How to have a dialogue with a text
How to have a dialogue with a text
Chatbots open a Bakhtinian opportunity with large source materials
I suppose it’s been a part of my internet/search practice since early on with the web—Ctrl+F functions in a doc. Maybe I have a fifteen-page scholarly article on something and I’m trying to find a highlight, maybe I’m working with an extensive teacher-facing PDF that’s part of a suite of curricular materials, maybe I’m just trying to find something in the Quixote to refer to. Either way, search functions with large text documents have become commonplace in lots of folks practice.
Enter large language models and chatbots. With Chat-GPT’s splash into the mainstream a little less than a year ago, folks have been excited and anxious about implications, capabilities, reliabilities, vulnerabilities, veracities, etc. (I’ll pause on the word association right now). I couldn’t resist tinkering with the tools—I’ve always been a relatively early tech adopter, fumbling with HTML in 1995, breaking dada-poetry generators, and most recently taxing out every single podcast application.
Of course, the first things folks do are generative and pretty straightforward—and I wasn’t any different: Write me five paragraphs about dialogue in leadership, write a haiku in the style of Oscar Wilde based on Daoist philosophy, etc. There was a period this spring when my Twitter feed was full of this kind of generative stuff. This is where folks anxiously linger, thinking about students generating essays or generally “cheating” their coursework. Yet, pretty early on, I had some questions about this limited usage.
After a bit of tinkering, the idea of a ‘context window’ began to gain prominence in my use of the tools. You’ve probably seen the new little icon emergent in our lexicon—at least if you’ve used a chatbot. It’s that little broom with the action sweep lines.
This is the floppy disk/Save File icon that everyone will know in the future. This tells your chatbot to clear the context from your previous dialogue—it frees up the large language model architecture for the next set of utterances.
Why am I telling you all this? Well, I’ve been way too deep in the Bakhtinian over the last few years—evidence of which is all over this blog. And at some point in late spring 2023, I started seeing how much text I could feed chatbots. This led to some fantastic findings about the close-read capacity of chatbots, their ability to summarize and describe sets of texts, and finally, an emerging understanding of what I call having a dialogue with a text.
Context Windows
Anthropic is an org very similar to OpenAI in its creation of large language models and dialogic interfaces. Poe.com is my favorite way to access a variety of language models and chat interfaces, and I started using it because I wanted one free daily shot at Chat-GPT-4. Yet, I’m a Poe evangelical not because of that but because they added a bot called Claude-2-100k in late spring, which totally changed the game. The 100k names the size of the “context window” of the bot’s dialogue. Let me try and explain—
Were you and I to sit down and have a conversation right now, we’d keep what was said—at each utterance—in mind as we listen, respond, develop our own utterances, build meaning, etc. Perhaps the most recent chronologically will be more prominent in our thinking. Yet, we have the capacity as humans to refer to other bits of information associated with our conversational partner and the topic at hand. For example, my wife (unfortunately for her) knows that I talk about Don Quixote all the time, so when we’re in a new dialogue, that context she’s bringing. It might not be as clear and detailed as our convo five minutes ago, but this context comes along.
This is something chatbots just really don’t do as well as humans—understandably. As far as I can discern from a practitioner's point of view, they don’t have this sort of ‘unlimited’ context because it would require burdensome computing and become unwieldy pretty quickly. I believe I’ve experienced this, and many others have too—if you’re chatting with a bot and adding tons of text and going for fifteen-plus minutes without clearing context, you may find that the gumming up of the language model results in confused conflations, mix-ups, etc. And I mean something way more than hallucination.
What this looked like with Chat-GPT-4 was that when you tried to feed it a bunch of text, it would just say, “Nope, can’t do that. Too large.” So I’d have to break texts into sections, work with the bot to closely read those parts, etc. Yet, Claude-2-100k’s huge context window changed that. Suddenly, I could give it twenty pages of scholarly literature on paradox and ask it to describe and summarize the findings. I can’t explain to you the boon this has been for my research and scholarly understanding—amongst other things.
Anchor Sources
Recalling what I said above about the cursory, superficial, generative uses of chatbots to create content, I now only use chatbots with what I’m starting to call anchor sources. If you follow my Twitter, you’ve seen me do it all the time. Shortly after the 100k context window emerged, functionality was added to give bots attachments or links to retrieve information from the internet. It didn’t take me long to realize the implications of this. Wait—I can send it a crazy long Stanford Encyclopedia of Philosophy entry and ask it to describe and summarize it? Wait—I can send it a twenty-page piece by Adam Lefstein on teacher collaborative discourse and get something more descriptive than an abstract? Wait—
Folks, leaders, educators, learners, everyone—do you understand what this means? Since then, I’ve been on an evangelical mission to demonstrate to people how chatbots are inestimable close-reading tools that can help you deal with massive volumes of text in meaningful ways. As I mentioned, you’ve probably been Ctrl-F-ing things to find document sections. Now, you can send the whole doc to a large-context chatbot and have a conversation with the bot about the text.
Let me give you an example. The classics book club I’m a part of is now reading Oscar Wilde. I’ve become a bit obsessed with the provenance of Wilde’s aphorisms, or at least the trail of his French influences. So what have I been doing? Scouring Google Scholar for released essays on Wilde and his French sourcing and tributaries. Previously, I would find these, filing them as bookmarks, maybe hopefully (hopelessly) printing them, and wishing I could find time to read them as life swirls around me. Now, I can work with the bot to develop and understand the text from a text-message-like interface, take that knowledge, and move forward. Who wouldn’t want to be doing this all day long?
Zoom In/Out
Returning to the Ctrl-F analog, here’s something a search and find cannot do for you—allow you to zoom in and out on text sections. Let me help you conceptualize this: think about how you interact with images, maps, etc., on touchscreens. It’s not much different than that with text bots once you’ve got the anchor source in the model.
Below is an example of a few things I’ve mentioned here. I start with an anchor source from Stanford on “evidence.” I get a description and a summary. I’m dealing with this source because I’ve come across some interesting ideas regarding teacher/educator belief and discourse around “criterial/constitutive evidence.” I wanted to understand this terminology’s place in the larger philosophical landscape of knowing, evidence, epistemology, etc. So I got the summary and description, and instead of having to Ctrl+F “criterial” to no avail—I asked the bot if the source said anything about it. Even though it doesn’t use those terms, it gives me something really helpful in my search.
Finally
En fin, I asked a colleague of mine to help me understand the capabilities of Code Interpreter with ChatGPT—as I mentioned above, I haven’t been a subscriber due to my affinity for Poe and Anthropic’s products. But seeing some demos of Code Interpreter helped me understand or conceptualize something about this idea of a dialogue with a text: When you use Code Interpreter, you’re giving it a .csv or data repository, asking it to understand what’s inside, then asking it to reorganize, study, analyze, manipulate whatever is there. This is in no way substantially different than working with an anchor source and a chatbot.
Let me explain—you know how you can sort columns, filter cells, and do all this other amazing data/digit manipulation within sheets? You’re doing the same thing once you introduce an anchor source as a third point of reference in a dialogue with a chatbot. You’re asking it to bring its large language model training to bear on a particular text organization—a set of graphemes, words, etc. Thinking about it in this way starts to reveal the capabilities of the tool—"take all the stuff about constitutive evidence out of this source,” “only report back on Wilde’s French influences,” etc.
Before, I was accumulating outrageous amounts of bookmarks, favorites, PDFs in the downloads folder, and archives in a Google Drive that I just couldn’t deal with at volume. Now, when I see relevant research from RAND emerge in the form of a forty-page PDF, I can ask my close-read partner to describe and summarize it, zoom in on parts that I find helpful, and check the source text for accuracy. Honestly, I dunno why everyone isn’t doing this right now. But I’m here to help you learn why and how you should!