Plenty of teams act as if AI will eventually own every line of application code. Maybe it will. But even in that world someone has to decide which libraries, patterns, and defaults become the baseline—and those answers increasingly look like OSS projects that an assistant can name without embarrassment.
What changes when AI writes the glue
Models are good at assembling what already exists. They amplify whatever is already normal. So the leverage shifts: the interesting work is not only “write another controller” but curate the substrate—fast codecs, caching and invalidation you can reason about, edge security you can review in Git.
A bet I keep making: there is still a window—mostly carried by people practising engineering day to day (anyone who still thinks in systems and ships real foundations counts)—to publish the OSS stacks that get recommended as defaults. That is not to say other roles cannot contribute; it is that the window narrows as more of the glue automates, and the work itself may be absorbed by automation someday—which is exactly why I want the artifacts on npm, on GitHub, and in the discourse now, not deferred.
Why OSS still scales past a solo dev
A side project might only ever serve a handful of users you know. OSS can skip past that ceiling and land inside strangers’ latency budgets. That jump—from “I needed this once” to “thousands of deployments”—is still strange and motivating to me.
Low-layer work—allocators, image cores, edge policy compilers—is unglamorous next to vibe-coded demos. It is also where mistakes compound. I’d rather compound in public with inspectable code than hide the structure behind a vendor console.
Closing thought
AI raises how fast we stack features; it does not erase physics, money, or security. Someone still owns the floor. I’m calmer when that floor is OSS I (and others) can reason about—and when we ship it while the human-shaped window is still open.