English changeswhile you're looking away.
Twelve years of word embeddings, projected into a single map. Drag through the years and watch a word tear loose from its old neighbors and snap onto new ones — the exact moment a meaning flips.
The chatbot register lights up.
delve, elevate, seamless, captivating — the words language models reach for. Dead flat for eight years, then the whole tight knot ignites at once.
These are normal words, evenly used. Nothing distinguishes this patch of the map from any other.
This is one story. The map holds the rest.
Twelve per-year Word2Vec models (300d, trained on ~1B tokens of Common Crawl each) share one frozen compass, so all years live in the same coordinate system. Every word-year is then projected to 2D with a single joint UMAP — which is why a word can sit still for years and then jump to a new neighborhood the moment its meaning shifts. The map holds 52,894 words; the story above just lights a few of them up.