Nando Muñoz

BlogCode and management

Is anyone inside? An introduction to artificial intelligence for everyone

In 1770 a machine won at chess with a man inside it. Where AI comes from, what is inside it, why it gets things wrong and what it does with what you tell it.

Originally written in Spanish. This translation is automatic and may contain mistakes. Read the original in Spanish

Old coloured engraving of the Mechanical Turk: an automaton in a turban and fur coat, seated at a chessboard, with the doors of its cabinet open to show gears and levers
Joseph Friedrich zu Racknitz, ilustración de «El Turco» (1789). Dominio público.

Vienna, 1770. In Schönbrunn Palace, in front of Empress Maria Theresa, a court official called Wolfgang von Kempelen opens the doors of a wooden cabinet. Inside you can see wheels, gears, levers. Sitting behind the cabinet is a life-size figure dressed as a Turk, with a turban and a long pipe in his hand, looking at a chessboard. Kempelen closes the doors, winds up the mechanism with a key and the figure starts moving the pieces.

And wins.

It will keep winning for decades, under different owners, across half of Europe and in America. According to the chronicles, Napoleon and Benjamin Franklin sat down opposite it. And all that time there was a man inside: a good chess player, tucked into a sliding seat that hid him whenever the doors were opened, who followed the moves thanks to magnets under the board and moved the figure's arm with a system of levers.

What is astonishing about the Turk is not the trick. It is how many people preferred to believe it.

More than two hundred and fifty years later you type a question to ChatGPT, it answers in impeccable Spanish, and the doubt is the same as in Schönbrunn: is anyone in there? This article is the long answer. The short one, so you are not left in suspense: there is no little man hidden away. There are a great many more.

Almost three thousand years of trying to make someone

Since the end of 2022, for a lot of people artificial intelligence is ChatGPT. That is understandable: it was the tool that took it out of the laboratories and put it on any phone. But the idea of building a mind is much older than that date, and than computers. This timeline gathers some of the milestones that have led here:

  1. 8th century BC

    Hephaestus's handmaidens

    In the Iliad, Homer tells how Hephaestus, the smith god, had made himself golden maidservants who moved and spoke on their own. The idea is older than the technology.

  2. 4th century BC

    The syllogism

    Aristotle describes the syllogism, which centuries later textbooks would sum up like this: “All men are mortal; Socrates is a man; therefore Socrates is mortal”. For the first time, whether an argument is valid depends on its form, not on what it says.

  3. 1st century

    Hero's theatres

    Hero of Alexandria describes in his treatise Automata mechanical theatres driven by weights, ropes and falling sand: the movements are written down before they happen, a very old way of programming.

  4. Al-Jazari's devices

    The engineer Al-Jazari completes his Book of Knowledge of Ingenious Mechanical Devices, with fifty machines; among them, a boat with four automaton musicians that entertained palace parties.

  5. Around 1305

    Llull's wheels

    The Majorcan Ramon Llull designs concentric wheels with written concepts that, when turned, combine ideas: an attempt to mechanise reasoning.

  6. 1673-1694

    Leibniz's calculator

    Leibniz builds a mechanical calculator and dreams of a language in which arguments are settled by doing sums. His motto: Calculemus!, “let us calculate!”.

  7. 1738-1739

    Vaucanson's duck

    Jacques de Vaucanson presents in Paris a flute player that plays twelve tunes and, the following year, his duck: it flaps its wings, eats grain, “digests” it and expels it. The digestion was a trick: what came out the back was already prepared inside.

  8. 1768-1774

    “The Writer”

    Pierre Jaquet-Droz builds “The Writer”, an automaton boy who dips his pen and writes any text of up to forty characters, chosen on a wheel: a machine that can be programmed. It still works, in Neuchâtel (Switzerland).

  9. The Turk

    The Turk of Kempelen, with its hidden player. More than two centuries later, Amazon called a service Mechanical Turk in which people do, on commission, tasks that look as if a machine had done them. The joke was no accident. Hold on to it, because it comes back.

  10. The first program

    Ada Lovelace publishes her notes on Babbage's analytical engine, with what is now considered the first program, and a warning: the machine “has no pretensions whatever to originate anything”. Turing will call it “Lady Lovelace's objection”.

  11. The laws of thought

    George Boole shows in The Laws of Thought that reasoning can be written as algebra, with true and false in place of numbers.

  12. 1888-1906

    Cajal's neurons

    Santiago Ramón y Cajal, looking down the microscope and drawing what he saw, shows that the brain is a set of independent cells that communicate with each other: neurons. Nobel Prize in Medicine in 1906.

  13. Logic with wires

    Claude Shannon proves in his master's thesis that circuits with switches can do that algebra. Logic can now be built with wires.

  14. The neuron as a switch

    Warren McCulloch and Walter Pitts describe a neuron as a logic switch that fires or not depending on what it receives: the brain, seen as a Boolean machine.

  15. 1946-1953

    The Macy conferences

    The Macy conferences sit mathematicians, neurologists, anthropologists and psychologists down in New York to talk about the same thing in different words.

  16. Cybernetics

    Norbert Wiener publishes Cybernetics, with a central idea, feedback: a system that measures what it does and corrects itself, like a thermostat or like your hand when it reaches for a glass.

  17. 1948-1949

    Grey Walter's tortoises

    William Grey Walter builds two robot tortoises in Bristol, Elmer and Elsie, that seek out light, dodge obstacles and go back to recharge by themselves. With next to nothing, they seem to have a will of their own.

  18. Hebb's rule

    Donald Hebb proposes a simple rule: when two neurons fire at the same time, the connection between them is strengthened. It is the precise, neuron-by-neuron version of an old idea about how a brain learns, and the one neural networks will inherit.

  19. The Turing test

    Alan Turing publishes “Computing Machinery and Intelligence”. Instead of arguing over whether a machine “thinks”, he proposes a game: if in writing you cannot tell it from a person, what does it matter? This is the Turing test, the Schönbrunn question turned into a method. And he suggests not programming an adult mind, but a “child machine” that can be educated.

  20. Dartmouth gives AI its name

    A group of researchers meets at Dartmouth College to found a discipline, artificial intelligence, a name John McCarthy had proposed the year before in the invitation. There Newell and Simon show the Logic Theorist, written with Cliff Shaw: a program that proves theorems on its own.

  21. 1957-1958

    The perceptron

    Frank Rosenblatt simulates on an IBM 704 the perceptron, the first artificial neuron that learns from examples. The New York Times reports that the Navy expects from it “the embryo of a computer able to walk, talk, see, write, reproduce itself and be conscious of its existence”.

  22. A machine that learns to play

    Arthur Samuel publishes a program that learns to play draughts by playing against itself. The title of his paper popularises the expression machine learning.

  23. Networks in doubt

    Minsky and Papert show in Perceptrons that those single-layer networks cannot learn some very basic things. The money for neural networks goes elsewhere.

  24. 1974-1980

    The first winter

    The Lighthill report (United Kingdom, 1973) and the cuts in the United States leave the whole of AI without money: the first “AI winter” arrives.

  25. Backpropagation

    Rumelhart, Hinton and Williams popularise backpropagation, a way of correcting networks of many layers by sharing out each error among the connections that caused it. It is the thaw.

  26. AlexNet

    AlexNet wins by a landslide an image-recognition contest with many layers, more than a million photos and video-game graphics cards doing the sums. A name takes hold: deep learning.

  27. The transformer

    A Google team presents the transformer, a way for the machine to learn which words matter for understanding each of the others. It is the T in GPT.

  28. 30 November 2022

    ChatGPT

    OpenAI opens ChatGPT to the public. The technique, which already existed, doesn't change: what changes is that anyone can talk to it from the browser.

  29. 8 October 2024

    The neural networks Nobel

    The Royal Swedish Academy of Sciences awards the Nobel Prize in Physics to John Hopfield and Geoffrey Hinton for making machine learning with neural networks possible. More than eighty years after McCulloch and Pitts.

Four paths cross in that list: imitating the body, imitating reasoning, imitating the brain and, lastly, letting the machine learn on its own. Almost everything we now call artificial intelligence takes the fourth, and it inherited its two pieces from neuroscience: the neuron as the building block and learning as the adjusting of connections.

What is inside?

Programs that learn from examples, instead of following instructions written one by one by a person. Notice what I am not saying: nothing about thinking or understanding. They learn in a neural network, a mesh of artificial neurons in which each neuron is a small switch that receives numbers and passes others on to the next ones. Each connection has a value, the weight, which measures how much it influences. Learning is moving those weights a little at a time until it gets things right, with the rules that already appeared in the timeline: Hebb's, the perceptron's error correction and backpropagation. The biggest models have trillions of weights, and nobody has set them by hand.

The lesson can be a photo labelled “cat” or a reward for winning a game, as in Samuel's draughts program. In a language model it comes from the text itself: the next word of a sentence is covered up and the model is asked to guess it. Since the text already carries the answer, nobody has to label anything, and that is why it could be trained on an enormous part of what is written on the internet.

What is left at the end is called a model: a recipe the machine has written by itself after trying millions of dishes, which works but which nobody, not even those who trained it, can read line by line. If it was trained on text and is very big, it is an LLM, short for large language model. ChatGPT, Claude and Gemini are LLMs with a chat window in front.

The only thing an LLM can do is guess what comes next in a text. It is what your phone's keyboard does when you type “good” and it suggests “morning”, but having read an entire library and taking the whole conversation into account. It guesses a piece, adds it and starts again; piece by piece, a paragraph comes out.

I say piece because the model does not read words but tokens: fragments that are sometimes a whole word, sometimes half of one and sometimes a punctuation mark. It looks like an engineering detail, but it explains three things any user notices: that companies charge by tokens; that each model can only handle a limited number at a time, and that is why it forgets the beginning of long conversations; and that it has trouble counting the letters in a word. It doesn't see letters, it sees pieces.

With this we can go back to the question of Schönbrunn. Inside the model there is nobody, but everything we people have written is there: books, articles, forums, arguments. And before opening it to the public, OpenAI turned to people who wrote example conversations and scored which answers were better, to teach it manners. Whoever answers is not a man hidden away. It is an average of a great many people, and that explains the good and, as we will see, the bad.

That average has been remade in recent years at a speed that is hard to keep up with. You don't need to memorise the names; it is enough to notice the pace:

  1. 11 June 2018

    GPT-1

    The first experiment: a transformer trained only to guess the next word learns some language along the way.

    OpenAI

  2. 2 November 2018

    BERT

    Not a chat: a model for understanding text, which Google releases openly and the following year puts into its search engine.

    Google

  3. 14 February 2019

    GPT-2

    It wrote texts so believable that OpenAI took months to release it in full, for fear of fake news.

    OpenAI

  4. 28 May 2020

    GPT-3

    More than a hundred times bigger than the last one: with a couple of examples it learns new tasks. From size alone.

    OpenAI

  5. 30 November 2022

    ChatGPT

    GPT-3.5 with a chat window and manners. In five days it has a million users: the day all this reaches your family.

    OpenAI

  6. 24 February 2023

    LLaMA

    Meta shares it with researchers, its weights end up circulating on the internet and open AI takes off.

    Meta

  7. 14 March 2023

    GPT-4

    It understands images and, in a simulation, passes the United States bar exam.

    OpenAI

  8. 14 March 2023

    Claude

    The same day, Anthropic's assistant is born, a company founded to make AI safer.

    Anthropic

  9. 11 May 2023

    Claude, cien mil tokens

    Its memory reaches a hundred thousand tokens, about 75,000 words: for the first time you can paste in a whole book and ask it about it.

    Anthropic

  10. 18 July 2023

    Llama 2

    Free even for commercial use: any company can set up its own assistant without paying licences.

    Meta

  11. 6 December 2023

    Gemini

    Born understanding text, image, audio and video at once. In February 2024, Bard, Google's chat, takes the same name.

    Google

  12. 11 December 2023

    Mixtral 8x7B

    Europe enters the race: a French company openly releases a model that shares out the work among internal “experts”.

    Mistral AI

  13. 4 March 2024

    Claude 3

    Three sizes, from the fast and cheap to the powerful, so that everyone pays only for what they need.

    Anthropic

  14. 18 April 2024

    Llama 3

    The most capable open model of its moment, and the basis of Meta's assistant on WhatsApp and Instagram.

    Meta

  15. 13 May 2024

    GPT-4o

    It sees, hears and speaks in real time, with a voice that lets itself be interrupted and laughs.

    OpenAI

  16. 20 June 2024

    Claude 3.5 Sonnet

    Smarter than the most expensive Claude, twice as fast and at a fifth of the price. It becomes the favourite for programming.

    Anthropic

  17. 12 September 2024

    o1

    The first model that “reasons” before answering, and that improves a lot in maths and science.

    OpenAI

  18. 20 January 2025

    DeepSeek-R1

    A little-known Chinese company matches o1 with far less money and publishes it for free. A week later, Nvidia loses nearly $600 billion on the stock market in a single day.

    DeepSeek

  19. 24 February 2025

    Claude 3.7 Sonnet

    You choose whether it answers at once or thinks it over. With it comes Claude Code, for programming from the terminal.

    Anthropic

  20. 27 February 2025

    GPT-4.5

    OpenAI's biggest model so far, more natural in conversation, but expensive: it shows the limits of just growing.

    OpenAI

  21. 25 March 2025

    Gemini 2.5 Pro

    Google takes the lead with a model that reasons before answering and reads very long documents.

    Google

  22. 5 April 2025

    Llama 4

    The first Llamas designed from the start for text and image. They get criticism and will be the last big Llama.

    Meta

  23. 29 April 2025

    Qwen3

    The most downloaded Chinese open family, in many sizes, from phone to server.

    Alibaba

  24. 22 May 2025

    Claude 4

    Able to work alone for hours on long programming tasks: the step from chat to agent.

    Anthropic

  25. 9 July 2025

    Grok 4

    Elon Musk's company enters the elite, with real-time search and a $300-a-month version.

    xAI

  26. 11 July 2025

    Kimi K2

    A trillion parameters, open and free, designed to use tools and act on its own.

    Moonshot AI

  27. 7 August 2025

    GPT-5

    A single system that decides for you whether to answer quickly or stop and think, and available even for free.

    OpenAI

  28. 18 November 2025

    Gemini 3

    The first Gemini to reach Google's search engine on its launch day.

    Google

  29. 2 December 2025

    Mistral 3

    The big open European bet: a huge, free model with a free licence, which also sees images.

    Mistral AI

  30. 7 April 2026

    Claude Mythos Preview

    So good at finding security flaws that it is not released: only used by companies that protect critical software.

    Anthropic

  31. 8 April 2026

    Muse Spark

    Meta abandons Llama and launches a closed model, the first from its “superintelligence” lab.

    Meta

  32. 24 April 2026

    DeepSeek V4

    Open and cheap again, and now it reads a million tokens at once, about 750,000 words.

    DeepSeek

  33. 9 June 2026

    Claude Fable 5 y Mythos 5

    The power of Mythos reaches the public as Fable, with safety filters; Mythos 5, the same model with looser filters on cybersecurity, stays for verified professionals.

    Anthropic

  34. 30 June 2026

    Claude Sonnet 5

    Almost as capable as the expensive models, cheaper, and the one used by default by those who don't pay.

    Anthropic

  35. 17 July 2026

    Kimi K3

    The biggest open model released to date, 2.8 trillion parameters, close to the best closed models.

    Moonshot AI

  36. 24 July 2026

    Claude Opus 5

    It comes close to Fable 5 at half the price: the most advanced becomes cheaper within weeks.

    Anthropic

  37. 1 September 2026

    Claude Fable 5.1 y Mythos 5.1

    Improvements in science and programming, and up to 45% cheaper on long tasks.

    Anthropic

  38. 3 September 2026

    GPT-6 Astra

    OpenAI's generational leap: it can operate a computer for you and is cutting-edge in cybersecurity and science.

    OpenAI

  39. 22 September 2026

    Claude Opus 5.5

    Performs like Fable 5.1 and costs 40% less than Opus 5.

    Anthropic

  40. 30 September 2026

    Gemini 4 Argon

    Google's new generation, on a par with Astra and Opus; for now, only for cybersecurity experts.

    Google

Between the first ChatGPT and the first model that reasons before answering, less than two years pass, and when one company launches an idea, the others have it within a few months. There are few companies, because training one of the big ones costs tens or hundreds of millions in computers and electricity, and it is best not to confuse them with their products: OpenAI makes the GPT models, and ChatGPT is the application for talking to them.

Why it gets things wrong with such confidence

Now the key sentence makes sense: an LLM produces the most probable text, not the truest. Nobody has taught it what truth is; it has been taught how a text usually continues. Almost always the two coincide. Almost.

When they don't coincide, the model doesn't fall silent or say “I don't know”: it writes something that sounds perfect and is false. It is called a hallucination: a book that doesn't exist, a quotation nobody said. And it says it with the same confidence as what is true, because for it there is no difference. It happens most where there was least text to learn from: exact figures and dates, proper names, the very recent (it knows nothing after its training, unless the application searches the internet for it) and the very local. The best way to measure it is to ask it, without letting it search, about something you know in depth: the Falla final of a particular year, the history of your neighbourhood, the details of your trade. The mistakes you find there show how far to trust it on what you don't know.

Biases are the other problem, and they are not a flaw of the machine but a mirror. If in the texts it learned from nurses are almost always “she” and engineers “he”, the model will tend to repeat it. It has no prejudices of its own: it has ours, and sometimes magnified.

None of this makes it useless. It is good at working with words, which is what it is made of: summarising, explaining something difficult another way, making a first draft, translating. And, since it understands images, reading you an invoice or a sign in another language. It is bad at what demands being sure: deciding for you, a medical or legal diagnosis, replacing someone who knows. I use it to prepare the questions, not as the final answer, and what matters to me I check in another source.

What you write goes into the cabinet too

What you write in a chatbot doesn't stay on your phone: it travels to a company's computers, almost always in another country, and in the best-known services, unless you switch it off, it may end up among the texts the next model is trained on. That is, inside the cabinet.

A simple rule: if you wouldn't say it out loud on the bus, don't write it. The Spanish Data Protection Agency published in January 2026 the decalogue “Cuidado con lo que le confIAs”, which goes along the same lines: no ID number, no address, no medical or bank details, no photos of other people, even less of minors, no confidential information from work. And it asks you to check the answers and not to replace a professional with a chatbot, not even in emotional matters.

The three most used services have settings for this. In ChatGPT, the option “Improve the model for everyone”, inside Data controls, can be switched off, and temporary chats are deleted after 30 days. In Claude, Anthropic asks each user since late 2025 whether they allow their conversations to be used for training: if they accept, they are kept for up to five years; if not, 30 days. In Gemini, the setting is called “Keep Activity” and temporary chats are deleted after 72 hours. It is worth spending five minutes looking at them.

In addition, the European Union has had since 2024 an artificial intelligence regulation that applies in phases. Since February 2025 it bans some uses, such as the social scoring of citizens, and since 2 August 2026 it requires a system to tell you that you are talking to a machine and to label generated images and videos. The obligations for high-risk systems, those that decide on a loan, a job or a public benefit, were postponed in July 2026 until December 2027.

Lady Lovelace's objection

In 1843, Ada Lovelace warned that Babbage's machine “has no pretensions whatever to originate anything”. A century later, Turing replied to her with another idea: not to program an adult mind, but a “child machine” that could be educated. Both were right, and together they describe fairly well what is behind a chat window today.

The child machine exists. It has been educated on almost everything we have written, and that is why it speaks like us, gets things right like us and gets things wrong like us. But Lovelace's objection still stands: what it says it does not originate. It was originated by the people who wrote before, those who corrected it afterwards and the one who now types the question. In Schönbrunn, the audience looked at the figure when they should have looked at whoever sat on the other side of the board, because that person decided which move was worth making. With artificial intelligence, that person is you.

Written by

Nando Muñoz

Frontend tech lead · AEM, EDS & SFCC · Cádiz · Madrid