How to play
Warning: Depending on the seed, this game may be DIFFICULT. We
are working on features to make it easier.
Your goal is to find five hidden words. You get 10 guesses for each
word.
- Similarity: The percentage shows how close your guess (or
another word) is in meaning to the hidden word. Higher is
better.
- Word equation: After making a guess, an equation combining
your guess with one or two other words is displayed. This is
a clue for which direction to explore.
- Decompose / auto guess: After four guesses, spend one more
to let the engine build the hidden word from two, or
occasionally three, clue words. The fit percentage measures
the complete combination.
- Concept guesses: These show where the hidden word falls
between two opposite ideas, such as singular ↔ plural. The
percentages show which side the hidden word is closer to.
For example, the game learns the idea of plurality from patterns
like cat → cats, dog → dogs, and house → houses.
More detail
Example: hidden word winter, guess holiday → X = holiday + 0.4 × snow + 0.3 × season. The
game will display the similarity of holiday, snow, and season to the hidden word winter.
Guessing the hidden word, or a close form of it like employed for employment, solves the round. Each word guess
scores its similarity, and solving early adds up to 2,000
points. Concept probes consume a guess but add no similarity
points. Decomposition probes also consume a guess and add no
points, but produce a combination chosen independently of your
previous word. Give up to skip a word, then share your five
results.
Similarity does not directly measure meaning or define
a category; it measures how similarly words are used in the
training corpus. This is a limitation, and results may reflect
biases in that training data. For example, a very
common concrete noun may not be close to tangible if those words rarely appear together.
Additionally opposite meanings can have very high similarity
scores if they are often used together for contrast e.g. hot and cold.
How it works
Words are 300-dimension vectors from Word2Vec.
Similar vectors represent words used in similar contexts.
Hints use vector arithmetic: the app builds the arrow from your
guess to the hidden word X and picks a relevant word with a
positive projection along it. The display shows this as X = your guess + the hint path. It can then fit
a distinct second word against the remaining error and refit
both positive coefficients together. The two-word form is shown
only when its nearest real-word landing is significantly closer
than the best one-word result. Hints come from a curated pool,
are at least 25% related to the answer, and prefer words at
least 30% related to your guess. Every candidate is capped at
your guess's answer similarity plus 30 percentage points
(within a 50–90% range). If none meet the preference, the engine
uses the best otherwise-safe fallback under the same cap. Hints
avoid word-form repeats and never name the hidden word.
How it works