A small thing happened this week: a Ruby developer announced, in a post titled “Why I’m no longer updating this repository”, that he is archiving all of his gems — the little open-source libraries he has been maintaining for years. His explanation is elegiac. He wrote the code, he says, in an era “when Ruby was ostensibly for joy, for creativity, for humans,” and “that age, of doing things for humans, has passed.” The machines write for machines now, and he doesn’t see his role in it.
Posted to Hacker News, the obituary turned into something more interesting: an autopsy of a promise. The argument, kicked off by commenter newprint, is that we were told large language models would set off a software renaissance, and with two or three years of accelerated agentic development behind us, the renaissance is not evident in the output. Where are the new CAD packages more capable than Blender, FreeCAD or KiCad? Where is the AI-native replacement for the Adobe Creative Suite? Where is the new operating system? Microsoft and Google, newprint points out, are as deep in AI as anyone alive, with frontier models and thousands of elite engineers, and yet Windows and Android are exactly as laggy and buggy as they were four or five years ago. Reverse engineering and vulnerability-hunting have gotten easier, he concedes. “But where is the new cool software?”
It’s a fair question, and the thread’s best answers all converge on the same idea, which I think of as the cheap-labour fallacy. If code is now nearly free, then everything that is not code becomes the price of software.
Commenter jerf describes the lived version. For months he ran multiple parallel AI-driven work streams at his job; lately he’s back to one stream, because every one of his streams eventually hits “some human blocker” — QA to test it, a colleague’s attention for the next step, a deployment. Code generation was never the whole pipeline, so once you’ve burned through the backlog of “all the stuff we know we needed but didn’t have the coding time to get to,” the frontier of work simply advances until it reaches a human again. His guess, invoking Amdahl’s Law — speed up one component of a system and your total speedup is capped by the share of time that component took — is that the steady-state gain from nearly-free code may be “only a maybe 2x or 3x improvement” even in software, the best possible domain for AI. “If this is true,” he adds, “goodness help the financial bubble which is assuming that something like 5x or 10x indefinitely is a lower bound.”
Others think even that is generous. Commenter uludag argues that a true 2x would show up as either half the software-engineering workforce being let go or a visible revenue surge at AI-heavy companies, and with about 1.7 million software engineers in the US alone, neither has happened on that order; his estimate is a single-digit-percentage total productivity increase. Commenter ryandrake does the arithmetic of an ordinary enterprise release: designing and writing code — “basically the part that LLMs can do” — was maybe 5 to 15 percent of the calendar time to ship. The rest was stakeholder alignment, security reviews, privacy reviews, legal reviews, QA, lining up marketing, waiting for people to return from vacation, waiting for someone to click “approved.” Make the 5-to-15-percent component ten times faster and the release takes 95 to 104 percent of what it used to.
Sure. And it isn’t only intuition. The research nonprofit METR ran a randomized controlled trial with experienced open-source developers working on their own repositories, and its early-2025 result remains the most quietly devastating data point in this debate: with AI tools, the developers took 19 percent longer to complete their tasks. The kicker is the perception gap — they had expected to be sped up by about 24 percent, and after the fact believed they had been about 20 percent faster. They felt the renaissance; the stopwatch said otherwise. METR revisited the experiment in February 2026 and found “some evidence for speedup” in its newer cohort, so the number is moving in AI’s favour as tools improve. But the careful, measured literature on this is currently arguing about whether the gain exists at all, not about whether it is 5x or 10x.
The renaissance, plausible deniability edition
Now, the strongest counterarguments in the thread are not that the renaissance failed, but that you’re looking for it in the wrong places. Commenter Doohickey-d suggests the real output is bespoke software built by people for themselves, “often with a user count of 1” — a huge change for that one user, and completely invisible to the rest of us. Commenter satvikpendem offers a concrete positive case: on his MacBook, games that never had native ports now run, via vibe coding and AI-assisted reverse engineering, work that “would have taken years before.” Commenter wvenable argues the whole complaint is premature — even six months ago the tooling was much cruder, so we’ve had perhaps a year of using it effectively — and adds the epistemological jab: “are you sure you even have a way to notice all the new cool software that exists?” Commenter djeastm thinks that’s a problem in itself: so much new software arrives that he bookmarks impressive projects and never returns, because something newer lands a week later.
My favourite objection comes from commenter Buttons840, who makes the renaissance falsifiable. A convincing sampled piano as a VST plugin sells for $100 a licence or more; a genuinely better geometry kernel for CAD software would be worth thousands of dollars a seat. Neither requires distributed systems or venture capital — the VST is just math behind a bare-bones UI. If code is nearly free and talent is abundant, where are the new pianos? “Who is vibe coding a geometry kernel?” That’s a market test: the bottleneck theories predict exactly this emptiness, because the hard part of a geometry kernel was never typing it — it was the decade of correctness, edge cases and trust that no one will buy without.
And lurking under it all is a bleaker possibility, raised by xorcist and wvenable: commercial software quality has been declining for a decade, LLMs or no, because fewer companies want to pay for QA and everything, software included, has been optimized to extract the most value at the least cost. If the prevailing use of nearly-free code is not to make better things but to make the same things with fewer people, then the renaissance was never going to appear in your applications menu anyway. It was always going to appear in a margin.
Anyway, one developer just retired his gems because the age of writing for humans has passed. If the economics work out the way the thread’s gloomier voices expect, the age of reading things written for humans may be in some trouble too.