true; i think this was just path of least resistance. i don't think the hansen material sets it up with XX' and potential p>n cases [that's for Bach or Wainwright, you go try that; too rich for my blood]

⅟ (Xᵀ * X) being the inverse operator applied to XᵀX is such a cursed combination of programming and math idioms it is kinda amazing.

extremist static typing and proofs are a match made in heaven - check out how we get OLS

GitHub - apoorvalal/lean-hansen-econometrics: formalizing econometrics formalizing econometrics. Contribute to apoorvalal/lean-hansen-econometrics development by creating an account on GitHub.

been trying to get my claw to teach me Lean - send PRs cuz misery loves company
github.com/apoorvalal/l...

AI Adoption (2026), colorised.

Anisotropy Is Inherent to Self-Attention in Transformers The representation degeneration problem is a phenomenon that is widely observed among self-supervised learning methods based on Transformers. In NLP, it takes the form of anisotropy, a singular property of hidden representations which makes them unexpectedly close to each other in terms of angular distance (cosine-similarity). Some recent works tend to show that anisotropy is a consequence of optimizing the cross-entropy loss on long-tailed distributions of tokens. We show in this paper that anisotropy can also be observed empirically in language models with specific objectives that should not suffer directly from the same consequences. We also show that the anisotropy problem extends to Transformers trained on other modalities. Our observations suggest that anisotropy is actually inherent to Transformers-based models.

arxiv.org/abs/2401.12143 forgot why i read this but i did

Bring back self censorship

That's what aperture science was. Portal was an RL env

Nice, yeah pretty funny it settled on quite similar ui too (codex likes its pastel backdrop html).
I think with appropriate metadata scouring gh for different variations on the same idea should be feasible, although licensing etc could be a minefield.
In this case, parallel branches+merge would do

to kick the tires, poke around with my edits of the cars dataset figure here and mangle it some more
lalten.org/vega-ui/char...

quite nice prototype built over telegram: apply finishing touches on your figures in a web-ui using the magic of vega json graphics and get back json/code that you can put back in your source-code [thereby maintaining reproducibility of output - usually my biggest bugbear with wysiwyg fig edits]

prototype seems to work; source here github.com/apoorvalal/v... and
deployed here lalten.org/vega-ui/

1) get starter plot from altair and extract json; then paste into vega-ui
2) edit [fig1:every 'apply' button mutates the vega json directly]
3) when done, extract python/json [fig2]

this is a nice idea, although i wouldn't use matplotlib but altair [since it has vega underneath and that fully contains the data + code so editing it can probably permit a clean round trip back into source code].
Set off a job on my openclaw; will report back with a link if promising.

GitHub - apoorvalal/mlodies: glues pretrained models to do stem separation and transcription to help learn music by ear. glues pretrained models to do stem separation and transcription to help learn music by ear. - apoorvalal/mlodies

github.com/apoorvalal/m... Demucs + basic pitch might help with what you want; I glued things together here for my own ear training and it works surprisingly well

Linear Programming for Fun and Profit How we use an eighty-year-old algorithm to find arbitrages in the cloud market.

modal.com/blog/resourc... linear programming done well

ha Zellij way to spook the normies. Ghostty tabs are fine.

Making sense of Power-tweets is a pretty good case for reasoning models.
chatgpt.com/share/69c609... is HCR an OG minimax result?
I‘d seen the Loewner order version of CR and the associated ellipsoid claim is natural. HCR as ’a secant body in chi-squared space’ makes my head hurt.

the aristotelian method is when you are such a boring teacher and wont shut up about rocks having pneuma that your ideal student runs away and conquers most of the known world

ha yeah the least squares solver is blazing fast [which came as a surprise to me - rust libraries rely way less on BLAS/LAPACK and perf opts from mkl].

Painstakingly detailed documentation; this is a good intro to binding rust

apoorvalal.github.io/crabbymetric...

Codex is great at Rust, i am not. I have strong preferences and tests for what a lean statistics package should do, it does not.
I'm not crazy enough to try to patch out numpy deps but that really is it.

Solid collaboration, now on pypi (uv add crabbymetrics)
github.com/apoorvalal/c...

allpoetry.com/colonel-faza...

now you have to write the paper.
Mondrian Orderings for Interactive Sequential Tests using
Sequential Hierarchical Additive Regression Trees

Structured Hierarchical Additive Regression Trees

based on recent history i think it would be called something wonderfully juvenile like "deezNUTS" or "YUTS"

agree it's tricky; i think for SEs the standard GMM style sandwich representation should be valid [Conley is written with generic moment conditions in mind IIRC]. Optimal bandwidth for RD is harder because the MSE representation in sth like CCT is specific to local linear.

no super clean solution i'm afraid; pandoc-ing the tex->markdown conversion and then getting an LLM to iterate on the conversion has reasonable odds of success but tex is too messy to reliably automate

bsky.app/profile/apoo...

Apropos of econometrics discourse du jour, people are often trying to model a multiplicative shift in the rate of some event, and OLS really is just bad for this. Poisson is good.

lalten.org/pages/counti...

(Aside, quarto papers are better than pdfs, esp for our clanker assistants)

Flip the (now somewhat hack) metaphor for human-AI augmentation.

AI used well with human in the lead: Centaur
AI used poorly with prose and code smells: horseface

Same ingredients and proportions, very different outcomes.

apoorvalal.github.io/pyensmallen/
Pyensmallen now has a nice website with benchmarks (it is very fast!) and API docs. An LLM agent also helped me figure out how to patch a weird BLAS error that cropped up on older mac metal; should now work out of the box.