Prompt: cthulu operating a computer, detailed and intricate, advertising, realistic, beautiful lighting, by Malika Favre; made with DiffusionBee
The DHCU Irregular has been a bit… more irregular than I would’ve liked, so far this year. Sorry about that. Normal service resuming soon.
Hey, what’s all this ChatGPT business I’ve been hearing about? ChatGPT, please write the introduction to this issue of the Irregular, but give it an eldritch slant.
Welcome to the first issue of the Digital Humanities Irregular at Carleton University! In this issue, we will be exploring the exciting and rapidly-evolving field of digital humanities, which uses digital technologies to enhance and transform research and teaching in the humanities. We will be featuring articles from faculty, students, and staff on a wide range of topics, including digital archives, text mining, data visualization, and online pedagogy. We hope you will join us in this journey as we delve into the intersections of technology and the humanities.
In this issue, we’ll use chatgpt3 to write a variety of texts, including articles, poetry, and fiction. We’ll also explore the capabilities of chatgpt3 as a language model, including its ability to generate human-like text, complete tasks based on natural language input, and engage in conversation with users. We hope you’ll join us as we discover the exciting potential of chatgpt3 and the role it could play in the future of digital humanities.
As we embark on this journey we can’t help but feel a sense of unease, as if some ancient and malevolent force is watching our every move. But we press on, determined to uncover the secrets that lie hidden in the circuitry and code of our devices. Will chatgpt3 be able to navigate the complex and sometimes unsettling world of digital humanities? Or will it be driven mad by the data and algorithms it uncovers? Only time will tell…
ChatGPT 3 and similar ‘large language models’ are extraordinarily high-dimensional models of the English language, with billions of parameters in that space. When we give such a model a prompt, it collapses the probability space such that we can end up with a digital humanities / Lovecraft mashup. It’s not great prose, but it’s not too bad, either; certainly at a par with, umm, newsletters put out by university departments and so on (Maggie Appleton calls this ‘the expanding dark forest’ and wonders what happens when the web becomes polluted with such output).
From a different direction, there’s been a lot of moral panic about students using such models to write essays. This I think undervalues students’ own agency. They know the game: if the class is about hoop-jumping, rote responses, then what do you expect will happen? If the work is meaningful, then students will do it, and will take pride in it. I had a GPT3 exercise in my own class ‘HIST3812: Playful Engagement with History’, just before ChatGPT3 emerged on the scene. The point of the class was to rethink how we tell history through digital media. I asked students to build something with GPT3, and to think through and reflect on that process. Process over product, that’s my mantra. Interestingly, the students found GPT3 too easy, instead preferring the friction and less polished results of working with GPT2 retrained on a corpus of historical writing. This is an approach that Carleton student Chantal Brousseau explores in a tutorial for The Programming Historian. Brousseau writes,
In a way, the output of GPT is various interpretations of a singular history being performed, and you are the critic meant to analyze and elucidate each performance. Are these performances completely accurate to their source material? No, but in their “inaccuracy” we can find new perspectives and expositions that differ from our own, which ultimately expands our understanding of the subject at hand.
Instead of forbidding these models to students, maybe we could work through with our students what these models teach us about being human? I’ll leave you to consider Alan Liu’s ‘ Theses on Large Language Models and ‘Good’ Writing’. He suggests that what we really need to be thinking about is what new ways of being creative/human will look like, as these models get ever bigger. How do we teach that? How do we scaffold that? Now that’s what I call digital humanities.
One last prompt to ChatGPT: explain large language models, in the style of a 1940s gangster film:
“Hey boss, what’s this ‘large language model’ you’re always talkin’ about?”
“It’s a new kind of artificial intelligence, kid. It’s trained on a lotta text data, and it can generate human-like language like it’s nothin’.”
“No kidding? What do you use it for?”
“Lots of things. Machine translation, language generation, question answering. It’s a game-changer, let me tell ya. Could revolutionize the way we do business.”
“I’ll say. Imagine never havin’ to deal with those dang translators again. We could expand our operations all over the globe!”
“That’s the idea, kid. The sky’s the limit with this technology. It’s gonna be big, real big.”
I recently came across elicit.org. This service uses GPT3 to handle the natural language processing, coupled with a search engine tied into the academic publishing ecosystem. Ask a plain-text question, “Why might bumblebees be more socially advanced than we expect?” and it will return results, summarizing their contents and how they interrelate to one another. It’s quite impressive, if not always spot on.
“These papers provide some insights into why bumblebees might be more socially advanced than we expect. Sadd 2015 found that bumblebees have many of the same genes for social organization as honeybees, indicating that the route to advanced eusociality in bees was mediated by many small changes in many genes and processes, while Libbrecht 2015 found that…”
The summary comes with the full citation, number of times a paper was cited, and a link to the DOI for each article. There’s also a table summarizing each paper ready for export as CSV. Would I use this for serious research? I might certainly use it to find a way into a literature that I don’t already know. For now, it’s free, but don’t count on that lasting for very long. I imagine they’re burning through their cash reserves very quickly, since each search is making a call to GPT3, for which someone has to pay.
Fair play to Elicit, they provide a lot of information about how the service actually works, how they prompt GPT3 (and which submodels), AND - which I found very impressive - how they mitigate GPT3’s willingness to bullshit when it doesn’t know the answer. Their search also surfaces papers that criticize the main results, too. You can read more about it here but I think you might very well find an interesting use case for Elicit in your own teaching or research.
You’ll notice that at the top of this newsletter I included an image that I generated using DiffusionBee. This is an interface for using the Stable Diffusion model, on your own computer (provided your machine has enough juice). You’ll notice that I included in the prompt, ‘by Malika Favre’. I like Favre’s art. Wikipedia characterizes it as, ‘where pop art meets op art’. The piece I generated looks nothing like Favre’s work. It seems to me that when I use that phrase, it steers the image generator towards flattened areas of colour. Here’s another image generated with the exact same prompt, but from a different random number as seed:
I reverse image search these too, to see if the machine is generating something new, or if it is merely copying something extant. So far, so good, there doesn’t seem to be anyone out there painting/drawing in this ‘style’. But that’s not the case for all painters. Just ask Greg Rutowski.
The best explanation of how these image generators work is maybe this piece by Eryk Salvaggio. It’s essentially a kind of archaeological process - the model is trained to understand what a fully decomposed image (with associated written description) looks like when rendered as noise. Then, when you give it a prompt, it triangulates between the patterns of noise it ‘knows’ for those words, and then recomposes an image from that particular pattern of noise. Thus, for popular painters like Rutowski, it knows a lot about how that painter paints… to the point where the work of the actual painter is lost in the, erm, noise of the internet and people posting their creations.
The folks at Nomic (a digital humanities start up!) have a map of the text labels for images used for one popular model (check it out here); we could profitably spend quite a lot of time working out how the model sees the world, from that atlas. Salvaggio proposes a different way to ‘read’ the output of these models, in a kind of aggregate approach - see this post - and I think there’d be a very interesting Master’s project in applying that method, pushing it further, for someone.
AI and these large language models are a challenge that I think Digital Humanities is well suited to meet… if things would just stat still for a gosh-darned moment. For my upcoming sabbatical I proposed exploring a variety of AI approaches for an archaeology project. I wrote my application sometime in September, I think. My project was already obsolete by the time I it was approved back in November. Things move fast in this space. Can DH keep up? If anyone can, it’s gotta be us, right? Otherwise, AI in the form of these ever-expanding models is going to give us all a very rough ride.
Remember folks, if you’re interested in a dh practicum or dh directed study (digh5011, digh5012) in the spring, you must clear such a thing with me first.
If there is any interest in a workshop on what a DH Thesis/MRE might look like, I’m thinking of holding such a beast in early February. Email me to indicate your interest.
The CUDHGSS are hard at work planning their second annual graduate student run virtual DH conference for this spring. If you’re interested in supporting the work of the Society, tax-deductible donations can be made at their FutureFunder.ca page.
I will be going on Sabbatical on July 1. In which case, there’s an opportunity for a faculty member already at Carleton to take over the reins for a while:
The Dean of FASS is accepting applications for an Interim Coordinator of the Collaborative Specialization (Master’s) in Digital Humanities (www.carleton.ca/dighum).
The Interim Digital Humanities Coordinator would serve a one-year term beginning July 1 2023 and receive an annual 0.5-credit course release. The Coordinator is responsible for all aspects of the Collaborative Specialization, including:
Please send questions or expressions of interest by February 1st 2023, to: firstname.lastname@example.org
cthulu wearing snowshoes in a snowy open field, detailed and intricate, advertising, realistic, beautiful lighting, by Malika Favre, made with DiffusionBee