On Government: A Distant Reader Guided Reading

Yesterday I published a curated study carrel named On Government, and it was my first "guided reading." I'm still not 100% sure about the validity and utility of the result.

The reading is rooted in a study carrel (think "data set") composed of 480 books and journal articles on the topic of government. It totals about 14 million words. (The Bible is about .8 million words long, so this study carrel is about the size of 17.5 Bibles.) For more detail regarding the size and scope of the carrel, see the rudimentary bibliography as well as the computed summary page. The companion web page also includes word clouds (unigrams, bigrams, and keywords) that begin to illustrate the carrel's scope.

After creating the carrel I used an MCP (Model Context Protocol) server of my own design to query the carrel. More specifically, I used the server to address a number of things:

1. What is the Distant Reader, the Distant Reader Toolbox, and a Distant Reader study carrel? - Because the Reader is probably foreign to you.
2. Tell me about this particular carrel. - Because it is a good idea to browse through a book before you actually begin reading it.
3. Compose a speech on the topic of government as if written by a professor emeritus from a college of science. - Just for fun; a gentle way to get your head around the size and scope of the carrel.
4. What types of government exist, and what are their strengths and weaknesses? - Not a complete list, but are you able to enumerate even a few?
5. What is the relationship between the power of government and the people being governed? - Because the large-language model generated this question, and it sounds interesting to me.
6. Outline a set of study questions on the topic of government. - These are intended to get your brain going.
7. Tell me a few jokes about different government types walking into a bar. - I think they are funny; they made me laugh.

All of the responses are rooted in the content of the carrel, not from the ether of the general Internet.

If you spend fifteen minutes reviewing the contents of the carrel's home page as well as its children, then you will increase your vocabulary regarding the concept of government. Hopefully the process will spark your curiosity and make you want to learn more. Want to learn more? Download the study carrel. For more detail, see the carrel's readme file.

Guided Reading

This whole process is what I call a "guided reading." Collect data. Curate, index, and model the collection. Implement an application programmer interface (API) intended to query and report on the collection. Use the API to create any number of static reports describing the collection. Use the traditional reading process to use and understand the reports, and in turn, the collection. Implement an MCP server intended to exploit the API, which employs a large-language model (LLM) to convert natural language commands into API calls, interpret the results, and save the results as HTML. This is a "guided reading" in that the reader - someone like me or you - submits additional natural language commands depending on the previous results; the reader is directing the reading process. In the end, the reader will have investigated, learned about, and become familiar with the given collection.

This whole thing is different from the traditional reading process. But then again, it is not. Imagine picking up a book, leafing through the pages, glancing at the table of contents, browsing the back-of-the-book index, following an index entry to a specific page, reading the result, and repeating the process until done (or tired). In the end, the reader - you or me - will have become more familiar with the content of the book, and you may have even answered or at least addressed a curiosity. I assert my guided reading is not a whole lot different from the process just outlined.

Validity and Utility

But does a guided reading have validity and utility? I guess the answer is rooted in a number of things. Is the given collection well-curated? Is it accurately and thoroughly described? To what degree is it complete, comprehensive, and well-balanced? Were the processes used to do the collection's feature extraction more correct than incorrect? (They are never 100% correct.) Are the reports created by the API just as accurate? Are they meaningful? And then the biggest question: "To what degree can an LLM accurately interpret natural language commands, submit API calls, and interpret the result?" Regarding this big question, the jury is still out.

Based on my experience, guided readings can have validity and utility, but just like everything else, the results need to be curated; guided readings need to be thoroughly read, verified, and edited. If I didn't believe this to be true, then I would not have published On Government. That said, this process, especially the generative-AI aspect, can too easily be regarded as outputting truth. Instead it outputs more than plausible generalizations that need to be consumed with more than a grain of salt. Can you say "information literacy"? Some people will equate guided readings with "AI slop." What is the thing I heard the other day? "Trust but validate." Sounds good to me. Such is why I do not think the result is slop.

A Bigger Picture

Finally, the process of reading has existed for thousands of years. Our reading process evolved with the advent of Gutenberg's printing press around 1460. With the advent of the Internet, reading is evolving again. I suggest Distant Reader guided readings are one example of such an evolution.

What do you think?