Generative Pen-trained Transformer
A few years ago, back when GPT-3.5 was first released, and I still spent most of my days sitting in a high school classroom, I had a project idea. Perhaps one of my best named project ideas to date: The Generative Pen-trained Transformer or GPenT for short.
GPenT was awesome, I was going to build out a color-changing pen plotter with some funky coreXY-esque kinematic, and alongside it, try my hand at writing one of those SVG Blob Generators to generate a bunch of shapes to draw. These ‘blob’ generators are simple enough, they take in some input variables from a user in the form of sliders and output a downloadable SVG, yet I wanted to try my hand at something new. Instead of a UX with sliders, I would pass in a raw string of numbers into this generator, numbers generated by GPT-3.5.
It’s funny to look back at my intuition here, to use an LLM to generate a string of numbers, to generate SVG blobs, to feed to a pen plotter, and to call the whole thing the Generative Pen-trained Transformer? Being the hardware engineer I always have been, I didn’t have any ML, or honestly, even much Python experience to fall back on. And for my first ever flirt with artificial intelligence, this felt like a pretty solid project. So, I got at it, starting to chip away at the kinematic and machine design, before quickly getting blindsided by the chess cheating scandal of the century, and with it, a newer, shinier project. As with most of my wonderfully named project ideas (I’m slightly biased here), GPenT was back on the shelf.
Failed first attempt at GPenT Below are the half-completed remains of GPenT v1. Rest in pieces.
A lot has changed since my high school days, for starters, we have GPT-5.2 now! and with it a whole bunch of gpt induced psychosis. Also, I have my own place in San Francisco! If you’ve seen any of my other recent projects, you know I’ve been on a kick to trick this place out, and while I’ve been making progress on the silly, whimsical projects front, the apartment walls have remained shockingly bare. So I figure it’s time to dust off the old Generative Pen-trained Transformer, see this, beautifully-named, project through, and with it finally get some awesome art for my apartment.
What is a polargraph? And why is it wall-mounted? Well, a polargraph is a vertically-mounted drawing machine, and I figure what better than art on your wall if not an epic machine that makes art on your wall!
The operating principle of the polargraph is simple enough. Take some rectangular work area with a motor in each of the top corners, string a belt over each of those motors, append a counterweight on one side of each belt, and attach the other side of both belts to a single gondola. This gondola will thus be moved when either motor rotates.
Plus, I already had a Fusion 360 folder for this project, so starting on the kinematic and machine design couldn’t have been easier. There are a bunch of easy DIY polargraphs out there on the internet that I referenced throughout this design process, notably the Makelangelo, whose software and firmware we’ll reference later in this piece.
Given I intend to hang this polargraph above the couch in my living room, I’m taking pains to make sure it looks pretty.
... continue reading