Your guesses and the secret word will be plotted here.
Guess
| # | Guess | Similarity | Proximity |
|---|
No guesses yet.
Loading 44,868 word2vec vectors
0%
Your guesses and the secret word will be plotted here.
| # | Guess | Similarity | Proximity |
|---|
No guesses yet.
Same rules: a secret word, unlimited guesses, and a score that is the cosine similarity between your guess and the answer in word2vec, scaled to 0–100. Proximity tells you where you land among the secret's thousand nearest words.
What is added is the geometry: every guess and the secret are drawn in 3D, with the secret as an unlabelled marker, so you can see which way to go instead of just how far away you are. Drag to orbit, zoom (pinch, or Ctrl/⌘ and scroll), double-click to reset. Point colour always comes from the true similarity, never from where a point lands on screen. The map offers three views:
PCA projects your guesses and the secret onto the three directions they spread out along most. One guess gives a line, two a plane, and from three on the map is truly 3D. Up to that point every distance is exact; after it the note under the map says how much of the 300-dimensional spread survived.
Local PCA puts the secret at the centre and picks the axes from your warmest guesses, which are the ones worth steering by. Every point sits at its true distance from the secret. Only its direction is approximate, and hovering shows how much of that direction the map captures.
t-SNE ignores global geometry and keeps each word next to the words it is most similar to, which shows clusters well. Distances between clusters and the axes mean nothing there. It needs a few guesses before it has anything to say.