Toward a Luddite Approach to LLMs

When it comes to "engines that fabricate articles in a fraudulent and deceitful manner", I agree with the luddites. I am concerned that generative tools powered by large language models (LLMs) -- such as OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, etc. -- often nudge people toward fabricating products in a fraudulent and deceitful manner.

So, when I published a new website that was partly "built" by LLM coding tools such as Claude Code, I felt the need to share some of my reasoning.

To get a little meta, let me start by saying that drafting (and/or editing) my writing is one of the tasks for which I refuse to use LLMs. I believe writing is thinking -- that is, (I think) one of the best ways to understand what is happening in my brain is to translate my thoughts into language (and/or visualizations). Within this framework, every word outsourced to a generative model is a missed opportunity to strengthen my own thinking.

The analogy of "bringing a forklift to the weight room" has been an effective way to communicate this principle to students. If you don't want to strengthen your thinking, then by all means, go ahead and have LLMs do your writing. And I have to ask, if you don't want to strengthen your thinking, what made you decide to enroll in school?

The same goes for my own work - everything I ask an LLM to do is something I will not do myself. I will not learn from any "writing" produced by an LLM - it's like bringing a forklift to the weight room.

There does seem to be a difference between outsourcing a thought process and outsourcing a building process. In the case of my website, for example, I have gone back and forth between managing it on my own versus outsourcing parts of it (i.e. to Ghost, Wordpress, etc.). Writing html and css is time-consuming. But when I let Ghost handle all the html/css, I lose some control. It also costs more money, and creates a dependency on their hosting services.

For my website, I wanted the best of multiple worlds:

• Quick editing
• File hosting
• Customizable/controllable html and css
• Fast and lightweight
• Tinkerable! Especially for privacy and security improvements

It's only a little personal website, but ultimately, all technologies are political. So I sketched out my idea for a new website - I actually drew it on paper! Then, I fired up Claude code and iteratively prompted it to generate the code that would bring my sketch to life. I checked all the outputted html, performed some tests, and kept track of my prompts. For the most part, it felt like working one layer of abstraction above existing high-level programming languages -- for example, I still used html vocabulary (e.g. div, span) to guide the output.

Why all the fuss? Who cares if a website is vibe-coded? Well I do, and if you are reading this, you probably do as well. We are the working-class nerds! We know it's not just about code and control, it's about freedom! And as Eliza Ludd wrote, "curs’d his the man that even lifts a straw against the sacred cause of Liberty." Dare I say, we are the Neo-luddites?

As you may have heard, Claude code is pretty capable. It made my website renovations pretty quick, and eliminated a lot of time that I would have spent searching online for different html and css snippets. Since it is my website, I do feel responsible for the code that is published -- understanding it well enough to maintain it and adjust it without an LLM.

Here is a summary of my current approach to using and not using LLMs, as of August 2026.

I avoid using LLMs for...

• Drafting or editing writing (again, writing is thinking)
• Summarizing sources I have not read
• Creating slides/teaching materials from scratch
  • e.g. "write ten slides for virtue ethics"
• Fabricating something that I could not meaningfully explain or modify myself

I sometimes use LLMs for...

• Drafting code, particularly code that will do mechanical / verifiable / deterministic chores (e.g. reformatting, renaming, converting one file format to another)
  • e.g. "write a python script that will generate an svg file that resembles example.png"
  • this way I can tinker with the script and/or the svg :-)
• Creating Html and css scaffolding (again, I see this as working one layer of abstraction above the code I would previously write by hand)
  • e.g. "add a quarto slide template in week0 that I can use for two-column slides"
• Searching for a source I can't find, or finding a quote from a text (both of which I of course verify)
• Checking for typos / inconsistencies
• Checking for security
• Converting or translating materials between formats/languages
  • e.g. "write this python code in R"
  • e.g. "turn this list into a markdown table"
  • e.g. "create a markdown file from the main text in this PDF"

There is admittedly some gray area here, and I might need another list, or at least some longer explanations. Namely, while I try to only "ship" things after I have read them, understood them, verified them, etc., I can't really guarantee that. In a perfect world I could maintain everything without an LLM, but that is difficult to verify, unless I "quit cold turkey" and stopped using LLMs... forever?

Another approach/safeguard I have adopted is to keep detailed version histories for basically everything I work on: class materials, papers, etc. This is intended to keep the slow, incremental, human work process visible, whereas fully LLM-generated content is more likely to be fabricated in bulk, out of thin air.

Another hesitation I have at this point is Anthropic itself. They're running on billions of dollars from Amazon and ruthlessly stealing intellectual property wherever they can. This is not to mention their partnerships with Palantir and the U.S. military. And don't forget about all the unnecessary resource-hungry data centers! There are similar issues with basically any major tech company these days (Meta, OpenAI, Microsoft, Google/Alphabet, etc.)

So, as of writing, I am exploring ollama launch in order to hopefully "get off" Anthropic and the other big tech companies. The struggle continues.

Thanks for reading, fellow nerd!