatesummarizewrite codereason & chat

Emergent properties

Past a certain size, the model starts doing things nobody programmed or could predict: translating, reasoning step by step, writing code. They weren't added one by one; they emerge from scale.

Illusion of complexity

Inside it's still the same: adding and multiplying numbers. But at that scale the conversation gets so fluid it feels like an expert is on the other side. There isn't: it's statistics at massive scale.

A bit of history

70 years in 10 milestones

The idea isn't new: it was born in the 1940s and advanced in leaps, with winters and springs.

  1. 1943

    The idea is born

    McCulloch and Pitts imagine the first mathematical 'neuron': a yes-or-no decision from several signals.

  2. 1958

    The perceptron

    Frank Rosenblatt builds the first machine that learns from examples. The press dreams of robots that will walk and talk.

  3. 1969

    The first setback

    Minsky and Papert prove the limits of the simple perceptron. Funding dries up: the 'AI winter' begins.

  4. 1986

    Learning from mistakes

    Backpropagation goes mainstream: networks finally tune their own weights, layer by layer. Momentum returns.

  5. 1989

    Artificial eyes

    Yann LeCun uses convolutional networks to read handwritten zip codes. Computer vision takes off.

  6. 1997

    Memory

    LSTM networks learn to remember the context of a sequence: a key leap for speech and text.

  7. 2012

    The big breakthrough

    AlexNet wins a vision contest by a mile. With GPUs and massive data, the 'deep learning' boom begins.

  8. 2017

    Attention

    The paper 'Attention is all you need' creates Transformers, the architecture behind almost all of today's AI.

  9. 2022

    AI for everyone

    ChatGPT puts large language models in the hands of the general public and shifts the conversation.

  10. Today

    Multimodal

    Models that combine text, image, audio and video in one system, and start acting as agents.

Want to apply AI in your business?

Let's talk