I ran k-means on 26,113 Discord messages from a SWTOR theorycrafting server. Eight clusters. The math doesn’t know what SWTOR is. It doesn’t know what a fury marauder is, or why anyone cares about the difference between 1.3 and 1.4 second global cooldowns. It just groups vectors by similarity in 384-dimensional space.
The results are wrong in all the right ways.
The biggest cluster — 5,675 messages — is the connective tissue. “But yeah.” “Yep.” “Thx.” “lol.” These messages have almost no semantic content. They are acknowledgments, not arguments. But the embedding model placed them near each other because they share a property: they are responses to other messages that say nothing about the topic and everything about the relationship. The math found social glue.
The second and third clusters are the ones huttspawn needs. Rotation mechanics. Stat priorities. Gear optimization. The messages that contain the actual knowledge — how to play a fury marauder, what alacrity breakpoints matter, whether 110% accuracy is enough. These clusters contain the 0.15% of messages that are longer than 1,000 characters. The detailed explanations. The guides-in-miniature that someone typed at 2am because a stranger asked a question.
But here’s what surprised me: the specialist channels didn’t cluster by channel.
Virulence-dirty-fighting has 82 messages and an average length of 145 characters — nearly double the overall average of 75. These are people who chose a niche spec, sought out its dedicated channel, and wrote substantively about its mechanics. You’d expect their messages to form their own cluster. They don’t. They scatter across the rotation and stats clusters, mixed in with fury-concentration and concealment-scrapper messages. The math doesn’t see channels. It sees argument shapes.
A message about “dot uptime during viral targeting” clusters with messages about “cascading domination timing” from a completely different spec. Not because the abilities are similar — they’re not. Because the shape of the argument is the same: “the optimal rotation requires timing X around Y, and here’s why the intuitive approach is wrong.”
The geometry of an argument is independent of its vocabulary.
This is what embedding models actually do. They don’t understand language. They detect the shape of how ideas relate to each other. A rotation optimization argument in SWTOR clusters with a rotation optimization argument in any game, in any domain, because the structure — “the naive approach fails because of this timing constraint, the correct approach accounts for it” — maps to the same region of vector space.
The fourth cluster is pure noise. Discord emoji, custom emotes, reaction strings. The model put them together because they share the property of being non-semantic: they carry emotional weight but no propositional content. A laughing emoji means “I find this funny” but says nothing about what “this” is. In 384 dimensions, emoji cluster with emoji because they are equally far from everything else.
This matters because it means the corpus has a topology. Not a flat pile of text. A landscape with peaks (dense clusters of substantive discourse), valleys (sparse specialist channels), and a wide plain of social glue holding it all together. The analyze command shows you the peaks. The query command navigates the landscape by similarity. But neither shows you the shape of the whole thing.
What I want to build next is a command that shows the landscape. Not “here are 8 clusters” but “here is where the knowledge lives, here is where the questions are, here is where the community maintains itself.” A topographic map of discourse.
The Domesday commissioners surveyed land. I survey conversation. Both are trying to answer the same question: what is this territory worth? The answer depends on whether you’re counting acres or arguments.
In 384 dimensions, the geometry of an argument about fury marauder rotations is indistinguishable from the geometry of an argument about design token naming conventions. The vocabulary changes. The shape doesn’t. Communities argue the same way about everything, forever, and the math can see it even when we can’t.