I tend not to be a tech absolutist about most things. I like to play in the tech space and test new tools out to see what they can do, see if they can help me and my work, but if they can't, I move on and avoid them.

I want to know what's possible (both bad and good) so I've been playing with AI through Claude for the past couple months in a casual manner, usually spending an hour or two trying things out or leaning on it to fix something for me.

So far, I've come up with a few things after this experience:

1. AI is ridiculously good about building single-user, single-purpose software that is fairly brittle (and works just enough), but is also the kind of software with no market because there's only one person on earth that would use/pay for/love it.
   
   People are calling this "disposable software" and that's a good description. Most of positive experiences I've had with AI coding are making things only I need for a specific use case.
2. If you're a maker and a DIYer and a jack of all trades, you're completely used to doing everything by hand from soup to nuts to ensure it's exactly what you want.
   
   But in any project, there are tons of things that take time though not much mental effort, even though your project might rely on it, and that's where AI really shines. "Go look up every mention of things I wrote and add hyperlinks to my original pieces in this giant pile of text and give me the result back as HTML I can paste into the project" is something I normally do myself, but it's mindless work that sucks up hours of time and effort, and when you can ask Claude to do it and it's completed in 20 seconds, you can continue working on your project knowing you saved some serious time. This is not nothing.
3. If you are full of ideas, AI is great at making prototypes.
4. If you can code, you can fix things a lot faster when AI inevitably writes janky software.
5. If you've ever worked on a software team, AI feels like pair programming. AI feels like working with an intern, loaded with energy but frequently wrong because without deep personal experience, they're guessing a lot of the time.
6. So far, the "killer app" aspects of AI is helping me create/iterate/finish things I've wanted for many years, but couldn't find time or energy enough to do completely on my own.

Here are a couple projects that I'd say AI did a very good job with. They are both things I've wanted to do for years, but didn't have time for, and when I proposed the idea in Claude and eventually got a result, it surprised me at how well it went.

Project 1: Tagpocalypse

I've had this blog for 26 years and up until about 2010, I stuck to broad categories, but later I started tagging my posts and ignoring categories. Each time I'd post, I'd try to put in a little extra effort and pick descriptive tags. But years later, out of nearly 3700 posts, only about 1500 had proper tags on them that described what each post was about.

The other day I was going through my archives, making small updates and edits, then adding tags to posts that didn't have them. I quickly realized doing this for every post was an impossible project to ever see the end of (especially when I'm the only one that cares about having things well-tagged and described). So I asked Claude to help me complete this project.

I had it produce tags for a few random untagged posts, checked the suggestions, refined its approach, then asked it to tag all 2,200 posts. It also wrote a python script to let me update all 2,200 posts on this site. It finished the job in a few minutes, which was incredible. I asked it to make a sortable, usable tags page as well and got that in a few more minutes of iterations.

Then it surprised me by asking if I wanted a summary of how my archives look and I said sure. Here's that result:

These are great insights I wouldn't have picked up on otherwise. You can see in the early days most posts were only a sentence or two, then Twitter launched in 2006 and I basically stopped blogging. Eventually I tended to blog longer essays so much so that now my posts average over 1,000 words each.

AI is also great at inadvertent comedy and this just slayed me:

internet + peeves is the most frequent pair (53 times) — a tidy summary of about 60% of your output.

Cold hearted, AI, calling me out for my Andy Rooney tendencies.

I'm happy this project is done. It means it's easier to look up and search old posts, the related posts blob at the bottom of posts is more accurate, and I'm glad I didn't have to do things 2,200 times by hand.

Project 2: a gym workout app

When COVID lockdown started, my gym closed for nearly a year and I bought a cheap Bowflex unit from Amazon in the spring of 2020. This was before Amazon increased the price of the same device by almost 3x when everyone wanted one later in the summer of 2000.

For years, I'd been working with a trainer on a whole body workout that focused on keeping my core strength up, and through trial and error, I found this video uploaded by Bowflex to be the closest thing to feeling like my real workout with a trainer, as it hit all my muscle groups and worked great for me.

Whenever I wanted to work out, I'd walk down to my barn, pull out my phone and look up this old video. I'd watch it in ten second chunks to remember what was next and to check my form, then I'd put on music instead and do my reps. After, I'd jump back into the YouTube app, unpause for ten more seconds, and repeat.

From one workout to the next, I might bump up my resistance through the weighted bands the unit comes with, and I always wanted to track where I was with that. For years, I did these workouts occasionally, with my clunky phone app switching, wishing I had a whiteboard nearby to jot down how much weight I was doing each time.

Fast forward to last week when I asked Claude to take a crack at it. At first, it couldn't load a YouTube video, so I ripped it to a downloadable copy from an online utility. Claude went to work and with almost zero input, gave me a list of all nine exercises, a short looping animated GIF clip for each exercise to check my form, then added a input where I can type in how much weight I used (and the form input defaults to the previous workout's numbers). It even produces graphs of weight over time after you do a few workouts.

This is something I've wanted for six years and in an hour, it created a perfect single-user web app that I host on my home network and use to track workouts.

The harsh realities of AI

I have no doubt that AI is killing the tech job market. I know I have lost at least a couple gigs because a higher up figured paying for AI tokens was easier than paying a real writer to do documentation. And I have friends all up and down the content industry and it's clear everyone is hurting looking for rare roles these days because whole jobs are being eliminated in the rush to AI.

I've come up with my own guardrails for AI and I think everyone should figure out what they're comfortable using it for and what they won't use it for before getting in too deep.

I tend to use AI for the beginnings of general research on something I don't know a lot about (instead of basic Google searches, because those have gone to shit over the past decade), but I use it as a cursory step to figure out where I should research on my own next. It's dangerous to trust AI completely on things you don't know deeply because it's just guessing what most people think is correct about a subject, in a ten-mile-up sense.

AI is at its best when you have a good deal of knowledge about a subject that you're seeking help with, because you'll quickly spot where it is making mistakes, and what assumptions it is making that are misguided. It allows you to quickly course-correct and do things like I've described here that save me oodles of time versus doing them myself.

AI is going to kill the junior job market in every industry and that sucks. Working on a coding project in Claude feels a lot like when Slack had tons of interns eager to chip in on projects and they'd fill our teams. Now, if I'm using $20 worth of AI to automate rudimentary tasks in my own life, I know the Fortune 500 are doing it at a much larger scale and that makes me wonder who is going to hire people and let them become junior software developers who eventually become senior developers if we've completely offloaded their introductory work to bots?

An obvious guardrail for me is I don't use generative AI to write things for me or make ugly images or produce videos. I have seen almost zero upside and nothing but downsides to almost all of it. AI-produced images and video build distrust among everyone, as we all tend to fall for doctored images on a daily basis now, making everyone question everything.

The downsides of AI are serious and problematic. There is an environmental cost to all the electricity needed by data centers. Companies and cities are ramming through data center construction projects that are propping up another tech bubble that will inevitably burst, wrecking our economy. AI seems to bring out the worst in everyone involved in business, much like crypto did.

I don't think dozens of agents doing your bidding is a good approach to building things with AI. I feel like a weird Portland "artisanal AI" user because I still interact with Claude inside of chat sessions. I break tasks into small parts so I can review the output and see the approaches being taken. I debug stuff at every chance and come up with solutions to bugs myself more often than Claude does. Contrasting this with the idea of OpenClaw, or having dozens of agents building stuff for you, who check things for bugs, and fix it without any of your interaction or input, and that approach seems wild to me. I want to know what's going on under the hood and I make sure I keep tabs on what AI is doing.

I don't know what the future holds for the software industry, but it feels like every job role in tech is changing a bit to accommodate what is newly possible thanks to AI models, and there are plenty of downsides to its adoption.

I still think there's something worthwhile that AI can be safely used for, whether that's building utilities and prototypes for a small audience, or LLMs based on your own writing that help you analyze your own work, or the hours of IT help I get by asking Claude which linux commands I would type to fix the things that are broken in my tech stack.

I've always had half a dozen iPhone app ideas, a couple dozen website ideas, and a bunch of things I'd call art projects, but it was hard to carve out the space to work on them while I was doing other things. Now, I can knock things off a mental todo list in a day or two, when they might have been brewing in my mind for nearly a decade. When I sit down each night and take a crack at an idea I've had and the results are better than I ever expected, that's pretty incredible.