For some reason I thought Emperor Penguins where like 6 feet tall. Learning that they're just a meter high is a bit disappointing.

I don't disagree with the career implications. I'm just annoyed by people claiming that what they really care about is highfalutin principles about "art" writ-large when they're really just worried about their own careers.
Those are legitimate worries on their own - why dress it up?

Has this kind of hyperfocus on "critique" always been a feature of American/progressive activism? Or is it new? It feels very academia-brained, like the best possible thing you can do is find new and creative ways to articulate "critiques" of how things are actually problematic.

So far, when I've used this code for internal, day-to-day analyses at work, speed has never been an issue (although I use the Gaussian estimators the most, where are naturally very efficient - I haven't really stress-tested the KNN estimators as much).

The neural estimators can almost certainly be improved w/ GPU integration (I just don't have one, so I never spent time with that). I'm not super familiar with Numba so I don't know much much extra juice it could squeeze out.

I do vectorization and hand off to Nupy/Scipy wherever possible for major bottlenecks (matrix operations, KNN network inference. There's some potential for further optimization if you can get paralleism over loops (esp. for the higher-order info measures), but Python makes that hard.

I have no personal interest in Julia or MATLAB, but I do have Opus 4.5 and might try and see if it can automate the porting/translation process, just to see if it works.

I am more than happy to continue expanding this package. If there's something you'd like to see, feel free to make a request, or roll it yourself. 6/N

The goal is not to supplant other packages (e.g. JIDT/IDTxl are still my go-tos for information dynamics), but my goal was a pure-python package (i.e. no MATLAB, ever) to make everything as accessible as possible. 5/N

I have classic measures (entropy, MI, CMI, etc), modern multivariate measures (O/S-information, TC, DTC, etc), and three different kinds of information decomposition (PID, PED, GID).
As far as I know, this is the most complete package on the market atm. 4/N

I've included estimators for different data types: discrete estimators, covariance-based estimators for Gaussian data, and two different non-parametric estimators (normalizing flow-based and KNN-based). Also mixed, for data with discrete and continuous components. 3/N

This package grew out of my work in graduate school and my postdoc - inspired by the networkx package, I wanted a one-stop-shop for discrete and continuous multivariate information theory, all in Python for maximum accessibility. Years later, here it is. 2/N

The way (some) professors will use rhetoric around "the environment" as a cudgel against AI and then thoughtlessly hop on a multi hour flight to present a 15 minute talk (and let's be real, socialize with their friends from grad school) is so disheartening.

This blows my mind living in MA all the time. The whole state is enthusiastically pushing out a whole generation of young professionals because it's impossible to afford to live here! Happily sending all that future growth (and tax revenue) to sunbelt states.

Very excited to live in a world where everyone has some form of dementia because we've outsourced all of the cognitive demands of daily living to the hungry ghosts in our phones.

I think they did the "H" in Northampton a bit too hard. We usually say "North-ampton". Ditto for Easthampton.