gist.github.com/apoorvalal/6... here are some test scripts and benchmarks comparing CPU w GPU for matrix multiplication, least squares, MLE, nnet operations. 50x+ speedups abound.

self-explanatory ones

Yeah doing this by hand in 2025 is definitely vehemently ignoring the bitter lesson but it's vaguely artisanal and i think it's fun lol. Let me think about the labelling problem; agree in principle

Tbd, really; i can already see some predictable ocr bleed (ad copy entering text) that can likely be postprocessed away. I'll text you once this run is done with some examples.

Still this surya library is miles ahead of layoutparser for this kind of text data.

(pixar didn't do very much)

here's a caveman linux searcher that works well. assuming your search service is on its last legs, you could write a little script that
0) pulls and extracts these zips into a folder
1) execute ./seeker.sh files/
2) regex search, fzf navigation, open in vim
gist.github.com/apoorvalal/2...

how big is that google drive folder? it's been zipping for like 20 mins for me

google scholar pdf reader's citation popups are sending me down interesting rabbit-holes. good job goog.

yeah i suspect they'll be quite parsimonious actually, you could even try a few simple options in spacy and spot-check performance before going for a big model off huggingface. unclear if an LLM running on top will add value but reasonable people seem to disagree with me and put everything in chat

was going to apologise for giving the gift horse a dental examination but since your other pasttime is a genuinely masochistic game i think debugging cuda issues is preferable bsky.app/profile/paul...

assuming you went ahead with postgres, i think you could generate embeddings with some reasonable model (openai, or huggingface.co/nomic-ai/nom...), insert it into your db, and then use pgvector to do nearest neighbours instead of a string search?
github.com/pgvector/pgv...

would be cool to have

arxiv.org/abs/2307.11255 i learned a lot from this review fwiw

applied econometrics basically restricts itself to this by sticking to stacking heaps of least squares and (with a few exceptions) has basically missed the boat on the last 40 years of computational advancements.

Lol to stretch the metaphor beyond breaking point: twine is used to move the wheels? Cut a chunk off a wheel of cheese?
twine.readthedocs.io/en/stable/

github.com/py-econometr... i've had to explain why the regression library that goes fast because of ducks has this logo

I use "scikit-learning" (complimentary)

not sure why What we do in the shadows hasn't crobarred in a set-piece here