It’s always the gradient quality huh

Asking Santa Claude for algorithmic improvements under the ICAmas tree

Independent mechanism analysis, a new concept? (NeurIPS 2021) proceedings.neurips.cc/paper_files/...

The 4th-order cumulant [aka Binder cumulant] is a measure used to analyze the kurtosis of a probability distrib’n and has been used as a contrast function in ICA (CuBICA, 2003 papers.ssrn.com/sol3/Deliver...) Cumulant en.wikipedia.org/wiki/Cumulant Binder parameter en.wikipedia.org/wiki/Binder_...

“Machine Learning meets Algebraic Combinatorics: A Suite of Datasets Capturing Research-level Conjecturing Ability in Pure Mathematics” (ICML 2025) arxiv.org/abs/2503.06366 ML4AlgComb github.com/pnnl/ML4AlgC...

Pondering how to apply algebraic combinatorics to ML (noticed Luca Sabatini’s paper is out in LMS Bulletin londmathsoc.onlinelibrary.wiley.com/doi/full/10.... a year after arXiv)

OSS highly invested social norms just tricked me into solving the technical contribution I was trying to walk away from smh

Pro tip: don’t forget to close any abandoned PRs you made this year or Santa won’t come ‼️

"Thin ICA" (2003) - from a whole symposium on ICA & BSS personal.us.es/sergio/PAPER...

People will really comment on an unrelated issue with several feature requests and a (perhaps LLM-generated?) offer of assistance towards them (which in hindsight nudged you to accept them) and then ghost after you make the issues and offer some scoping Congrats you now have unactioned extra tasks

TIL you can compute just part of an SVD ("thin"/"economy" SVD), via the -x variants of the LAPACK operations www.intel.com/content/www/...

cool: the internals of a SVD LAPACK call github.com/rust-ndarray...

All I want for CI-mas is OpenBLAS builds

"It does not matter if you stop developing a library, as long as it is not slow" — Confucius Modern Day

boo, rewrote my ICA lib in faer and it got 3x slower (I think requires more careful use of pre-allocation / special zipping routines / views)

Pure Rust SVD computation with faer gist.github.com/lmmx/7b0dcda... 📃 Docs docs.rs/faer/latest/... ✍️ Guide faer.veganb.tw/docs/getting...

Woah, faer-ext crate has a Polars feature docs.rs/faer-ext/lat...

picard: pre-conditioned ICA (faster than FastICA) pypi.org/project/pyth...

gm to everyone except [checks notes] [edits notes] [re-checks notes] pytest ORT Rust extension dylib path configuration race condition induced segfault?

use of the word “slop” is becoming a tell for lazy criticism. Why was something “slop”? OK sure it was nondescript mush, but what else was it? (Ugly, poorly performant, malformed, … ?) The fact that the term has been taken up to refer to non-machine generated outputs gives the game away. It’s a punt

Came for the FastICA impl, stayed for the outrageously cute cyber Ferris github.com/rust-ml/linfa

Not sure I ever posted polars-nmf (non-negative matrix factorisation plugin github.com/lmmx/polars-...) but now Polars has persistent metadata, I think it’ll enable storing the H matrices you decompose into [while W goes into the DataFrame itself]—discarding them was so unsatisfying I abandoned ship

🧙‍♂️ topicwizard — a modern, model-agnostic framework for topic model visualisation and interpretation arxiv.org/abs/2505.13034 sklearn style topic model pipelines and plotly viewers github.com/x-tabdevelop...

S3: Semantic Signal Separation (2025) arxiv.org/abs/2406.09556 The fastest contextual topic model [4.5x faster than the runner-up BERTopic], notably higher coherence than FASTopic 📦 turftopic github.com/x-tabdevelop...

The atomic ensemble time scale at the NIST Boulder campus has failed.

BERTrend (2024) arxiv.org/abs/2411.05930

“Scatter/Gather Browsing” by Cutting et al. (including Tukey), SIGIR92 dl.acm.org/doi/pdf/10.1...

questioning whether there’d be any genuine intellectual merit in doing clustering on the arXiv embeddings // time series analysis to see them shift

Delighted to see that TOML has released a new version! TOML is overall a great standard but I understand they had limited people with the power to cut a new release. github.com/toml-lang/to...

Made an arXiv explorer so I can catch up on the papers I missed this year 😆 github.com/lmmx/arxiv_e... Allows you to select a given subject area and just download a year's data at a time (searches in real-time!)

New dataset acquired! 🗃️ 🤗 arxiv-papers-by-subject: arXiv catalogue (2.5M papers) split by primary subject area huggingface.co/datasets/per... (Currently topping up 1M+ rows missed on initial run in physics subject areas – if you're just interested in maths/stats/CS it's ready to use)

Graph embedding of a nightingale song www.instagram.com/reel/DSVUb03... Each node = a time frame, distances reflect acoustic similarity. Node size indicates overall energy/salience, color encodes {chroma energy / spectral centroid / spectral flux}, edges = temporal transitions between song elements

Over/under on me regretting letting Claude vibe upload the *squints* 29,000 parquet files to HF Hub and leaving the house…

Writing Python with Claude is wretched wretched wretched wretched

A cool feature for HF Datasets would be to repartition large datasets to accelerate the ‘saved queries’ (autogenerated from frequent use), Polars pushdown queries are still not fast enough to be practical [compared to just downloading] when the dataset is completely unpartitioned

to cook, one must first become fried

“I'll stop perseverating when I'm rate limited” is the new “I'll sleep when I'm dead”

Sometimes people will repair the guilt they feel about expressing their frustration* at code being difficult to write/run by framing their expression as constructive criticism *[which is not necessarily a reflection of the fault of a maintainer, e.g. the person may have been rushing or vibe coding]

Dangerously bored rn

a good text dataset would fix me… I got a DMCA on the BBC News dataset I filtered out of FineWeb (which was an official web scrape from HuggingFace themself… but you can’t rejoinder without entering full on legal proceedings lmao nope) feeling a need to embed something but out of ideas

my new year’s resolution is to be more Playfully Obstreperous

Made a little local real time emoji search app with polars-fastembed to test drive the Snowflake Arctic Embed XS embeddings (emojis plotted in 2D courtesy UMAP) 🎑 github.com/lmmx/emboji [Model card mentions you're meant to add a query prefix, was giving junk before I did that]

how are there five thousand emojis now

Haircut kinda looks like a dragon today

Yes yes it had a good year, now time for "slop" to die

Classic

Luxical One being no use for semantic search with a query fragment could be a blessing to think outside the box... For the Python PEP corpus Claude suggests grouping paragraphs into concepts, identifying 'orphan' PEPs, concept evolution (e.g. type hints), clustering (outliers being unique proposals)

There's a sort of homeostatic withdrawal to frictive development processes wherein, rather than actively noping out of a process [reactance] a developer will simply let the plate spin slightly too slowly to be kept in motion and ultimately fall off in a way that gives plausible deniability of intent

Going to upload my first m*del to H*ggingF*ce [dies of embarassment]

ONNX Model Explorer - a Hugging Face Space by onnx-community huggingface.co/spaces/onnx-...