Re: The AI has come for my code

GitHub appears to be struggling with measly three nines availability

GitHub AI Agents: 275M Commits Are Breaking The Platform

GitHub's availability has tanked the last few years, with all kinds of outages. This'd be fine if it were still the mostly-hobbyists code hosting platform it started as, but in the last decade, it's worked its way into enterprise organizations, and downtimes absolutely wreak havoc.

The culprit seems to be AI commits, up 10x so far in 2026 from 2025. At the current pace, GitHub expects AI commits to total more than 14 billion this year. I'm well past the point of hoping my fellow developers' bad decisions are going to go away; it's become increasingly clear that we're collectively going to have to learn to deal with the tsunami of shit, because apparently writing and understanding your own code is now a thing of the past.

The AI has come for my code

I found Sean's gemlog interesting: he got a PR on his 6809 assembler, an automated one by some AI company. He breaks it down: the commit message, if you weren't familiar with the project, might seem reasonable. memcpy calls in C source files from (ostensibly) attacker-controlled source buffers.

But when he investigates it, having a holistic understanding of his project (because, you know, he actually wrote all the code), the message and the changes break down. The size param is not solely defined by attacker-controlled assembly, and the fix proposed doesn't even fix the problem-that-ain't. Worst of all, just reviewing that PR took up valuable time. There is a cost to everything, and AI code is antisocial. It makes work for people to read and understand the code. Automatically writing PRs and slinging them at people's repos is deeply antisocial behaviour.

I've been lucky so far: I've got some public repos, and so far, nobody's sent me automated PRs. Not that I'd accept them, anyway: I'm clear in the README that I'm not likely to accept any pull requests (that was written a decade before LLMs ate my colleagues' brains, so applies to both human and AI contributions).

But I'm also active in literary spaces, and have seen, first- and second-hand, just what LLMs are doing to submissions over there. Earlier this year I got a ton of AI submissions from submitters with Indian names (not sure if the perp was actually Indian, or just impersonating), and I could tell they were AI because there were a few different "base" poems from submission to submission that I could tell were based off the same prompt. They were unexceptional. Better than my awful teenage poems from the mid to late 90s, but in no way good, or publishable. Probably they came my way because I actually pay contributors to my little magazine, and maybe someone thought, it's a numbers game, if I send enough submissions that look like poetry, one of them will be accepted (yeah, wrong).

I hate AI code and I hate AI agents and I hate how it's, to borrow from my younger coworkers, cooking everyone. There are stink lines above this shit. It's not awful, but it's definitely not good, and when you're looking for excellence, or correctness, it only looks like such if you're dumb. That's what I keep coming back to: with one or two exceptions, the people I know who're most excited about AI code are the dumbest, most boorish people I've ever met in my life. There's the vainglorious friend-of-a-friend who works at DraftKings, writing code to break families and ruin lives; there's the talented musician who can't keep a job as a band teacher because he's egotistical and unreliable and lately has been teaching high school computer science and advocating AI coding to his students (yikes); and then there's the guy I know who moved back to Israel from Canada in his 20s, who has some sort of company-of-one, and who's been all-in, in order, on online gambling, NFTs, the metaverse, and now AI.

I'm sure it looks great if you don't know what the hell you're doing (or, in the case of the DraftKings guy, if you don't care). If it doesn't matter to you what you're making, just that it's been made. In the coming decades, assuming this rock is still habitable, there's going to be endless articles written about legacy code in the age of AI, how people used it for everthing, and how they understand none of it.

gemlog