55 lines
4.6 KiB
Markdown
55 lines
4.6 KiB
Markdown
+++
|
||
title = "How I (hardly ever) use AI"
|
||
date = 2025-10-22
|
||
description = "Outlining my AI usage philosophy by answering some questions for my friend."
|
||
|
||
[taxonomies]
|
||
tags = ["Random", "AI"]
|
||
+++
|
||
|
||
A couple days ago, [my friend Ryan](https://ryantrimble.com) asked fellow developers [on Bluesky](https://bsky.app/profile/ryantrimble.com/post/3m3n2vopg322g) about their approach to using AI-powered tools and LLMs for work. I thought I'd use that opportunity to prepare my personal AI statement of sorts, especially since Ryan's given me a complete list of questions for me to reply to in full.
|
||
|
||
## Are you dabbling with it, or are you making full use of it in your work?
|
||
|
||
No, I don't really dabble with (let alone utilise) it on a regular basis. There is one very specific use-case I can think of, where AI and LLMs can indeed prove quite useful, but I'll elaborate on that in a later section.
|
||
|
||
## Is your work part of a collaborative team? Any impact there?
|
||
|
||
Practically every single line of code I've written so far is a part of a solo project of mine, therefore I can't say I've been impacted by AI's output there.
|
||
|
||
## Has it inspired you to build bigger and better things than you may have found difficult or impossible without it?
|
||
|
||
Not at all, because I've got enough confidence in my own set of technical skills to build virtually any project myself. Should I need to expand that set, I've got a wide range of significantly more reliable resources at my disposal - official documentation, books, videos, courses, forums, you name it.
|
||
|
||
## How does it feel to give up control to a virtual agent? Does it require a different mindset or mental model than when you’re handwriting the code yourself?
|
||
|
||
All the aforementioned confidence goes out of the window straightaway. I feel like I constantly need to double-check the output and point out all the mistakes I've found or just rewrite the whole prompt, which gets more tedious as the complexity of a given task increases.
|
||
|
||
I reckon it's more productive to tackle the problem head on with the power of some of the resources I listed earlier and maybe a debugger on your side. That way not only do you manage to solve it, but also gain a proper understanding of its root cause.
|
||
|
||
## What’s been good about it? What benefits have you gained from it?
|
||
|
||
Here's where I have to give LLMs some credit. I believe they do a rather good job when it comes to performing simple transformations over numerous lines of text that would quickly prove cumbersome when done manually.
|
||
|
||
I remember having to convert a [Rust enum](https://doc.rust-lang.org/rust-by-example/custom_types/enum.html) to a module of [constants](https://doc.rust-lang.org/rust-by-example/custom_types/constants.html), because some of the new entries I needed to introduce had their values conflict with existing ones, but the enum in question contained well over a hundred options.
|
||
|
||
This is where I've opted to delegate the conversion to an LLM, since doing it by hand would be much more time-consuming and typo-prone. As much as their generative capabilities are way too hit-or-miss for my liking, I've found their transformative skills reliable enough for me to apply them on large amounts of text.
|
||
|
||
## What’s been difficult about it? What drawbacks have you experienced. Do those drawbacks outweigh the benefits?
|
||
|
||
I think I've already answered the first two questions, so I'll focus on the last one. The fact that an LLM's output is only designed to pass off as written by a human without providing any guarantees of its factual correctness makes it a huge deal-breaker for me.
|
||
|
||
## What’s your overall take? Does it help you enjoy your work more, or less? Does it make you more or less productive?
|
||
|
||
Using an LLM to build a project or solve a programming problem for me feels dishonest. It takes away all the fun from doing these two things, which is completely missing the point.
|
||
|
||
And like I said earlier, having to proofread the reply to my prompt, editing the prompt to provide feedback, and so on until eventually landing at a decent enough solution... that seems quite counterproductive to me.
|
||
|
||
## Does it change the way you view yourself as a developer?
|
||
|
||
Some claim that LLMs are bound to replace us, but judging by all the vibe-coding horror stories I've heard so far, I might as well join the ever-growing market of _cleanup specialists_, where my prescription would always be tearing down the sorry mess of an app/website/game/whatever and writing one from scratch.
|
||
|
||
## How do you plan to use it going forward?
|
||
|
||
I'll stick with employing LLMs to do the dirty work of text transformation if necessary. I can handle writing actual code/articles and problem-solving just fine.
|