Yes, i agree that unix commands should be first class. I did this for the super common stuff like ls and cp. As for substitution, I did exactly $ for substitution. You'd do something like ($ rsync -avP $src $dst), but I don't think I ever got around to implementing $() to evaluate forms. If you really need to do that then you have to quasiquote the whole expression and unquote the form you need to evaluate. This has been relatively ok for me though. I never implemented anything like pipes or redirection, I instead just send everything like that to bash.
This is not really relevant to your question, but I regret choosing janet for this, it's too opinionated and hacking on C is not as fun as lisp. I started writing my own version of schemesh in racket, but I never got far enough.
awesome! I have wanted something like this for a long time. Currently I use a janet fork <https://github.com/eshrh/matsurika> with some trivial additions, the most important of which is a `$` macro that does what the `sh` does here. I have two questions:
- I see that `sh` does not take in strings but instead lisp forms. How do you distinguish between variables that need to be substituted and commands? In my fork, the way to do variable substitution involves quasiquoting/unquoting.
- Almost all of the features that make your language good for shell scripting are essentially syntactic features that can easily be implemented as a macro library for say, scheme. Why'd you choose to write in C++? Surely performance is not an important factor here. (I'm interested because I am currently working on a scheme-based shell scripting language).
A lot of wifi sensing results that have high-dimensional outputs are usually using wideband links... your average wifi connection uses 20MHz of bandwidth and is transmitting on 48 spaced out frequencies. In the paper, we use 160MHz with effectively 1992 input data points. This still isn't enough to predict a 3x512x512 image well enough, which motivated predicting 4x64x64 latent embeddings instead.
The more space you take up in the frequency domain, the higher your resolution in the time domain is. Wifi sensing results that detect heart rate or breathing, for example, use even larger bandwidth, to the point where it'd be more accurate to call them radars than wifi access points.
Think of it as an img2img stable diffusion process, except instead of starting with an image you want to transform, you start with CSI.
The encoder itself is trained on latent embeddings of images in the same environment with the same subject, so it learns visual details (that are preserved through the original autoencoder; this is why the model can't overfit on, say, text or faces).
Now many people use https://ps.zpj.io/ which supports some hardware including SDRs, but I must discourage using it, especially for research, as it's not free software and has a restrictive license. I used https://feitcsi.kuskosoft.com/ which uses a slightly modified iwlwifi driver, since iwlwifi needs to compute CSI anyway. There are free software alternatives for SDR CSI extraction as well; it's not hard to build an OFDM chain with GNUradio and extract CSI, although this might require a slightly more in-depth understanding of how wifi works.
I think the results here are much less important and surprising than what some people seem to be thinking. To summarize the core of the paper, we took stable diffusion (which is a 3-part system of an encoder, u-net, decoder), and replaced the encoder to use WiFi data instead of images. This gives you two advantages: you get text-based guidance for free, and the encoder model can be smaller. The smaller model combined with the semantic compression from the autoencoder gives you better (SOTA resolution) results, much faster.
I noticed a lot of discussion about how the model can possibly be so accurate. It wouldn't be wrong to consider the model overfit, in the sense that the visual details of the scene are moved from the training data to the model weights. These kinds of models are meant to be trained & deployed in a single environment. What's interesting about this work is that learning the environment well has become really fast because the output dimension is smaller than image space. In fact, it's so fast that you can basically do it in real time... you turn on a data collection node and can train a model from scratch online, in a new environment that gets decent results with at least a little bit of interesting generalization in ~10min. I'm presenting a demonstration of this at Mobicom 2025 next month in Hong Kong.
What people call "WiFi sensing" is now mostly CSI (channel state information) sensing. When you transmit a packet on many subcarriers (frequencies), the CSI represents how the data on each frequency changed during transmission. So, CSI is inherently quite sensitive to environmental changes.
I want to point out something that most everybody working in the CSI sensing/general ISAC space seems to know: generalization is hard and most definitely unsolved for any reasonably high-dimensional sensing problem (like image generation and to some extent pose estimation). I see a lot of fearmongering online about wifi sensing killing privacy for good, but in my opinion we're still quite far off.
This is pretty neat, I've also encountered the same problem. Personally, I take my notes in emacs org-mode, and I've built up a workflow around anki-editor[0], some snippets and elisp.
I did this for just about a year, and i can confirm this works if you put in the time. I'm at around 3000 recognizable kanji and can watch/read pretty much anything that's not particularly hard.