“Maimonides entering a query into ChatGPT on a laptop” (Image created using ChatGPT/DALL-E)
by Ian Jensen, Department of English
Introduction
The title of this guide alludes to Maimonides’ influential development of apophatic/negative Jewish theology in the 11th century. While for Maimonides God can only be described in negative terms, here I’m concerned to develop a working account of the kinds of “AI” we teachers of writing are likely to face from now on. As I hope to demonstrate below, however, what large language models (LLMs) aren’t is equally as important as what they are.
I’ll begin by explaining how LLMs work without a lot of jargon or computer science. I’ll then discuss what they are (text generating machines), what they aren’t (intelligent in any meaningful way), and what all this means for us.
Finally, two things to keep in mind. First, I am no expert here, so this is one person’s pieced together account. As a result its accuracy may not be 100%. Second, this will be an opinionated guide. Caveat lector!
What are Large Language Models?
It’s important to remember that generative artificial intelligence (AI) is a genus, and that large language models (LLMs) are a species of AI. Here I’ll focus on the biggest issue for writing teachers, large language models.
Elastic notes that “ A large language model (LLM) is a deep learning algorithm that can perform a variety of natural language processing (NLP) tasks. Large language models use transformer models and are trained using massive datasets — hence, large. This enables them to recognize, translate, predict, or generate text or other content.” It is important to understand that LLMs are essentially algorithms, not so different from the predictive text typing we find on our phone messaging apps or the music predictions produced by Spotify etc.
Ok, so let’s define algorithms. The International Institute in Geneva says that “[A]n algorithm is a set of instructions, used to solve problems or perform tasks, based on the understanding of available alternatives.” Algorithms process data and apply rule-based frameworks (e.g. English grammar and syntax) to that data. From these, the algorithm is able to “predict” what is likely to follow a given relevant input. So e.g. when you “like” a post on Instagram, your interest is fed into the algorithm which will then populate your feed with posts it deems to be similar. The same goes for advertisements we see in Google, Facebook, and so forth.
Generative AI LLMs take these “predictive” algorithms and trains them on more or less the entire free internet while exponentially expanding the parameters that the algorithms are able to take into account. Rather than “predicting” a given word based only on the word before it, which is what many of our messaging apps seem to do, LLMs are able to widen the parameters of how much data is taken into account. So instead of using a single word, an LLM might be able to use a full sentence, a paragraph, or even an entire essay or book “in mind” as it “predicts” what is likely to come next. I’m using the example of text here but AI algorithms can be trained on images, sounds, biological data, and so forth. Google’s AI division has already developed a multimodal generative AI they are calling, unsettlingly, “Gemini.”
Now back to deep learning, which simply means that rather than having only one data processing “layer” there are multiple layers that increasingly refine the data processing that is used as the basis of “predictive” generation. These layers are called neural networks. The point is that the “predictive” algorithms of large language models have processed inconceivably large amounts of data (i.e. text) in a very sophisticated way called “natural language processing.” From that data and the parameters both embedded into them by humans and self-produced from data analysis, the LLM can then produce astonishingly coherent text.
OpenAI’s GPT-4, the one that powers paid ChatGPT and Bing Chat Microsoft Copilot, famously used a method called “crawl-and-scrape” to train. Essentially, OpenAI built script-running “bots” that index and scrape data from publically available websites (and other resources). If this sounds ethically questionable, that’s because it is. Indeed copyright lawsuits have been and are being filed, some by big name authors & publishers, as well as the New York Times: their complaint is well-worth a read. In addition, publicly accessible large repositories of human-produced text like Wikipedia and Reddit are beginning to close ranks against LLM-training crawl-and-scrape bots, in one case causing Reddit to alienate a number of contributors (redditors) and communities over the summer.
NB: Anything you have written that has been posted to a publicly accessible website has likely been used to train at least one LLM. This includes, but is not limited to, open access papers, your CV, syllabi, profiles at Academia.edu/ResearchGate/Google Scholar, personal blogs, Substacks, lectures posted on YouTube, comments you’ve posted, and so much more. Welcome to the brave new world.
Are Large Language Models Actually Intelligent?
The short answer is no, they are not. The reason for this leads me to the longer answer, which is absolutely crucial for both teachers and students to understand: Large language models, and indeed all generative AI so far available to the public, does not in any sense understand what it processes or produces.
In fact, in LLMs the text generated is produced through something called “tokenization.” In tokenization, the letters and words produced (and processed) by the LLM are assigned non-semantic data values, with non-semantic here meaning having reference to meaning. From there, the LLM produces the next letter or word token, which is based on the prior tokens in the model.
What this means for teachers is that generative AI produces pure text devoid of understanding, intent, or semantic content. This empty text based on probability is nothing other than a product of tokens assigned lexical skins, as it were, being strung together. As a result, the text produced is not in any usual sense a message nor does it have content. Any meaning that the text appears to have comes entirely from the reader. There is no ghost in the machine.
When initially made available, ChatGPT was not connected to the internet and relied on training data from its initial “crawl and scrape” of the internet. Increasingly, LLM chatbots are being connected to the internet—ChatGPT’s paid service, Google’s Bard (i.e. Gemini), and Microsoft’s CoPilot (formerly Bing Chat & powered by OpenAI’s GPT-4) are all able to integrate real time data from the internet. This is bad for instructors because it mitigates some of the errors seen in early LLM output, which in turn makes IDing LLM text more difficult.
All of this is important to remember when pundits and AI CEOs tell us about all of the creative work LLMs will help us do, and helps us understand why LLMs “hallucinate” false references and factual errors. In the first case, and keeping in mind that generative AI has been and will be useful in many capacities, it is simply the case that in today’s LLM-generated text is produced out of thin air as it were–a thousand neural network monkeys on typewriters so to speak. Certainly this doesn’t foreclose the possibility of inspiration striking a reader of LLM generated text, and perhaps more generously we could think of LLM text as something closer to Burroughs’ cut-up experiments in such a use case, but that inspiration is not the product of any kind of linguistic function Jakobson would recognize–there is no sender and there is no message; there is no line of communication to open, etc.
In the second case, LLM hallucinations occur precisely because of predictive tokenization and the absence of knowledge or understanding. So because Stephen Greenblatt and Julia Lupton have written a great deal about Shakespeare, if we ask an LLM to add citations to our essay about The Tempest it is completely possible that it will do so by simply generating a tokenized name like “Greenblatt” or “Lupton” and a page number that fits the pattern it trained on: so never a 1, 2, or 3; never exceeding 600; falling between say 200-500, etc. It may also very well generate the title of a book or paper that fits its recognized pattern for something like “titles associated with Shakespeare and including an author reference to Lupton.” You could well get something like [“Affects and Affordances in Hamlet” (2022)] as a reference, despite no such article existing.
For more on what LLMs can do and how we might think about them in the college classroom, look out for part 2 of this guide on The CraFT next week!