Nobody programmed the machine to know what a mango looks like. It worked it out — from examples, and from being wrong thousands of times. Today you'll hear the 80-year story, then train a real model with your own hands and watch it learn while you look at it. 👀
Artificial intelligence didn't arrive in 2022. It's older than your grandparents, it has been declared dead twice, and it was rescued each time by people everyone else had stopped listening to. Tap each year.
Two of the years on that line changed everything. So instead of telling you about them, here they are as games. Play 1950. Then play 1969 — and find out for yourself why the whole world gave up on AI for seventeen years.
Turing said: stop asking "can it think?" Ask instead — can you tell? Six messages. You decide. (Careful — this is harder than it looks.)
Four mangoes. Two sweet 🟢, two sour 🔴. Your job is simple: separate them with ONE straight line. Sweet on one side, sour on the other. Turn the two dials until you get it.
Every AI on Earth — the one that spots faces, the one that writes essays — is built from billions of copies of the tiny thing below. It does exactly three things: it multiplies, it adds up, and it decides. That's all. Move the sliders and watch it make up its mind.
🥭 You're buying a mango. You look at two clues — but they don't matter equally. How much each clue matters is called its weight. Set the weights, then feed it a mango:
👩🏫 The grown-up name for what you just did: you built a perceptron — invented in 1958. In real notation it's written
output = step( w₁x₁ + w₂x₂ + b ).
Your sliders were the weights (w), the mango's clues were the inputs (x), and the threshold was the bias (b).
You did the whole thing with your fingers. That's not a simplification of AI — that is AI, at its smallest size.
That last bit was you being clever. Machine learning is the machine being clever instead. Below is a real model — not a video, not a pretend one. You show it examples. It guesses. It measures how wrong it was. It nudges its own weights. Then it does that a few hundred times a second while you watch.
The line started random. The model made a prediction for every mango, compared it to the truth, and computed the loss — one number saying how wrong it was overall. Then it worked out which direction to nudge each weight to make that number smaller, and took a small step that way. That is gradient descent, and the size of the step is the learning rate. Repeat until the loss stops falling.
ChatGPT was trained the exact same way. Same loop, same idea — just with billions of weights instead of three, and text instead of mangoes. You have now seen the actual engine. 🚂
Wipe it. Now drop only sweet mangoes on the left side and only sour ones on the right, and train. Easy — it learns instantly.
Now wipe again and scatter both kinds mixed up everywhere. Train. Watch the loss refuse to drop.
The machine is not stupid. Your data is. One straight line cannot separate a mess, and no amount of training fixes bad examples. This is the single biggest reason real AI projects fail — and you just proved it with your own fingers. 🧠
This surprises people: a language model doesn't plan a sentence. It predicts one word, sticks it on the end, then looks at everything again and predicts the next one. Build a sentence and watch it guess.
Where the percentages come from: during training the model read an enormous amount of text and counted, in effect, which words tend to follow which. Now, given everything so far, it produces a probability for every word it knows. Picking the top word every time makes it dull and repetitive; picking with a little randomness makes it lively. That dial is called temperature — and it's why the same question can get you slightly different answers twice.
Anyone who tells you AI is pure magic is selling something. Anyone who tells you it's useless is not paying attention. Here's the truthful middle.
A language model has no idea what a mango tastes like. It has seen the word mango beside the words sweet, juicy and yellow, millions of times. That's not the same as knowing — and it's why it can write a beautiful, confident paragraph that is completely wrong. People call that hallucination. Always check anything that matters.
A model learns from whatever examples it's given. If the examples were unfair, the model becomes unfair — quietly, at enormous speed, while looking like neutral mathematics. Think about it this way. If a bank trains a loan model only on customers who already had bank accounts, it quietly learns that market women who trade in cash are "risky" — not because they are, but because they were never in the examples. Face systems trained mostly on light skin have failed badly on dark skin for the same reason. Bias in, bias out. So always ask the Nigerian question: who was counted, and who was left out?
Your model found a line between sweet and sour mangoes. It did not care. It will never care. Deciding what is worth building, who it should serve, and when to switch it off — those stay human jobs, forever. That's not a small consolation prize. That's the actual work.
Ask most AI models a question in Yorùbá, or Igbo, or Hausa, or Pidgin, and watch them struggle. Not because our languages are hard. Because we were not in the examples.
Over 500 languages are spoken in this country. Millions of Nigerians think, pray, trade and dream in a language that the biggest models on Earth barely hear. A grandmother in Kano who speaks only Hausa cannot ask a machine anything. A trader in Onitsha cannot get farm advice in Igbo. That is not a technical accident. It is a gap where nobody built anything yet.
And here is the part that should make you sit up: Nigerians are already closing it. Nigerian researchers have recorded over 1,800 hours of speech from more than 5,000 native Yorùbá, Igbo and Hausa speakers and given the whole thing away free so anyone can build with it. Nigerian teams have built working text-to-speech for our languages — the very button you have been pressing on this page is meant to run on their work.
They started exactly where you are standing right now: understanding weights, and loss, and training loops. Which you now do. 🧠
The mathematics on this page belongs to everybody.
The question of what we build with it belongs to you.