How AI changed what I can build

In late summer 2022, I got hooked on Midjourney, but kept running into its limits and costs. I did some research and found Stable Diffusion, which I could run myself. Luckily, the 3090 I’d bought for gaming back in 2020 could run it.

When I showed my friends, only one had a GPU that could run it. I built a Discord bot so everyone else could use mine, giving our server its own little Midjourney at home.

“An astronaut riding a horse on the moon.” One of my first Stable Diffusion images

Generating images was basically all we did in that server for the rest of the year. My room was always hot, and the power bill went up.

I signed up for OpenAI around the same time. DALL-E was what I got access to first. Between that and the experiments in Discord, I was hooked on making funny images with my friends. When ChatGPT came along later that year, it felt like science fiction.

Even here in Silicon Valley, working at Verkada, I was far from an AI expert. Getting to ask it questions myself was different. I used it to learn about things I didn’t understand, write funny copypasta, and troubleshoot problems at work. It gradually became my Google replacement and proofreader.

In summer 2025, I started using Claude. Opus 4 and Claude Code set off a fire in my brain.

I’d been adjacent to code for a long time. I’d made simple web pages, utility scripts, and a few Discord bots, but building full-stack software had never really clicked. I already liked making things for myself and my friends. Claude made it possible to take on much more complicated projects.

Working with a coding agent, I could describe a problem and have it work on the code, then try it and figure out what needed to change. I could keep going on projects where I’d previously have gotten stuck.

Homelabbing became fun again. Before, the thought of configuring a Plex server after a workday felt like a slog. Now I try things I wouldn’t have attempted on my own, and getting stuck doesn’t automatically mean spending the rest of the evening on one configuration problem.

I’m more willing to start a project when I don’t have to know everything upfront. I can get something running and use it for a while before deciding how much more work I want to put into it.

Learning to delegate

A lot of the work starts to feel familiar if you’ve managed people. You have to explain what you’re trying to accomplish and what a good result would look like. You also have to notice when the agent is doing a perfectly reasonable job of solving the wrong problem.

In that sense, more of us are becoming managers. Even without direct reports, we’ll be responsible for work we didn’t do by hand.

An agent can produce a lot of plausible work very quickly. If I don’t understand it well enough to evaluate it, I can end up with a bigger problem than the one I started with.

The curious engineers seem especially well placed here, whatever their role. If you understand how a system works, you have a better chance of noticing a bad assumption before the agent builds on it. Knowing where to look when something breaks is still useful, even if you’re writing less of the code yourself.

I also expect people’s scope to grow. There’s plenty of work you might understand well enough to help with but have to hand off to another team. Then it sits in their queue until it becomes important enough to pick up.

A support person who understands a recurring problem may be able to help fix it rather than only explain it to engineering. An operations person may be able to build the tool they’ve been waiting for. Once that’s possible, I doubt the expectations of those roles will stay the same.

For the kind of technical work I know, I’d prepare for a much bigger scope. I don’t know how this will shake out for jobs overall, but I expect people who are good at figuring things out to have a lot more they can take on.

Some things I’m watching: benchmarks, usage, and employment.

“An astronaut riding a horse on the moon.” The same prompt four years later, using OpenAI’s latest image model.

Personal software

At home, I get to choose how much of that extra work is worth taking on. A project still needs maintenance, and I have to think about what access it should have and what happens when it breaks. But getting a first version running is cheap enough now that I’ll try things I wouldn’t have bothered with before.

A tool can fit an oddly specific workflow without needing to be useful to thousands of people. Usually we settle for whatever existing software comes closest. If I can make the thing I actually want without it turning into a huge project, I’d rather do that.