The essentials · five minutes
Seven ideas to keep
- 1
Two rival answers.
For fifty years, “how do you build a machine that thinks?” had two answers: write the rules, or let the machine find them in examples. The rules camp produced useful systems and made a great deal of money. The rival camp drove its supporters to despair for decades: it lacked oceans of data and cheap computing power. When both arrived, it won everything: in 2012 a neural network crushed an image-recognition contest; in 2017 the transformer made training industrial; in 2022 ChatGPT put these machines in everyone’s hands.
- 2
Learning is turning dials.
A learning machine is a box with millions of settings — far too many for any human to adjust one by one. So an automatic procedure does it: it measures the machine’s error, works out which way that error shrinks, moves every setting a small step that way, and starts again. Picture a hiker in fog: they cannot see the valley, but they feel the slope under their feet, and they walk down. That is not a metaphor; it is literally the algorithm, repeated billions of times.
- 3
Attention, or reading the whole sentence at once.
To understand a word, a transformer looks at every other word in the sentence at the same time and works out which ones matter to it: in “the animal didn’t cross the street because it was too tired”, the little word “it” goes looking for “the animal”. Trained on the dumbest task imaginable — guess the next word — over oceans of text, it learns far more than words: to guess right you have to understand what the text is about, and so, little by little, the model builds a representation of the world.
- 4
They generalise, and nobody completely knows why.
These models do not recite past exam papers: they cope with questions nobody has asked. We can now find, inside the machine, concepts, plans, and explanations that do not match what it actually did. What we cannot tell: whether there is a consciousness.
- 5
Eight labs, eight bets.
The big labs differ not by their rankings, reshuffled monthly, but by their bets — and one axis organises them all: publish your weights, or keep them.
- 6
Heavy industry.
There is nothing immaterial about AI: one company makes the machines that etch the finest chips, one island makes nearly all of them, three firms share the memory — and electricity has become the binding constraint. Those few chokepoints explain most of the geopolitics around AI.
- 7
Neither camp.
Enthusiasts or doomsayers? We are at the very beginning, and the limits, like the risks, are still to be discovered. Understanding how these machines work is what lets you form an opinion that holds — and the essay will ask you for yours.
Want the mechanism behind each of these ideas? Read the full essay — in Discovery (≈ 29 min) or in Immersion (≈ 33 min).