Vector Tug-of-War

Score 0 Streak 0 --

How Vector Tug-of-War Works

This is word2vec cosplay. Back in 2013 a neural net discovered that you could do arithmetic on meaning: take the vector for "king", subtract "man", add "woman", and the nearest neighbor was "queen". Spooky. We turned that party trick into a puzzle and removed the GPU.

  1. Most rounds are an analogy: A is to B as C is to ? — tap the word that completes the pattern.
  2. Harder rounds are equations: drag the given word tiles into the + and − slots so the result lands on the target concept, then hit Embed.
  3. Solve before the embedding budget runs out. Speed and streaks multiply your points.
  4. One wrong guess breaks the streak. Three misses on the clock and the run collapses.

Why Is This Hard?

The relationships hide in latent space: capital-of, gender, plural, tense, antonym, category. Your brain has the embeddings; you just have to traverse them faster than gradient descent. Allegedly you have a CS degree, so this should be trivial. Allegedly.

Slop Fact: The original word2vec analogies also confidently produced things like "doctor − man + woman = nurse", which is how the field learned that your training corpus is, in fact, your personality. We kept only the wholesome ones.

Back to the Slop