LLM-Lensed Iterative Blogging

NOTE: All AI-generated prose on this site is clearly marked as such.

The Problem

I've wanted to write a blog for a long time. Everyone can use practice writing and organizing their thoughts and I've had many small thoughts about math, tech, and life that I wish I had written down.

The primary obstructions are obvious: busyness and laziness. There are too many fun things to do in Philly and DC and too many cool people to do them with!

The secondary obstructions are less obvious. For bullet-list items, I've maintained an iPhone note for years, but without sharing these, my writing doesn't improve. For full blogs/essays, I can't share the idea without several hours of dedicated focus. Can we thread the needle? Is there a way to document and share an idea clearly before having followed each point to its natural conclusion? No. At least not as well as if you had fully edited the work; the writing itself is valuable. But what if we don't let the perfect be the enemy of the good?

There are existing solutions. Twitter allows users to share half-baked but still valuable thoughts. Even listservs largely handle the kind of communication I'm envisioning. But these are push methods; is there a pull method? A fine solution is just posting summaries of your ideas.

But thinking through posting just a summary of an idea illuminates the secondary, perhaps silly, obstructions to my having blogged:

• I want readers to have all of the context in front of them without my having to write the very familiar details,
• I want a certain aesthetic for the content, including length, and
• I feel acute dread looking at an artifact that communicates not only the intended content, but also my own lack of time.

A Solution

This problem is likely very old, but I'm optimistic I've found something that works for me: "LLM-Lensed Iterative Blogging". The process is shallowly identical to a flavor of iterative blogging; I'll write a summary of the blog, expand it a little, expand a little more, and continue until I have a blog I'm proud of. The difference is that readers will see not only the partially written post, but will also see a progressively clearer view of the entire post, including relevant background inline from the start.

We do this with good ol' LLMs. I write the growing summaries, and AI writes its best guess at what a final version would look like.

I don't want to read AI-generated content. You probably don't either. It's unethical to have Claude or Jippity write a blog and claim it's yours. This is for many reasons, ranging from deceiving the readers to intellectual property. But, for now, I feel comfortable hosting AI-generated content if it's clearly marked and attributed to a model.

Concretely, an LLM-Lensed Iterative Blog will contain a demarcated summary at the top with the content I have written by hand. After this is the body of the blog, as written by AI with the exact summary as the meat of the prompt. This is labeled "Generated by <model>" and the reader can decide if it looks interesting enough to read.

I secretly hope (and expect) that most readers will view only the hand-written summary. The entire post most functionally serves as a reference for readers unfamiliar with the domain, which will be mostly technical math and logic written for a general audience. But for me, the existence of the generated post is a fun gimmick that allows my head to hit the pillow at night thinking that I haven't released something offensively incomplete into the world.

The New Problem

The process of writing

• organizes your thoughts and
• manifests original self-expression.

Both of these goals are threatened by LLM-Lensed Iterative Blogging. I'll need to edit the AI-generated post to ensure its correctness. This in particular means that I'll need to read this content. Reading this will certainly change how I think about my own idea. I'm likely to write it more like an LLM would and less like I would have in a vacuum.

I read a lot of AI-generated summaries for work and use Codex and Claude Code extensively for both fun and profit. A friend recently told me some of my SMS prose sounded AI-generated. I don't know how to avoid this while still keeping up with the expected pace of modern work.

My best makeshift solution for now is a scientific one. I will 'cleanroom' some blogs, writing by hand without ever reading an AI extension. If my friends, colleagues, or I see a notable difference in character between the 'cleanroom' blogs and the LLM-Lensed Iterative Blogs, then we'll revisit the drawing board.

For now, I'm glad to finally be writing.