First screen, captured 4 days ago. Open the video.

E-means: Outperforming k-means in one-tenth the epochs

What HN thinks

Mini-batches, maxi-debate: half the thread is benchmarking it, the other half is asking what k-means is for.

Impressive practical speedups
Several commenters tried e-means in their own indexing projects and saw real gains, such as faster IVF index builds and 2.5x faster bulk reindexing.
Not that novel, scaling oversold
Skeptics say mini-batch k-means with EMA updates is well established and the parallel scaling claims ignore communication costs; the author replied with baselines and multi-node results.
Wrong tool for some problems
For mesh relaxation, where centroids can be computed exactly, sampled mini-batches are only noisier, and others want cluster-quality metrics like ARI or NMI rather than inertia alone.

Consensus: Most agree it is a well-presented, useful speedup for large-scale vector clustering, but choosing k is still unsolved and the novelty is disputed.

Recommended in the comments

HDBSCAN (library): Suggested for cases where you don't know how many clusters to pick.
X-means ↗ (other): Proposed as a principled way to select k.
scikit-learn (library): Its tuned mini-batch k-means was named as the baseline e-means should be compared against.

Summarised by Claude from the comment thread 3 days ago.

Points: 484
Comments: 71
Best rank: #2
Posted: 2026-11-09 by porridgeraisin
Category: AI & ML

k-means clustering e-means meta-fair mini-batch-optimization gpu-kernels faiss

All captures:

Full page

Captured 4 days ago. Click for the whole screenshot.

Looks similar

Screenshot of norvig.com: Chomsky and the Two Cultures of Statistical Learning (2011)

Chomsky and the Two Cultures of Statistical Learning (2011)(norvig.com)

88 points by atomicnature 11 months ago | 68 comments

Screenshot of davpinto.github.io: Feature Extraction with KNN

Feature Extraction with KNN (davpinto.github.io)

18 points by RicoElectrico 1 year ago | 3 comments

Screenshot of edwardbenson.com: Embedding boolean circuits in a neural network

Embedding boolean circuits in a neural network (edwardbenson.com)

19 points by eob 12 days ago | 5 comments

Ted Benson's blog post sketching how to embed a boolean circuit into a trained neural network using threshold probes for input and steering stims for output.

mechanistic-interpretability neural-networks boolean-circuits llm

Screenshot of sqrtminusone.xyz: 916 Days of Emacs

916 Days of Emacs (sqrtminusone.xyz)

238 points by d-s 3 years ago | 265 comments

Screenshot of seriot.ch: A Minsky machine in ncurses terminfo

A Minsky machine in ncurses terminfo (seriot.ch)

29 points by beefburger 1 month ago | 9 comments

Nicolas Seriot's article showing ncurses terminfo parameter expansion can simulate a two-counter Minsky machine, making it computationally universal, with a Fibonacci demo.

terminfo ncurses minsky-machine turing-completeness