February 27, 2026
Two weeks ago Spotify’s CEO said the company’s top developers “have not written a single line of code” in 2026. Around the same time, Stripe claimed they were merging 1,000 PRs per week completely generated by AI. Claude Cowork was supposedly written entirely by Claude Code. YC startups boast that 95% of their code is AI generated. There are countless reports of people vibe-coding useful software without any experience. And lots of current/former programmers are happily churning through their backlog of side projects.
Yet It’s hard to reconcile these stories with the day-to-day reality for engineers in the trenches. Some things are easier and faster to build but many software projects remain just as challenging. What’s going on? Is everyone lying?
Not quite. Sometimes there’s a financial interest (merely mentioning AI can move your stock price/valuation/# of applicants to your startup accelerator).
But the real reason for the disconnect is that not all programming tasks are the same. And coding was never the hard part of shipping complex software.
In June of 2009 a Hacker News commenter responded to a launch announcement with typical HN dismissiveness: “it looks like something that can be thrown together in a weekend.”
The launch announcement was for StackExchange, the generalized version of StackOverflow.
One of the engineers on the team saw the comment and pointed out that even if you had access to the full 96k+ lines of code and started retyping them verbatim at 100 words-per-minute, you still would not be able to recreate the project in a weekend. Not to mention the work required to design, implement, polish, and deploy everything.
A few years later one of the maintainers of the Guava open source project explained why they never accepted patches (external contributions to the codebase). The work done to produce a patch is minuscule compared to what’s required to integrate it internally: reviews, compatibility, migration of legacy code, documentation, testing, long-term maintenance, etc. And after all that grunt work, writing code was the fun reward. Why give that up?
Would these 15+ year-old anecdotes play out differently today? Not really. AI saves you from typing a lot of the boilerplate code but coding agents would fail to produce a scalable, polished, production-ready version of StackExchange. You could build a passable clone quickly but you will overlook key features and your AI implementation will collapse under real world traffic.
AI is great at writing code. It does make engineers faster. It’s incredibly valuable for teams shipping software with low-stakes (e.g., the first version a startup uses to validate an idea).
But for most complex applications this was never the bottleneck.