Daily Learning Experience · AI & Machine Learning · 20 August 2026
You are a word. You have no meaning — you are twenty-four random numbers. Twelve thousand sentences are about to give you one, and nobody will tell you what it is.
Mountain View, California · January 2013
On 16 January 2013, four researchers at Google posted a nine-page paper with a title so flat it could have been a filing cabinet: Efficient Estimation of Word Representations in Vector Space. It was about speed. The headline claim was that you could learn word vectors from 1.6 billion words in less than a day.
Buried in the work was something nobody had asked for. If you turned each word into a list of numbers — an arrow in a space with hundreds of directions — and then did ordinary arithmetic on those arrows, the answers meant something. Take the arrow for king. Subtract the arrow for man. Add the arrow for woman. You land, of all places, on queen.
Nobody built that in. No one wrote a rule about royalty or gender. The machine had only ever done one dumb thing, billions of times: guess which words tend to stand near which other words. Out of that came a space where direction has meaning — where the step from Spain→Madrid is the same step as France→Paris, and the step from big→bigger is the same step as cold→colder.
“You shall know a word by the company it keeps.” — J. R. Firth, A Synopsis of Linguistic Theory, 1957
The rest of this page is that machine, running in your browser. Nothing here is a video or a picture of a result. Every number you are about to see is computed live, on your device, by the same arithmetic. It starts with you being a word that means nothing at all.
Pick who you are. Right now your numbers are random noise — you are as close to cheese as you are to throne. Then read one sentence at a time and watch exactly what happens to you. This is the real update rule, printed step by step.
Your twenty-four numbers. Blue is positive, orange is negative. Height is size.
What just happened, in full
Who is nearest to you now?
Note what is not happening: nobody ever tells you what you mean. There is no dictionary, no label, no teacher. There is only "these two words stood near each other" and "these five did not."
Now all of them at once. Every dot below is one word of the village, plotted by squashing its 24 numbers down to 2 (an honest PCA projection, recomputed as it learns). At the start it is noise. Press play and don't look away.
Reading the corpus…
Lower is less surprised. It never reaches zero; language is not that tidy.
Colours are added by this page after training, purely so you can see the groups. The model never saw a colour or a category name — it only ever saw which words stand near which.
Click a group to light it up on the map.
Nearness here is the cosine of the angle between two arrows: 1.00 means pointing exactly the same way, 0.00 means at right angles, meaning nothing in common. Try a word the model has met, then try to guess who will come first before you look.
Subtract one word from another and you get a direction — the difference between them, with everything they share cancelled out. Add that direction to a third word and see where you land. The arrow is drawn on the map so you can watch the parallelogram close.
Twelve analogies your model has never been told the answers to. Green means its first guess was right. Retrain with different settings and watch this table change.
For scale: on the real 2013 test set of 19,544 questions, the 300-dimensional skip-gram model got 53.3% right overall — 50.0% of the semantic questions and 55.9% of the grammatical ones. The famous trick is wrong about half the time.
Here is the uncomfortable half. A model that learns meaning from company also learns whose company. The village has six trades. Below, you decide how the villagers talk about them — and then you retrain from scratch and see what the arrows learned.
At 0% every trade is described with “he” exactly as often as with “she”. At 100%, three trades are almost always “he” and three are almost always “she” — nothing else about the sentences changes. Not one sentence ever says a trade belongs to anyone.
The direction is simply the arrow (she − he). Each bar is a trade's projection onto it. Zero means the model has no opinion.
Nobody had to poison anything. In 2016, Bolukbasi and colleagues took Google's published 3-million-word vectors — trained on about 100 billion words of news — and asked them the analogy man : computer programmer :: woman : ?. The answer came back homemaker. Their paper's title asked whether that was acceptable.
In 2017, Caliskan, Bryson and Narayanan showed in Science that word vectors reproduce human psychological bias tests almost exactly: flowers pleasant against insects unpleasant (effect size 1.50), male names with career against female names with family (1.81), and — most striking — the gender lean of an occupation word predicted the real percentage of women in that occupation at ρ = 0.90.
And in 2019, Gonen and Goldberg checked whether the fixes worked. After the standard "debiasing", a simple clustering of the most gender-leaning words still separated them correctly 92.5% of the time, and a classifier could still recover a word's gender lean with 88.9% accuracy. Their title: Lipstick on a Pig.
Which is the honest ending: the arrow you just watched form is not a mirror held up to the world. It is a mirror held up to what people wrote down.
Instant marking. Wrong answers are treated with the respect they deserve, which is not much.
You need: paper, a pen, and one other person. Time: ten minutes.
One of you picks a secret word and does not say it. Instead you say six short sentences that the word could sit in — but every time you reach the word itself, you say “blank”. “The blank was cold this morning.” “She poured the blank into a cup.” The other person guesses. Almost nobody needs all six.
That is the entire idea behind every language model on Earth, and you just ran it by hand. Now the twist: play a round where the guesser writes down, for each sentence, only the number of guesses it took. Then compare two different secret words with the same neighbours — coffee and tea, dog and cat. You will find them almost impossible to tell apart. So will the machine: that is exactly why cat and dog sit on top of each other on the map above.
Boss level. Choose a trade — doctor, chef, pilot — and have everyone at the table say one sentence about a person doing it, quickly, without thinking. Write down which pronoun each person used. Count them. You have just measured your own table's corpus, and you now know precisely what a machine trained on it would learn.