An NSFW Filter for Marginalia Search

Marginalia Search just flipped on an optional “keep it clean” filter, and the comments instantly turned into a spicy taste test. The indie engine’s creator says they tried everything—old-school word lists, a fast Facebook-era classifier, even letting open‑source bots label tens of thousands of pages—before settling on a tiny, from‑scratch neural network that runs fast on everyday CPUs. Why? Speed over splashy tech.

That “no AI” slogan on Marginalia Search lit the fuse. One newcomer quipped that with this change, the promise is broken, sparking a semantics brawl: is a simple neural net “AI,” or just clever math? Others piled on with armchair solutions. One camp waved in modern tricks—“use embeddings and decision trees,” “fine‑tune a BERT,”—while the maintainer reminded folks those won’t fly when you need results in milliseconds.

Down another thread, a pragmatic voice asked why sites don’t just label themselves as adult. The chorus reply: cute idea, almost nobody does. Meanwhile, the builder called the whole effort a “meandering project,” which the crowd translated as: it’s hard to teach a robot what’s risqué without slowing search to a crawl. Oh, and the first fast model flopped because the training samples were too “saucy‑adjacent,” proving garbage in, garbage out. The vibe? Admiring the hustle, debating the label, and joking that robots are now taste‑testing the internet so we don’t have to.