← NetWords

About NetWords

What the game is, why it exists, and how the puzzles are made.

NetWords is a daily word puzzle built around a single question: how far apart are two words, really? Most word games are about letters — which ones, in what order, how many. NetWords is about meaning. You are handed two words that seem unrelated and asked to build a chain between them, one word at a time, until the two sides meet.

Play today's puzzle →

Why it exists

The idea came out of a property of language models that is genuinely fun to play with. To a model, every word is a point in a high-dimensional space, and the distance between two points is a usable measure of how alike their meanings are. That measure often agrees with intuition — and the times it disagrees are the interesting part. Two words you would swear belong together score poorly; two you would never have paired turn out to be neighbours.

A puzzle built on that measure rewards a different kind of thinking than a crossword or an anagram. There is no dictionary trick to learn and no letter pattern to spot. You have to ask what a word actually means, and what other words live near it. The best moves in NetWords are usually the ones that feel slightly too obvious to bother typing.

How a puzzle is made

Every puzzle is generated automatically, one day ahead, by a job that runs once a night. Nothing is hand-authored, and no puzzle ships until the generator has proved it is solvable.

1. A curated word list

The raw material is a frequency-ranked English word list, but the raw list makes poor puzzles, so it gets a curation pass. Out go words shorter than four letters, inflections whose base word is already present (MEANS next to MEAN is a plural, not a step in meaning), proper nouns — names, places, brands — and corpus debris like HTTP and PREV.

Words that are ordinary English but happen to double as names deliberately stay in: BILL, FRANK, MARK, ROSE, BROWN, JACK, TURKEY. Removing them would cost the game more than the occasional odd anchor gains it.

2. A model that measures meaning

Each word is turned into a vector by a pinned, open-source sentence-embedding model (all-MiniLM-L6-v2). The similarity between two words is the cosine of the angle between their vectors, and two words are considered linked when that value reaches 40%. The model weights are fixed at a specific version on purpose: swapping them would silently change the answer to every puzzle ever generated.

The percentage on the board is that raw cosine value, printed to one decimal. The link rule reads the raw number too, so a pair displayed as 39.8% genuinely does not link — the game never rounds a near miss into a connection.

3. Choosing the day's two anchors

The generator builds a graph over the day's vocabulary — every word a node, every above-threshold pair an edge — and then searches for two anchor words whose shortest path through that graph is an interesting length. It scores every candidate pair and takes the one closest to its target difficulty, breaking ties toward the harder pair. Deliberately not the hardest pair available: the most extreme pair on any given day is usually extreme because one of its anchors is a semantic oddity, which is difficult but unsatisfying to solve.

Because the path is computed rather than assumed, the number the game shows you as Shortest path is a real guarantee. A pair with no path between them is rejected outright rather than shipped with a plausible-looking number attached.

4. Keeping the solution honest

One rule exists purely for the feel of the solve. Some words are universal connectors — abstract umbrella nouns like SOMETHING, ITEM or ACTIVITY that sit close to 15–28% of the entire vocabulary. Technically they bridge almost any two words; in practice, routing a solution through them reads as arbitrary, and it makes every puzzle the same puzzle. So a non-anchor word that links to more than 15% of the vocabulary is barred from sitting on a solution path. Measured against the alternative, this reroutes about a third of paths onto specific, meaningful words while leaving difficulty essentially unchanged — the hubs were redundant shortcuts, not load-bearing.

Puzzles are also checked against previous days so the same anchors don't come back around, and the finished graph is precomputed and shipped with the puzzle, which is why the board responds instantly. Words you invent that aren't in the day's vocabulary are scored live by a small service when you type them.

Where it's honestly imperfect

The model measures usage, not dictionary sense, and that leaks in ways worth knowing about. It scores similarly spelled words as similar — HAVEN and HAVE link, which is a spelling coincidence wearing a meaning's clothes. The proper-noun filter was built by hand and isn't exhaustive, so a rare name occasionally slips through as an anchor. And the curated list still contains the odd acronym.

These are known and tracked rather than hidden. A word game whose engine is a statistical model of language is going to have edges, and pretending otherwise would be less interesting than admitting it.

Privacy, in short

You can play without an account, and if you do, your progress and statistics stay in your own browser rather than on a server. There are no analytics or behavioural tracking scripts of our own. The full detail — including advertising and what signing in stores — is in the Privacy Policy.

Who makes it

NetWords is made by Titus Whang, an independent developer. It runs as a web app at networds.app and works the same on phones and desktops. Questions, bug reports and puzzles you thought were unfair are all welcome at tituswhang0930@gmail.com.

Play today's puzzle →