Created on the Spot

I remember an epiphany I had as a kid downloading mp3s at a friend’s house who had DSL while I was still on dialup. You could download music faster than you could listen to it, and the infinity of it blew my mind. To say that AI is leaving my jaw on the floor on most days is an understatement, but maybe what I’m realizing today is it’s potential for creating faster than you can acquire. It’s not just that I’m having it write programs that replace a ton of apps I’m dragging around for only a thing or 2 they do well. It’s that at the time I need a program, it might be faster to make AI produce it from nothing than to otherwise acquire it.

The family’s been into Solitaire and Minesweeper, tons of sites, tons of ads, looking looking, bleh. Fuck it, I’ll just have AI make me one.

Click to pop out

Click to pop out

Inherently Programmable Pi

I’m teaching a small robotics class, we’re making the small Etch-a-Sketch plotter. Just a couple of kids, good eggs from the 5th grade X,Y coordinates class I teach every year. I’ve already done similar classes in various contexts, it’s a ton of work but very rewarding. One thing I always seek to improve is my Pi image management. Depending on the kids & material, we’ll go through some Linux CLI, or I won’t even want them to touch Linux but we’ll still have to click through some things to get going on Python to control GPIO pins.

What I’ve been wishing for for a while, is a Pi that’s always online everywhere, with the quickest way to get to just Python. Hopefully something web based so you don’t need anything other than your old faithful browser to start throwing code at GPIO pins.

Introducing IPP, the Inherently Programmable Pi! An image you can download, or build.

Step 1 – Python Forking Web Code Editor

And so step 1 is obvious, I want the Ace web based code editor, preset to Python settings, and served by the very Pi it’s meant to write code for. With a few buttons for running code, stopping it, and checking the output.

AI made that a breeze to code, it always blows my mind how well it understands even convoluted assignments. There’s some pretty gnarly stuff going on there. Python runs a web server that serves a coding environment that can fork another Python process, and kill it with varying degrees of prejudice when it won’t go away on its own. I want this packaged a a single Python file with all the bits and pieces bundled in. AI got all this with little help.

Step 2 – Zerotier & Public Bridge

For step 2, I want the Pi to be easily accessible online, no matter what crazy wifi it’s connected to. For this we’re using Zerotier, and an online bridge that is publicly available to forward traffic into it. In my case this is done with Traefik and a Docker container dedicated to forwarding into Zerotier. I unfortunately can’t offer this part of the stack with the image, but you can at least specify a Zerotier network ID as a tunable when building it.

This way, no matter what Wifi network your Pi is connected to, it will always be available via a public URL.

Step 3 – Local Wifi when Nothing Else

I wanted to take the script that turns PlottyBot into a hotspot if it’s not connected to the internet and modernize it for the newest Rapios (Trixie). I also want the web interface to let you connect to existing Wifi networks. This way you are either in a spot with no Wifi, and your Pi spawns a local one which doesn’t route to the internet but you can still program. Or there is pre-configured known internet Wifi within range and your Pi connects to that.

Of course AI yet again turned a multi-hour endeavor into a 30 minutes one. Well, it did mess up some fundamental things but still, I was spared hours of grinding.

Step 4 – Easy Imaging

Finally, I want to create Raspios images with all this and a few tunables baked in. No booting each one to install stuff & tune it for each kid. For this I’m using something I built years ago for mounting *.img files and tweaking them. It’s a Docker container you can pipe an image file into and out stderr comes the customized image. This might be worth its own post, I’ve used this for years to great effect, but that’s beyond the scope of this post. The point is it’s available to use, and here’s how:

Build Your Own Image

First download & decompress the latest Raspios:

wget https://downloads.raspberrypi.com/raspios_lite_arm64/images/raspios_lite_arm64-2025-10-02/2025-10-01-raspios-trixie-arm64-lite.img.xz
unxz 2025-10-01-raspios-trixie-arm64-lite.img.xz

Then download & decompress this project’s files to tweak it with:

wget https://ben.akrin.com/downloads/ipp_2025-11-11.zip
unzip ipp_2025-11-11.zip

Finally, create your customized Raspios image with:

sudo docker run --rm -i \
    --name=raspi-image-customizer \
    --cap-add SYS_ADMIN --privileged \
    --device /dev/loop0 \
    --platform linux/amd64 \
    --cpus="1" \
    --memory="500m" \
    -e HOSTNAME=mypi \
    -e HOTSPOT_WIFI_SSID=mypiwifi \
    -e HOTSPOT_WIFI_KEY=raspberry \
    -e ZEROTIER_NETWORK_ID="<ZEROTIER_NETWORK_ID>" \
    -e PIPASSWORD="raspberry" \
    -v ./ipp:/data registry.akrin.com/raspi-image-customizer:latest < 2025-10-01-raspios-trixie-arm64-lite.img 2> ipp.img

–cap-add SYS_ADMIN, –priviledged, and the loop0 device are necessary to mount a file as a disk inside the container.
HOSTNAME is evident
HOTSPOT_WIFI_SSID & HOTSPOT_WIFI_KEY are the settings for the local wifi network that the pi will spawn when it has no known networks to connect to.
ZEROTIER_NETWORK_ID is optional, if specified the pi will join it first chance it gets (when connecting to a managed wifi that routes to the internet).
PIPASSWORD is the password for the pi user, careful SSH is enabled in this image.

You can burn the resulting ipp.img to an SD card. The first boot will be longer than the subsequent ones as stuff gets initialized, but eventually you’ll see a new wifi network pop up. Connect to it, and point your browser to http://hostname.local or if that doesn’t work http://192.168.50.1, you will see the coding web interface. Now that might be enough for your to start coding, but if you want you can click on the wifi network icon at the top right to specify a standard network to connect to. And if you do so, I recommend you have a way to “follow” your Pi, either via Zerotier, or because you manage that network and are able to see what IPs devices get assigned. And if you setup a Zerotier forwarder on top of all this, your pi essentially becomes online to the world the minute it’s on. Now that isn’t just convenient, it also significantly lowers the bar to start coding with kids.

Or… Download a prebuilt one

Here’s the link.

parameters used for building it are:

HOSTNAME=ipp
HOTSPOT_WIFI_SSID=ippwifi
HOTSPOT_WIFI_KEY=raspberry
PIPASSWORD=”raspberry”

No Zerotier, but still everything else.

Plots

I’m plotting again after an August hiatus. The new Gondola Plotter is super quiet, so nice! I’ve made a few refinements to the software stack of course… The usual.

I found a bunch of really cool engravings in some very old books I found in the abandoned house. These books are worthy of some discussion on their own much like the old vinyls, but this isn’t the time.

I got ChatGPT to make me a few SVG handling tools that are really incredibly well done and will help me shed some of the apps I was using. With SVGs there really is not a single app that does all the things, so I drag a collection around just to use 1 or 2 of their function. AI is helping me write scripts to explore algorithms and replace more trivial functionality from apps.

Case in point: this super cool recursive inset fill algorithm to turn fills into something a plotter can do.

I’ve been wanting to write this one for a while, but I knew it’d take me a good few days to nail it. That got turned into 30 minutes with AI. Jaw dropping. I love the effect as it echoes a lot of bored fillings of shapes I’d do as a kid getting bored in class. The cookstove above is filled with it, but it’s far too dense at this size and so it really looks like a actual fill rather than a particular effect.

A lot of my tooling these days is Python which has good libraries for handling SVGs. I used to dislike Python for their dumb purist move from 2.7 to 3 which wasted everyone’s time, and because they had managed to reimplement library hell. But I have to say venv is a successful redress to the later.

Cool AI Use

Last time Esther wanted some coloring pages, I guess it’s time to battle with a bazillion sites trying to monetize everything. Jump through hoops here, web inspector there, the usual “this will take a minute” turns into another tech ordeal. Wait… I could just ask ChatGPT for some black and white stuff made for coloring.

The cool part? Esther gets a say into what she’d like to color. Well, she likes cats.

Then I realize that instead of asking for the generic stuff we’re used to seeing on coloring pages, we can ask for specific moments of her life.

Playing with an umbrella on a pile of woodchips? No problem:

Jumping on a trampoline full of Duplos? You got it!
And quite a few more:

Beyond the cool factor, I wonder if coloring her lived experience has a similar effect to journaling. Maybe, this helps process things?

More could be done to clean up the SVGs but this is quite sufficient for us. Evidently though, artists who were in the kid’s coloring niche have some competition :\.

One of the things that blows my mind with AI is all the things it can already do but have yet to be discovered.

All images above clickable for SVG, if you end up coloring one please send a pic 🙂

Stable Diffusion

I’ve played with Stable Diffusion a few months back, but the learning curve is steep, and easy access to DallE & MidJourney made it not worth it to press on. Until… I got excited about a super cool extension:

https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster

I just love the idea of having the same image have one analog meaning to a human and discrete meaning to a machine. And the same extension can be used for general masking other than just QR codes.

Without further ado, here’s the eye candy:

A benign QR code

General masking for an almost subliminal effect

There’s much to be refined in my incantation, but I really like how the mask isn’t just overlayed on top of a drawing in a subtle way, rather it’s the basis for growth of AI hallucinations.