A version of this post first appeared in my newsletter. Subscribe to receive posts like this in your inbox every other Sunday.

In a previous post, I shared one way I use LLMs: as a partner for writing. I’ve heard from several folks who thought that post was helpful, so in this post, I’ll share another use for LLMs: to augment reading notes.

Reading is my favorite hobby. During the day, I often read nonfiction for work (e.g., doing research for my book) and pleasure (e.g., personal growth.) At night, before turning in, I also read fiction (primarily novels.)

Around 15 years ago, I started a reading log. Whenever I finish a book, I write a note about it: a plain text file with high-level thoughts on the work. Later — sometimes years later — I revisit these notes to recall what I read.

At first, they were short notes: a line or two at most. Over the years, they’ve become more elaborate. (You can see some in my blog.) Recently, I’ve started using GPT-3.5 to help with the process. Here’s how I do it.

When I finish a book, I start a plain text note using a template to capture essential information: the book’s title, author, publisher, and year of publication. Then I write a summary followed by my observations.

For nonfiction, I write the summary myself. But for novels, I now let the LLM do it. The robot’s summary reminds me of the work’s plot and themes, making it easier for me to recall outstanding ideas.

For example, earlier this year, I finished War and Peace. It’s a long book; at my leisurely pace, it took several months. By the end, I was overwhelmed by details and characters.

I started the note with the usual template. Then, I asked the LLM for a summary. This is what it produced:

War and Peace is a novel by Leo Tolstoy, first published in 1869. It chronicles the lives of five aristocratic families—the Rostovs, Bolkonskys, Bezukhovs, Kuragins and Drubetskoys—during the French invasion of Russia. The novel explores themes of Russian culture and society, politics, family life, religion, and the meaning of history. It also features a large cast of characters in various relationships to one another and presents a detailed portrait of 19th-century Russian society at the time of Napoleon’s invasion.

This text went into the beginning of my note. (Whenever I do this, I precede the block with a tag that indicates that what follows was written by a robot, which keeps me from inadvertently passing it off as my own in the future.)

This snippet helped me recall the names of the families chronicled in the book and its themes. (“Russian culture and society, politics, family life, religion, and the meaning of history.”) The last theme, in particular, reminded me of a thought I had while reading the book: that historical developments (such as battles) are highly contingent; history assigns more agency to figures like Napoleon than they merit. So, I captured that idea in my note as well.

This is a good use for LLMs. For one thing, I’m not asking the system to read for me. There’d be little point in that. For another, this use case is less susceptible to hallucinations; I’d know if the system invented stuff since I know what the book is about. And even if the LLM spouted nonsense, I’d catch it. (I’ve already read the book!) Correcting the robot’s mistakes forces me to check the original or other sources, which is valuable per se.

So far, it’s worked well, but I’ve only started recently. I’m also reading classics like War and Peace; this approach likely works best for well-known works. The quality of the LLM’s summary depends on its training corpus, and there’s more written about classics than more obscure works.

Some details about the process: even though you could use ChatGPT to do this, I’m using the Text Generator plugin in Obsidian. This open-source third-party plugin allows me to call GPT-3.5 from my preferred note-taking app.

With the plugin installed, I type a line of text in a note (e.g., “The novel War and Peace by Leo Tolstoy is about”) and then press a button. After a few seconds, the summary appears in my note. (Similar features exist for other note-taking apps, such as Notion.)

Again, the point isn’t having the robot read for me. After all, I get pleasure from reading. And obviously, the summary doesn’t do justice to the entire work. But that doesn’t mean there’s no value here. It’s great to start with more than a blinking cursor on a blank screen. Not a draft but a memory nudge — an augmentation, not a replacement.