Embedding Tangram

Arrange the concepts so the embedding geometry holds. You are now a t-SNE plot with commitment issues.

How Embedding Tangram Works

An embedding is just a map: similar things land close together, unrelated things drift apart. Here you are the model. Drag the word tiles around a 2D board until every relation in the checklist is satisfied. When a rule holds, it turns green. Turn them all green to project to the next level.

  1. Drag word tiles anywhere on the board (touch or mouse)
  2. NEAR: the two words must sit close together — synonyms hug
  3. FAR: the two words must be pushed apart — unrelated concepts repel
  4. OPPOSITE: the two words must be far and flung to opposite ends of the board
  5. Satisfy every rule to clear the level. Fewer total moves is a better score.

Why Is This Hard?

Real embeddings have hundreds of dimensions to spread concepts out. You get two. Squeezing every relation into a flat plane means some constraints fight each other — welcome to the curse of dimensionality, served on a single sheet of paper.

Slop Fact: Word2vec famously implied "king − man + woman ≈ queen." It also implied a lot of things nobody wanted to publish. The latent space contains multitudes, most of them embarrassing.

Back to the Slop