Mildly difficult = generics + trait bounds + lifetimes. So everyday rust. And it hallucinated like an absolute fucking champion.

For audio, log-signature features complement Fourier-style features (e.g. spectrograms) in the sense that certain characteristics that are noisy or poorly captured by Fourier features are cleanly captured with certain log-signature features. Particularly interesting for audio source separation.
2/2

There are some languages (see also the bracket-fetishist’s 20-load weekend that is Lisp) that I am convinced are PsyOps.

It has an “object of type ‘closure’ is not subsettable”-system

My day to day language is rust rn and today I am arse deep in trait bounds and types were a mistake.

Strong choices. Probably the best at the time. But yeah.

Personally, I’m quite fond of polars in python, but I wouldn’t exactly call it easy to use or ergonomic. But if I were writing a pipeline that needed to be fast and stable that’s what I’d use.

It’s not good.

Econometricians don’t know how to program

Econometricians are applied mathematicians with delusions of grandeur

Oh every method that is only published in AoS is useless in practice. That is not their purpose.

Does anyone really care?

I’ve been told that’s how you know it’s good. Great statisticians don’t explain. They lemma.

My mildest take.

Indeed. The posterior mass is on signals about as dense as a neutron star. And it’s very easy maths to show it dansblog.netlify.app/posts/2021-1...