A 25,000-Line Pull Request Just Killed the Mid-Level Engineer

Who reviews a 25,000-line pull request?

Nobody. That's the point. That's the whole terrifying point of Florian Herrengt's post that just torched Hacker News with 600 points and 487 comments. A senior engineer opens a PR generated with AI assistance, teammates are making design decisions inside chat logs instead of docs, and the codebase is quietly transforming into what he calls an unmaintainable mess. Not because anyone is lazy. Because the machine can generate code faster than any human can comprehend it.

I've felt this. Maybe you have too. There's a specific dread that comes from reviewing a diff so large you start rubber-stamping it out of exhaustion rather than judgment.

The thesis isn't "AI kills programmers." It's sneakier than that. AI collapses the cost of producing code to nearly zero, which means the people who used to get paid a decent salary for reliable, competent, unglamorous implementation work are suddenly competing with a tool that does that specific thing for pennies. Meanwhile the people who can architect systems, catch the AI's confident nonsense, and orchestrate the whole mess become dramatically more valuable.

That's Greg Isenberg, CEO of Late Checkout, and it's blunt enough to make HN commenters furious, which is usually a sign someone said something true and uncomfortable. Allie K. Miller's version is even sharper: companies would rather hire one engineer who knows how to use AI than five who don't. Do the math on that. It's not a productivity story. It's a headcount story.

And then there's Dario Amodei — cofounder of Anthropic, a man with every incentive to hype his own product — suggesting the title "software engineer" might just... stop being a thing, replaced by some universal builder-slash-product-manager role. Coming from the outside that sounds like marketing. Coming from Anthropic's cofounder it sounds like a warning shot.

The data backs some of this up. Stanford's Digital Economy Lab found employment in AI-exposed roles for workers aged 22-25 dropped 4.3% year over year as of June, part of a contraction that's been running for 33 straight months. That's not vibes. That's payroll data.

But CNN's counterargument deserves airtime too: maybe AI isn't deleting the job, it's just relocating the difficulty. Developers now spend more time supervising agent swarms, validating output, and doing architecture instead of typing semicolons. That's real. Anyone who's tried to actually ship a large AI-assisted feature knows the bottleneck moved from writing to trusting.

Hot Take

Here's my controversial opinion: the middle class isn't dying, it's being strip-mined by companies too cheap to pay for proper senior review. The 25,000-line PR problem isn't an AI problem — it's a management problem wearing an AI costume. Teams were already skipping code review rigor before Copilot existed; AI just gave them a faster way to generate the mess and a more convenient scapegoat.

The real divide forming isn't senior-vs-junior or AI-savvy-vs-not. It's companies that treat AI output as a first draft versus companies that treat it as a finished product. Guess which ones end up with unmaintainable systems and which ones end up hiring the expensive engineers to clean it up?

If you're early-career right now, the advice isn't "learn to prompt better." It's learn to be the person who catches the 25,000-line disaster before it merges. That skill was always scarce. AI just made it obvious.