think like an llmdaily vector puzzle

what is an embedding?

1 · words become numbers

language models can't read. before a model can do anything with a word, it converts it into a long list of numbers called an embedding. the model used by this game turns every word into 3,072 numbers. that list is the word, as far as the machine is concerned.

"sushi"[0.12, −0.48, 0.91, …]3,072 numbers
an embedding is just a coordinate — a very long one.

2 · similar meanings land nearby

the numbers aren't random. the model learns them by reading absurd amounts of text, and words used in similar ways end up with similar numbers. if you could see all 3,072 dimensions at once, you'd see a map of meaning — foods near foods, animals near animals. distance on this map is similarity.

close = similar meaningfar = unrelatedsushipastapizzatacodogcatwolfmoonstar
a 2D shadow of the real 3,072-dimensional map. real embeddings cluster exactly like this, just in more dimensions than we can draw.

3 · directions mean things

because embeddings are coordinates, you can do arithmetic on them. subtract man from king and add woman, and you land near queen. the "maleness → femaleness" direction is a real, reusable arrow in the space. so is "japan → italy", which is why sushi − japan + italy lands near pasta.

− man + womansame directionmanwomankingqueen
concepts like gender, nationality, or cuisine become consistent directions you can add and subtract.

4 · how this game scores you

every daily puzzle computes a target vector from its words — an arithmetic result, a midpoint, a shared neighborhood. your guess gets embedded too, and scored by the angle between the two vectors (cosine similarity). we stretch that onto a 0–100 scale where 100 is the best answer we know of.

targetspaghetti · 100pasta · 93stapler · 0smaller angle → higher score
your score is the angle between your guess and the target — computed, not judged.