The “cite-a-tron” (I need a better name, but “quotes device” is even worse and that’s what I’ve been using so far) is a small standalone artifact in a wooden case that uses AI to generate a random inspirational quote – or, rather, to generate language that echoes an inspirational quote, about once a day.
It’s in the spirit of “Slow AI”, and is very much intended to get people thinking about rhetoric and language, and on what we expect from quotes by famous people, and how much of ourselves is involved in the way we interpret language. The core of it is a simple but entirely embedded AI model to generate natural language within the device itself. There are no networking functions of any kind.
There is also an allusion to “Love Letters”, a resurrection of an early Manchester text generator from the 1950s, built originally by Chris Strachey. That was more or less the first artificial intelligence natural language generation application.
A few samples, for demonstration:
The device (today) consists of a Raspberry Pi Pico controlling a small, 4.2 inch, e-Paper display, with a piezo buzzer for sound effects, and possibly an LED for status display. The Pico was chosen primarily because it has a big enough memory space and enough processing oomph to run an interesting model without requiring too much power. I also consciously wanted the device to not be networkable, for safety reasons, and as a reaction against the large language models with expensive cloud costs. The Pico has proven a good fit, and is cheap enough to make the full setup very cost effective. The actual costs of building the AI model and the device itself are pennies – apart from the human time in curating and building. The processes of hand curation, as well as hand construction, are important parts of the work.
Here’s the basic Pico generating a quote and updating the display.
The AI part of the technology works as follows.
There are two separate language models, one for names, and one for inspirational quotes. I curated the two separately, then shuffled them into two text files, with newline characters between items, and then used them to train two separate CharRNN language models, implemented in Pytorch and running on an NVIDIA Jetson Nano. Both took some tuning to get acceptable values.
The architectures of both networks are more or less the same: an embedding layer, a set of recurrent layers, and an output layer. For names, two recurrent layers was sufficient; for the quotes, three works better. And for names, the sequence length was short (20 characters), where for quotes it worked better much longer, reflecting the long-range dependencies in the quote text.
I chose inspirational quotes because they have a very specific rhetorical style. They tend to use repetition (e.g., The most effective way to do it, is to do it”), and contrasts (e.g., “Falsehood is easy, truth so difficult”), especially between positives and negatives, as well as an assortment of common patterns like the rule of three (e.g., “Do what you can. Want what you have. Be who you are”). What I am trying to do here is explore the extent to which the style can be abstracted from the substance – and to look at the style more or less in isolation. It is not completely successful, but for my purposes, for now, it’s good enough. And the same hardware has a little capacity yet, especially if I can quantize the models, to improve and still nicely fit inside the Pico.
I have also mixed multiple sources, including very specific quotes on coding. For now, I’ve removed these, for several reasons: they’re not completely of a piece with inspirational quotes, and curation is hard. They’re longer, and a non-negligible proportion are misogynistic and problematic in other ways. While they often did generate intriguing combinations, for now, I’ve simply dropped them out of the collection. However, the principle of mixing languages and sources proved intriguing, and i will come back to it.
Anyway, having build two trained models, and chosen reasonably good checkpoints, I exported them as ONNX files, which make them easier to run on different platforms.
I finally used
onnx2c to translate them
into pure C code, which compiles in as
part of the Pico build process. Admittedly, the models use a lot of floating
point calculations which the Pico emulates, and it’s relatively slow, generating
1-2 character per second.
The original plan was to use model quantization, to turn the floating point numbers into integers, which are smaller, and would be very much faster on a Pico. But when I saw the live generation, I liked the rhythm of 1-2 characters per second, so I decided to keep it as is for now. This also inspired the use of a piezo component to ‘click’ per character, echoing a teletype.
So the main controller, when it starts, generates a quote using the first model, and a name using the second. If they’re too long, we simply generate another. I then use simple line breaks and format them for use on an e-Paper display, and when the layout is complete, generate a refresh for the display. e-Paper being e-Paper, we can then shut down all the power and sleep tight, until next time, with the display remaining.
The software isn’t yet done for the piezo, but I’ll be doing a small click more or less for each character, when it’s actually generating, sort of echoing an old fashioned teletype sound. That may change to a better PWM later, but the wiring won’t need to change for that. To avoid it being too annoying, there’ll be a jumper to turn off the sound. That’s very nearly as much configuration as I hope to allow.
For testing, I’m generating quotes every few minutes, but in the spirit of “Slow AI”, it’ll maybe be 6-8 hours in the long run.
Seeing that this is a Pico, which will simply be powered by battery, and has no network access whatsoever, configuration will be limited to a small set of jumpers. Probably as follows.
The initial version will be a small wooden frame, around 9 inches by 7 inches, and around 1.5 inches in depth, with the e-Paper in the middle. The electronics will be hand-soldered and wired on a prototyping board, and visible at the back. The e-Paper will plug into a connector on the board, so it’s easy to remove.
However, I will have to hand chisel some of the frame, since I don’t have a routing tool. My hope is to get a couple of these built, send them to people I trust to live with for a few months for feedback, and in the meantime, work on ways to make nicer boxes that are more based on local and sustainable materials. I especially want to keep the natural aspect. If I could make “woodland steampunk” into an aesthetic, that’s where I’d be heading. Rustic, but AI.
I’m planning to power it directly off three AA batteries, at least at first. My calculations suggest this might last a few months, because the Pico isn’t very very low power in sleep. A second iteration may add a real time clock to reduce power usage during sleep.
I’m also looking into:
These are character-level models, so they can and do literally create new words. This means that I have to complete a filter of sorts, to stop it re-generating words that will not be acceptable. For example, the language models can and do generate names like “Andrew Fucker” that riff off words that aren’t in the training data. At the moment, fortunately, that is fairly rare. For now, it’s only me that sees these, but I won’t be gifting any until I can be sure that the texts are going to be safe – and more importantly, not derogatory, even by accident.
There are gender issues too, with the names. On the plus side, the network has a delightful ability to create mixed names like “Jennifer John Ferriani”. On the negative side, first, the network does have a habit of generating real people, even ones not in the data set (e.g., “Tom Jones”). I’m not yet decided whether this is a problem or not. However, secondly, the training data for the names (originally from Wikiquote) has a significant gender bias. It generates names like those of famously quoted people, which includes a long historical dimension and is not equally distributed. I have taken great care to minimize the impact of this, but women’s names are less frequent than men’s, in line with the training data. But they’re no less frequent, which is actually an achievement with these techniques.
There may be other options for variants with different models inside. One that definitely appeals is a kind of “fictitious tree” version, which generates a tree name, a tree description, and a small image (which will need different technology, but I’ve played with generative techniques for this in the past, so I’m thinking “L systems” will be the main engine there.) For this one, a natural tree exterior is a must-have.