On AI

When I’m writing this, the date is July 27th, 2026. Claude Fable 5 came out, got restricted by the US Government, and came back, all inside the last couple of months. For about three weeks the best programming tool I have ever used was just gone, and I want to be honest about what I actually felt, because it wasn’t anything profound. I was annoyed. I dropped back to Opus 4.8 and it felt like a downgrade. That model would have been witchcraft to me two years ago. That’s the part I keep getting stuck on. Not the capability, the speed at which I stop noticing the capability.

I’ve been using LLMs since GPT-3.5 in 2022 and programming with them since the Claude 3 models in 2024. I remember when getting a decently optimized Python algorithm out of GPT-4 Turbo felt like getting away with something. I remember contributing to AutoGPT in 2023 and thinking all of this was maybe 5 years away. 5 years from AI actually being able to take autonomous actions for you in ways that mattered. Everyone in the circle of contributors thought that. We were wrong about the timeline and right about everything else, which is a pretty weird outcome.

Here’s what my day to day actually is now. Frontier models write 99% of the code my projects need, and the thing I provide is the shape. I sit there for an hour or two figuring out how a system should fit with the dozen systems around it, because that’s the part no model holds in its head, and then I hand it off, and most of the time it comes back right. A year ago it came back wrong constantly, and I caught it because I know architecture and design patterns, and honestly, catching it felt good. It was proof I was necessary. That proof shows up less and less, and I’ve had to admit some part of me misses being needed at that layer, even though the work I do now, the architecture, the verification systems, the deciding what gets built at all, is genuinely more interesting.

I don’t know what to do with that except say it plainly. I’m not scared for myself. The more the models can do, the more my particular way of working seems to matter, defining things precisely, testing everything, having to be creative with project ideas, being paranoid about what I ship. But I’d be lying if I said the ground feels stable. I think radical honesty is the only way through this period for software engineers, and the radically honest version is that I’m having the best time of my life as a developer and I genuinely could not tell you what my job looks like in three years.

If you’re a developer still refusing to touch frontier models, I feel like you’re choosing to forgo efficiency. Whatever the reason may be, if you aren’t making navigation code for NASA spaceships, then I don’t think that’s a virtue. But I get it more than I used to. Some people don’t want the ground moving under the thing they spent ten years learning.

I talk a lot more about this kinda stuff on linkedin, but it felt like something that still needs a solid opinion piece on my portfolio considering how many of my projects are built with these frontier models. Thanks for reading!