What programming/technical projects have you been working on?

I recently completed "building" an AI chatbot as a final project for my coursework.

As most people here would probably say, it left me feeling pretty "meh."

The technical meat and potatoes was pretty basic. Install Ollama and a few models to try, throw together a pretty small chunk of python code to read in some documents and feed the model a prompt, and that's about it.

The part that took longest was selecting a model and fine tuning the prompt. My initial prompt for it may have been more lines than the actual code. I had to attempt to give it instructions that it would never completely follow, because AI is just supposed to have a mind of its own I guess?

Seeing my GPU max out for 1-2 minutes to generate nonsensical responses was a bit depressing. I could have put the same amount of time into training a human support agent and they would work with a higher level of accuracy and cause less frustration to everyone involved.

I'm still writing new stuff in my esoteric programming language, Funciton. It uses Unicode box drawing characters and looks a bit like flow charts. Here's what the factorial function looks like:

          ╓───╖
          ║ ! ║
          ╙─┬─╜   ┌───╖  ╔═══╗
     ┌──────┴─────┤ > ╟──╢ 2 ║
  ┌──┴──┐         ╘═╤═╝  ╚═══╝
┌─┴─╖   │    ╔═══╗  │
│ ♭ ║   │    ║ 1 ║  │
╘═╤═╝   │    ╚═╤═╝  │
┌─┴─╖ ┌─┴─╖  ┌─┴─╖  │
│ ! ║ │ × ╟──┤ ‽ ╟──┘
╘═╤═╝ ╘═╤═╝  ╘═╤═╝
  └─────┘      │

Each of the function invocations, including the multiplication (×) and even the if-then-else (‽), are themselves written in Funciton. The only built-in operations are bit shifting, less-than, bitwise nand, function invocation and lambdas. The only data type is the arbitrary-size integer.

Here's a YouTube playlist describing the language and then detailing all of the library functionality I've already implemented, starting with simple arithmetic and building up to more advanced structures like lists. The last video at time of writing is regular expressions, which I created about a month ago.

I am now working on a new data structure which I will call “arrays”. On the surface it sounds basically identical to lists, but lists cannot support arbitrary indexing and therefore you can't have binary search. The implementation of arrays is considerably more elaborate than lists. Stay tuned. I estimate the video will come out within the next 2 weeks (but it's a hobby project so I make no promises).

So I have made some progress but also some setbacks on my audio player. I have all the parts I ordered arrive, and yesterday I did the first battery capacity test, which went poorly. The 14500 battery says it has a capacity of 2500 mAh and lasted 35min. I am for comparison using an old battery bank as a benchmark that is 10000 mAh, and I am at 3:30 runtime, and it has not indicated yet that it is at 75% (only has four indicator lights for charge remaining, all 4 are still on). I was doubtful of the capacity of the 14500 cell, since its capacity seemed to be over double what is normally the max capacity of a 14500, but I thought maybe there has been capacity improvements since whenever the articles I was reading were published. Turns out, probably not and instead it was false advertising so I will need to return them. I am going to look at some 18650 cells in person later this week to get a rough idea for physical size, and then maybe order a charge controller for it and switch to 18650 if it is not too bulky. According to my current capacity tests, I can get at least a 3000 mAh 18650 cell, which should get several hours of playback.

I am also considering switching from running the device with a Pi 3a+ to a Pi Zero 2W, since I do not need the slight performance boost of the 3a+, and a Zero 2W will give me space savings.

I have been working on some other slight improvements to the code as well while waiting for batteries to arrive. I have also been researching transistors, since that seems to be the only way to cut off power via software to my screen backlight. My screen is running on one of the 3.3v output GPIO pins, so I am considering using another GPIO pin and a transistor to handle turning off the screen back light. I have also done some code cleanup and documentation, which will help prepare for some changes and features I want to add in later.

It's a tiny thing, but I registered gotb.li for the TheatreBunch client project I've been working on - I'll use it for shortlinks as well as what's set up now, which is links to the project's website and various social media.

The project is to set up a community for theatres. The first client of NullusAnxietas. Whee!

this is really stupid --- but every day I watch Countdown off the channel4 site. I want to go direct to the latest episode, so why not right a bash script that fetches the URL and generates a little html redirect? Overkill.

#!/bin/bash
ep=$(curl -s https://www.channel4.com/programmes/countdown/ | grep -oP '/on-demand/[^"]+' | head -n1)
printf '<meta http-equiv="refresh" content="0; url=https://www.channel4.com/programmes/countdown%s">\n' "$ep" > index.html

run that crap every day after the episode is posted and I'm made in the shade.

I've just started reading through the river Wayland compositor's upcoming river-window-management protocol to start working on my own wm for it. I'm planning to write it in Zig, but that's about all I've got done so far :P I use river currently and I think the idea of an entire wm being able to be built into a Wayland protocol is a super cool idea so I can't wait to see how things turn out.

I have begun to believe that Khronos just likes designing spec objects and functions. I think they might also just enjoy making Vulkan users allocate objects just for the fun of it.

Edited in this paragraph because I realized I forgot an example: I wanted to just bind a storage buffer to basically 0,0 (set 0, bind 0). The juggling of descriptor sets and set pools and layout objects and all that noise to just say "bind buffer x with offset 0 to 0, 0" is crazy. Even though I have it working, it's so crazy that I think I'll ultimately abandon that nonsense and just use a push constant to push a device buffer address.

I'm slightly into a project where my real goal is just to learn some Vulkan stuff through building a silly little thing. The silly little thing I've decided to do is trying to push textureless voxel rendering as far as I reasonably can. The "fun" part will hopefully start once I get into meshing and such. One of my goals with meshing is to not ever have 2 vertices in the same location even though the triangles connected to a single location may have up to 3 different normals and up to 4 different colors. I think this is doable for as long as I carefully lay out nointerpolates.

I also have some really dumb additional ideas that are probably slower because of conversions even though they're amusing from the perspective of trying to minimize data size. For example, could I represent every face in a 256x256x256 cube of colors using 3 bytes for the vertex positions? The issue is that I'd need 257 positions on each axis, but maybe I could squeak by somehow using data from the normals. Helpfully there are only 6 distinct normals in a cube if I assume all the raw vertex data only composes unit-size axis aligned cubes. I think the smallest I could reasonably make each vertex is 5 bytes, and that's using a 256 color palette. Maybe 4 bytes if I drop to 32, but that feels kind of low.

I'm currenlty working on completing Land of Lisp. It's really fascinating so far, and has broken my expectations at almost every turn.

I'm also working on a first person multiplayer cooking video game in godot :D