Words as Rows
This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.
An embedding gives every word a row of numbers positioned so that words used similarly get similar rows — turning "unseen but plausible" into ordinary nearness arithmetic.
Tally, for three words, how often each appears near weather-talk and near kitchen-talk: rain lands at (9, 0), snow at (8, 1), flour at (0, 7). Plot the rows as Unit 8's graph-paper points. Rain and snow sit close together; flour sits far away — pure co-occurrence, no meanings defined, no dictionary consulted. Now the payoff. A model that has seen "the snow melted" can half-trust "the rain melted," because prediction can lean on a near neighbor's statistics instead of demanding the exact string. The zero that killed the counting model becomes a small honest number borrowed from the neighborhood — which is what a similarity ruler is for.
Two columns made the toy work; real embeddings use hundreds — the geometry survives the width.