November 7, 2025
Math + Magic = Comment Chaos
Word2Vec-style vector arithmetic on docs embeddings
Mathy text mashups spark wild fanfic ideas and skeptic side‑eye
TLDR: Doc math landed an Angular testing match when task types were tuned. Fans cheered cross‑modal tricks and book mashups, while skeptics questioned verification and relying on a single 0.75 score.
Cue the “King − Man + Woman = Queen” meme: a researcher tried similar word math on whole documents and got “Supabase testing − Supabase + Angular” landing closest to Angular’s Testing docs (cosine similarity 0.75) when they tuned task types. Translation for non‑nerds: they turned docs into numbers, did simple addition/subtraction, and checked which real doc it resembled most.
The crowd erupted in three flavors. Hype squad: antirez jumped in to say it works with images too, dropping a YouTube demo. Chaos squad: londons_explore pitched a decoder that could turn a mashed‑up vector back into an original story—cue the viral joke, “Harry Potter + Lord of the Rings” = Frodo going to wizard school to yeet Voldemort into Mount Doom. Fanfic math is officially trending.
Then came the skeptics. They side‑eyed the single 0.75 result and the “no direct verification” caveat, asking if custom task types were cherry‑picking and pointing to previous threads and more debate. Jokes flew about “vector potions” and “linear algebra cosplay,” while optimists imagined remixing entire docs like recipes. Bottom line: the experiment worked once, the internet wants more receipts, and everyone’s dreaming about pressing the add button on entire universes.
Key Points
- •The article tests whether word2vec-style vector arithmetic applies to modern document embeddings.
- •Two experiments are defined: (1) ‘Same topic, different domain’ and (2) ‘Different topic, same domain.’
- •EmbeddingGemma is used, with both default and customized task types, acknowledging task-type effects noted in Gemini Embedding.
- •Verification uses cosine similarity against embeddings of selected docs from Angular, Supabase, Playwright, CockroachDB, and Skylib.
- •Reported result: With customized task types, the first experiment’s resultant vector is closest to Angular ‘Testing’ docs (e.g., 0.75 similarity for “Testing”).