December 16, 2025

Sine and whine in the browser line

A linear-time alternative for Dimensionality Reduction and fast visualisation

Browser-fast charts claim victory; commenters fact-check, meme-test, and demand a pip

TLDR: A new method claims near-instant browser visualizations by using landmark “radio towers” instead of heavy comparisons. Commenters cheer the speed but challenge the claims, test it with AI, and ask for packaging, sparking a lively debate over accuracy vs. interactivity.

A new browser-friendly trick called Sine Landmark Reduction (SLR) swoops in promising instant, low-lag visuals for big data, and the comments turned into a tech reality show. The author, romanfll, says he built it because dragging a file into a web app shouldn’t freeze your screen: think “GPS-style” positioning against a few “radio towers” instead of comparing everything to everything. Cue applause… and objections. One camp is hyped about linear-time speed and the demo claim: 9,000 points into 3D in under two seconds. Another camp checks the fine print. jmpeax fires the first shot: “UMAP is not O(n^2) — it’s O(n log n),” calling out the article’s swipe at a beloved tool. Meanwhile, memming nudges the method toward reality: sample a few real landmarks, not just sine squiggles. The practical crowd piles in with the classic dev question: “Is there a pip install?” And then the memes arrive. aw123 asked an AI to try it on the classic digits dataset and posted results, turning the thread into a vibe check between “wow, fast!” and “but is it good?” It’s speed vs. fidelity, browser vs. backend, and the community is gleefully split—big promises, bigger fact-checks.

Key Points

  • SLR is a deterministic, linear-time dimensionality reduction method designed for browser-based visualization.
  • It replaces all-pairs distance comparisons with distances to a small set of landmarks, achieving O(N × k) complexity.
  • SLR uses a synthetic sine-based skeleton or a data-derived skeleton and linearized trilateration to compute embeddings.
  • The method includes refinements such as alpha scaling and distance warping to improve results.
  • A compact Python implementation is provided, and SLR reportedly embeds 9,000 points (50D) to 3D in under two seconds on a CPU.

Hottest takes

“I needed this to run entirely in the browser” — romanfll
“UMAP is not O(n^2) it is O(n log n)” — jmpeax
“i asked an llm to test it on the digits dataset…” — aw123
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.