Behind the Words

It began with an obsession with the intricate relationships between words. Traditional thesauri offer lists of synonyms, but we wanted to capture how words actually work in human thought.

A thesaurus might tell you that “sprint” relates to “dash” and “run,” but it won’t show you how it connects to “athlete,” “track,” “competition,” or “speed” — the broader semantic network that gives the word its full meaning in our minds. We wanted to map how ‘hiking’ simultaneously evokes the serenity of nature and the vigor of physical exertion.

A language that absorbs all.

English is great for wordplay, thanks to its thousand-year history as a linguistic mutt. Our foundation comes from Old English — about 25-30 thousand Germanic words like 'heart,' 'love,' and 'life' that still form our emotional core.

After the Norman Conquest of 1066, England's ruling classes spoke Anglo-Norman French. Over several centuries, this added roughly 10 thousand words, enriching our vocabulary for governance, law, and culture. The result was a unique dual vocabulary—we can say both "freedom" (Germanic) and "liberty" (French), "think" and "ponder," each with distinct shades of meaning. Without this merger, we'd use "skycraft" instead of "aviation," "wordlore" instead of "grammar."

Through centuries of empire-building and colonization, English became an extraordinary word-collector, absorbing terms from Hindi, Chinese, Japanese, Arabic, and dozens more languages. Its flexible pronunciation patterns and lack of central authority made it uniquely good at naturalizing foreign words. That's why "yoga" (Hindi) and "tsunami" (Japanese) feel thoroughly English while keeping their original meanings.

Building our million-word map.

It started by compiling a vast word list. Beyond traditional dictionary entries, we included terms that lexicographers typically exclude, like encyclopedic entries, and thousands of proper nouns. Without physical space constraints, we could embrace language’s natural messiness.

We set out in 2017 to build a database of word associations. The scale proved daunting — creating relationship lists for a million words would have cost millions of dollars. The project began with Wikipedia’s vast network of interlinked articles and Wiktionary’s detailed entries. We hired freelance linguistics graduate students to create themed word pools, drawing inspiration from the Library of Congress Classification system to create thousands of topical groupings inspired by millions of American books. We used word frequency and adjacencies from Google Ngrams. Ngrams. But even with this foundation, we had solid coverage for only a tenth of our words—mostly common terms.

The breakthrough came in 2023, with GPT-4. Not for real-time generation (too slow, too expensive), but for expanding our database. We generated lists of word senses and contextual flavors, then fed these back along with our in-house data to generate comprehensive relationship lists. The success rate was high, with only a few thousand problematic terms (mostly vulgarities) and some occasionally prudish responses from the model.

The result was a complete semantic network with nearly a hundred million cross-links between individual words.

Seven steps to anywhere.

Exploring these relationships, we discovered something fascinating: virtually any word in English can reach any other through a chain of meaningful connections, typically in seven steps or fewer. Like Six Degrees of Kevin Bacon for language, you could get from “ocean” to “democracy” through conceptual stepping stones, from “butterfly” to “skyscraper” through chains of associated meanings.

From network to game.

This discovery coincided with an existential challenge. GPT-4 had transformed our capabilities. But this same technology destroyed our market, it made traditional reference tools obsolete. Why would consumers pay for an app to explore word relationships in an era of powerful language models?

During National Science Foundation’s I-CORPS entrepreneurial bootcamp, supported by a $295k seed grant we received in Summer 2023, we conducted mamy consumer interviews. We noticed something striking: while AI could generate endless word associations, people craved structured experiences that revealed patterns in their own thinking. And we heard how deeply New York Times word puzzles had become a cherished daily ritual for many people.

At this crossroads, co-founders Michael Douma and Greg Ligierko found a possible pivot: What if, instead of competing with AI, we turned our word network into something entirely different — a daily puzzle about navigating between ideas?

Finding the sweet spot.

The core mechanic emerged rapidly, but achieving the right balance demanded months of refinement. While words could theoretically connect in seven steps, this created meandering experiences. Four hops proved equally problematic — players reached such unrelated concepts that success felt random. Three hops revealed the perfect tension between challenge and discovery. The challenge became Goldilocks-esque: too many obvious connections made winning arbitrary; too few left players stranded. The best puzzles emerged around semantic bottlenecks—those lateral leaps where players discover connections like ‘ocean’ to ‘moon’ via ‘tide.’ These moments of connection became the heart of the game. Here are examples of different routes:

• sugar → sweet → harmony → peace
• sugar → dissolve → solution → peace
• sugar → crystal → clarity → peace

Building good puzzles meant learning what to avoid. We discovered certain words like ‘wisdom’, ‘invention’, ‘originality’, and ‘clarity’ tend to eat up conceptual space — they’re too easily reachable from most starting points. We banned these as targets and de-prioritized other overrepresented connectors like ‘navigation’, ‘guidance’, and ‘animation.’

Engineering the experience.

We took inspiration from the New York Times' daily word games, especially their category-matching game Connections. But where Connections has one correct solution each day, we wanted to celebrate multiple paths. Our puzzles typically offer a dozen ways to win in three hops, and thousands of solutions for longer paths. The core principle: let players win by making their own paths, following how they naturally connect ideas.

Creating puzzles at scale revealed two key challenges. First, vector-based semantic distances, while computationally efficient, proved inadequate for modeling human-perceived relationships — instead we traversed weighted, directed graphs. Second, cultural bias in word embeddings skewed the solution space toward Western associations, which we addressed through iterative GPT-4 passes with varied demographic parameters.

The hint system faces a sharp computational cliff: near targets, it evaluates 17³ possibilities in microseconds, but when players venture toward longer paths, the complexity grows exponentially with semantic distance. We implemented a ray-tracing inspired sampling algorithm to handle these distant explorations efficiently.

The iOS frontend uses Metal graphics and custom shaders, with a physics engine handling word cloud layout through repulsion forces. Each puzzle starts with a database query that yields roughly 100 thematically relevant words, from which we select seventeen using on-device embedding vectors. The system avoids root-word clutter except for strategic variations ('automatic' to 'automate') that enable part-of-speech transitions. Motion blur masks cloud-generation latency while maintaining 60fps.

Forging new paths.

We hope this game will begin a new category of idea-linking games, where the joy comes from discovering how concepts connect in unexpected ways. Each day brings new puzzles. That original dream of a visual thesaurus? It’s still there in exploration mode. And we’re excited about what routes you will discover, what connections surprise you, and how you navigate through the vast web of meaning that connects all words.