I am writing this at what I hope is the peak of AI assisted software engineering. Anthropic’s Fable 5 has just released, and somewhere deep down I am glad it’s ridiculously expensive.
On the one hand, it has never been easier to build that bespoke application you’ve always wanted, or to solve an esoteric bug in a huge codebase that’s actually costing your company real money every second that it persists. But on the other hand, it feels like the art of programming is slowly disappearing. Not because we don’t want to put in the effort — but more so because the architecture you would’ve jammed out in your editor during a 4-hour non-stop coding session, fuelled only by coffee and an insatiable desire to build something nice, can also be vibe-coded in fifteen minutes. With some extra prompting, exactly how you would’ve done it.
I’m in both camps simultaneously
The irony is real. Only a year or so ago I’d be shitposting on LinkedIn, making fun of the AI hype train. Saying things like “computers shouldn’t be glorified autocomplete engines.” I still stand behind that, because LLMs back then were genuinely bad at writing code. I’m the kind of person that’s both cynical and lighthearted enough to poke fun at that — and the fact that “hype bros” were essentially praising LLMs to be some kind of godsend.
Those days are over.
I’ll admit it: with Claude Opus 4.8 or GPT 5.5, you can get something decent within a weekend. On large-scale software, that’s a different story — and we do need real software engineers for that. Don’t get me wrong: this post was never about some kind of self-induced fear for job security. This is about not losing the joy of what got us (or me, at the very least) into software engineering in the first place.
It has been days since I’ve last opened up Neovim to do some actual hand-crafted programming, other than learning Rust in my free time (which admittedly I am partially doing to get out of the AI psychosis). The people I work with expect an end-to-end feature to be built out in days, rather than weeks. They expect bugs to be fixed in minutes, rather than hours. Part of that is my own fault. My obsession over speed has gotten the better of me. It’s time to go back to hand-writing code. Slowly. Deliberately.
At the same time, I can now jam out a fully vibe-coded internal tool or the piece of software I’ve always been missing. Some examples are: a Mattermost TUI, a git GUI client that supports a keyboard-driven workflow and this very website. Yes, this website is built with AI assistance — unapologetically, but this post is still hand-written.
The silver lining
Let’s be honest: LLM inference is expensive. It has been subsidized by frontier labs for the better part of three years now, and I’m expecting it to get more expensive before it gets cheaper. Maybe when the time comes, we’ve gone back to the good old days of using AI as a tool instead of outsourcing all critical thinking to it. Maybe lawmakers will realize running data centers at full throttle 24/7 isn’t doing the planet a service. And maybe restrictions from private companies and open source initiatives will force us out of the AI bubble. (Thanks, Flathub <3)
Or maybe… just maybe… we’ll all be harvesting potatoes this time next year.