General AI Update

I haven’t posted in a bit, but I feel bad leaving my last post at the top of my blog. I think my position at the time was reasonable: in mid-2025, “I’m skeptical but curious and experimenting” was a defensible stance, but that was a million years ago.

In 2026 (and really, as of late 2025), approximately 100% of the code I write, professionally and personally, is written by an LLM.

I feel like this is the point in every post about LLMs where I walk through the history of the last few years: tab complete -> chat interfaces -> “something changed” in the fall of 2025. But let’s skip all that.

Here’s where I’m at right now:

• Hand-writing code is basically done. Or, at least, it’s done in the sense that hand-building chairs is done: people will still do it, but only as a hobby or point of pride. Effectively all chairs that most people sit in today are built by machines and effectively all software written from here on out will be written by LLMs.
• I’m really sad about that because I love writing code: the craft of it, the puzzle of it, the feeling when an elegant abstraction snaps together… I’ll probably never get paid to do that again.
• I’m really happy about that because I love producing software: the things it allows people to do, the problems it can solve, the creativity of turning ideas into living, dancing screens. This part of the job has never been better.
• I spend most of my time in an agent interface juggling a few different conversations, followed by reviewing code diffs. I use Cursor and Claude Code almost interchangeably for the agent interface.
• Code is cheap, so new code review is the bottleneck. Figuring out how reduce or remove code review - to produce a true dark factory - is the problem of the software engineering industry for the next few years. If you can safely remove that bottleneck, your team can fly.
• The answer to the dark factory problem definitely starts with tools for AI self-evaluation: tests, linters, formatters, type checkers, visual diffs. Smarter AIs will be a component too, as will AI review tools like Greptile or CodeRabbit. There’ll probably be a psychological component too: senior devs just “getting over the hump” of accepting no-review merges. I don’t think we’re there yet, but this seems pretty clearly the direction we’re headed.

So, all that to say: I was skeptical but interested in mid-2025. Today, I’ve long-since accepted that software is written by AIs now and I’m excited to see how the industry changes as a result.