Paul Dix dropped a bombshell last week, and almost nobody noticed. An AI wrote a million lines of code, refined it over months, and shipped it to millions of developers. The software works. It’s reliable. It’s running in production right now.

The industry’s response? Crickets. Or worse, dismissals. “Well, it had an oracle to compare against,” people say. “It was just translating from one language to another.”

That’s exactly the wrong reaction.

The Translation That Changes Everything

Here’s what actually happened, based on Dix’s account: a team used AI to perform a massive codebase translation. Not a trivial script conversion, but a full million-line rewrite. They built a verification system to check the AI’s work against the original. Then they spent months refining the output until it was production-ready. Now it’s deployed on millions of machines.

Meanwhile, over at EVE Online, CCP Games is embarking on their own translation project. They’re migrating 2.4 million lines of Stackless Python 2.7 code to Python 3. The codebase has been running since 2003. Their last major upgrade was 16 years ago. They’re planning to use the futurize script, then manually review the ~20,000 places where Python 2 and 3 behavior differs.

One of these projects used AI. One is doing it the traditional way. And the gap between them tells you everything about where software development is heading.

Why Dismissing This Is Dangerous

When people hear “AI wrote a million lines of code,” they immediately reach for reasons it doesn’t count. It had an oracle. It was just translation. There was a verification system.

But stop and think about what those objections actually mean. Having an oracle means you can verify correctness. Building a verification system means you can iterate confidently. Translation work means the problem is well-defined and testable.

In other words: this is exactly the kind of software engineering work where AI absolutely should excel. And apparently it does.

The EVE Online migration is instructive here. This is hard work that requires deep understanding. You’re dealing with two decades of accumulated complexity, Stackless Python’s unique characteristics, and thousands of behavioral differences between language versions. Getting it wrong means breaking a game that hundreds of thousands of people play simultaneously.

CCP is doing it the careful, manual way. And they should. But the fact that we now have an alternative approach, one that demonstrably works at even larger scale, is something the industry needs to reckon with.

What Nobody Wants to Admit

Here’s the uncomfortable truth: if AI can reliably translate a million lines of code with proper verification, it can probably handle a lot more software engineering tasks than we’re ready to acknowledge.

The pattern Dix describes is actually generalizable. Build a verification system. Give clear direction. Let the AI generate and refine. The fact that translation work has a built-in oracle doesn’t make this trivial. It makes it tractable. And there are a lot of software engineering problems that can be made tractable with the right verification approach.

Refactoring? You have tests as an oracle. API updates? You have the old API’s behavior. Data migrations? You can verify the output against the input. Performance optimization? You can benchmark.

None of this means AI is going to automagically solve all software problems. But it does mean that a whole class of engineering work, work that currently requires senior developers spending weeks or months, might be getting dramatically faster.

The Migration Industrial Complex

Consider what this means for the massive backlog of technical debt sitting in every organization. All those “we should really upgrade this someday” projects. All those legacy systems that nobody wants to touch because the original developers left five years ago and the institutional knowledge is gone.

Python 2 reached end-of-life in 2020. Six years later, EVE Online is just now starting their migration. That’s not because CCP is slow or incompetent. It’s because this kind of work is expensive, risky, and requires deep expertise. You can’t just throw junior developers at it. You need people who understand both versions of the language, the codebase’s history, and the subtle ways behavior can diverge.

Or you need AI with a good verification system.

Why the Skepticism Matters

The industry’s immediate dismissal of this achievement is revealing. We’re not taking it seriously because we don’t want to take it seriously. Admitting that AI can handle million-line rewrites means admitting that a lot of senior engineering work might be getting commoditized.

But that skepticism is dangerous. Not because AI is going to replace all developers (it’s not), but because we’re not preparing for the shifts that are already happening.

If your company has a six-month Python 2 to 3 migration on the roadmap, you should be asking whether AI could do it in six weeks. If you’re planning a language translation or framework upgrade or major refactoring, you should be investigating whether the pattern Dix describes applies.

And if you’re teaching computer science or training new developers, you should be thinking hard about what skills actually matter when AI can handle massive code translations reliably.

The Verification Breakthrough

The real story here isn’t just that AI wrote a million lines of code. It’s that someone figured out how to verify and refine AI-generated code at that scale. That’s the engineering breakthrough that matters.

Verification systems turn AI from a novelty into a tool. They let you iterate. They give you confidence. They make the probabilistic nature of AI output compatible with the deterministic requirements of production software.

EVE Online’s manual review of 20,000 behavioral differences is exactly the kind of work that AI plus verification should be able to handle. You can programmatically identify the differences. You can verify that the behavior matches. You can catch the edge cases through testing.

Will it catch everything? No. Will it need human review? Yes. But “AI does 95% of the work and humans verify” is a radically different project timeline than “humans do 100% of the work.”

What This Means

We’re at an inflection point in software development, and most people are looking the other way. An AI rewrote a million lines of production code, and the industry shrugged it off because it “just” did translation work with verification.

But translation work with verification describes a massive portion of software engineering. And if AI can handle it reliably at million-line scale, we need to stop pretending this doesn’t change anything.

The question isn’t whether AI can write code. It demonstrably can. The question is what we do with that capability. Do we keep running expensive, slow, risky migrations the old way? Or do we figure out how to build verification systems that let AI handle the grunt work while humans focus on the problems that actually require creativity and judgment?

EVE Online’s migration is going to be fascinating to watch. But somewhere out there, someone is probably already running the same 2.4 million line codebase through an AI with a verification harness, just to see what happens.

I’d bet on them finishing first.