Jack's Rules for LLM Use

"Ultimately, the key question remains the one posed by Saint John Paul II: does AI “make human life on earth ‘more human’ in every aspect of that life? Does it make it more worthy of man?” If the answer is yes, then we can recognize it as an opportunity to be embraced responsibly, on a path of patient, shared reconstruction, akin to the rebuilding of Jerusalem narrated in the Book of Nehemiah. If, however, power grows while the heart withers and human bonds fray, then we are faced with a new form of Babel — a construction that is grandiose, yet fundamentally dehumanizing." — Pope Leo XIV, Magnifica Humanitas

I don't fear LLMs becoming some sort of god or world-ending paperclip monster. As Corey Doctorow said, expecting a text generator to become a sentience is as absurd as expecting a horse to become a train. You can't feed a horse more carrots and expect its fundamental mechanics to change.

Instead, I fear LLMs flattening our relationships with information. Rather than engaging critically with sources, problems, opportunities for intellectual growth, we can just shovel those things into an LLM and call the output a solution.

These tools reduce our critical thinking skills much as elevators reduce our ability to take the stairs. Riding the elevator might save some time, but take every elevator, every time, and all of a sudden a three-story walkup becomes daunting. Of course, in the perfect analogy, this elevator would also occasionally deliver you to the backrooms while insisting you're in the penthouse.

There is little value in the use of LLMs for most tasks. Human-oriented generated content is usually slop and always tainted by lack of effort. One of the greatest byproducts of the "AI Era" is the resurging value of authenticity by way of imperfection: it is human to make a mistake, to have an unpolished recording. The greatest jazz records crackle a bit, accumulating distortion from the instrument to the microphone to the tape. Using an LLM for human-intended action is taking a shortcut that doesn't end at quite the right destination, but at an uncanny spot around the corner.

The rare exception, though, is LLM use for software development. There, authenticity is achievable because the primary intended audience is the compiler, not the human. In code, there are clear and irrefutable definitions of truth and quality. The program runs, or it doesn't. When an LLM spews out lines of logic, that output can be automatically and objectively evaluated in ways that a PowerPoint, a blog post, a summary could never be.

While an LLM can generate solid code, that doesn't mean you should let it do all the work. A developer with access to high-end models can churn out hundreds of thousands of lines a day without learning anything.

Using an LLM for a task stagnates your ability to accomplish the same task manually. Anything that doesn't grow, decays. As someone interested in my personal growth as a software architect and invested in the continued improvement of my team, here are my rules for using LLMs as a software developer wanting to push back against stagnation.

Use LLMs For:

• Spinning up code to implement human-described, thoroughly understood solutions that address specific, finite problems
• Implementing bug fixes that you understand with your human brain
• Generating tests for existing software

Generally, moving faster than you can to extrude code you understand.

Don’t Use LLMs For:

• Making notes or thinking – use paper. It’s better for your brain.
• “Brainstorming” – do not mistake generated content for actual creativity
• Writing – anything worth writing out is worth writing out yourself.
  • This includes using an LLM to "improve" your writing, "clean up" grammar, etc. – buy Strunk & White. Write a journal. Write an essay. Write a letter.
• Any act of production intended to communicate concepts from your human brain to another human brain. Documents, presentations, emails.
• Summarizing requirements for a complex system or problem. You should be able to understand and articulate the task at hand without an LLM.
• Solving hard bugs on its own (“Fix this error” doesn’t teach you anything.)

If you need a text-generation machine to communicate effectively or to understand a problem, you’re doing something wrong. Go touch grass, read a book and call a friend.

To Preserve and Promote Your Own Thinking:

• Install a human-only day in your weekly schedule. No LLM, all you. Your quantifiable output might go down, but you'll be better for it.
• Don't address the same kind of problem with an LLM every time. A .jsx page doesn't properly compile after a change? Go hunt down that missing curly bracket yourself.
• Don't center your workflow around the LLM. Is Claude pinned to your taskbar? Launching on boot? The companies who run these models have a simple end goal: establishing themselves as the provider of not only information, but of "thinking" as a commodity. Blech.
• Create opportunities to reflect on your own abilities and accomplishments. If the LLM's output encroaches on the bounds of your understanding, put in the work to expand that radius.
• Don't use it for side projects (at least not for most of them).
  • It can be really fun to generate a side project from a loose scaffold and a prompt, but you won't learn anything and you won't care about the end product. Why should you? You didn't work for it.
  • A good side project pushes you, teaches you something, and then maybe turns into a useful tool. The journey is the point, not the destination.

Why follow these rules?

Setting aside the ethical issues of environmental impact, nonconsensual extraction of training data from countless sources of intellectual property and human creativity, the dangerous systematic erosion of the fundamental existence of truth…

Some fundamental principles:

• Friction is good for your body, mind, and soul. Not everything should be easy. Growth comes from the stairs, not the escalator.
• “Use It or Lose It” applies to everything: muscle mass, memory, language skills, cognitive ability, lung capacity.
  • Whatever you give to AI, you take away from your own set of abilities. Are you comfortable giving up what you’re giving up right now?
• Without mindfulness and intention, AI will stagnate your growth as a developer. Your ability to solve problems, learn on the fly, and work hard got you to where you are today. Where will outsourcing those abilities get you?

Before Henry Ford, cars were bespoke, highly crafted pieces of technology. His factory lines made cars faster to produce, but stripped the process of its creativity, removed the artisans and craftsmen from their workshops and replaced them with generic workers executing rote functions again and again. In the interest of efficiency, individual creativity and soul were sanded off the job.

Which is the more human life, the one where we solve problems with our grit, our experience, our gumption? Or the one where we copy output from one text generator into another and back again?

Don’t let AI erode the contributions that only you can make.