On July 8, Jarred Sumner published a postmortem titled “Rewriting Bun in Rust.” The numbers were staggering: 535,000 lines of code, ported from Zig to Rust, in 11 days, using 64 Claude agents running in parallel. By July 19, Simon Willison had confirmed that Claude Code — Anthropic’s own developer tool — was already shipping with the new Rust-based Bun runtime inside it. The loop was closed: Anthropic’s AI rewrote Anthropic’s runtime, and Anthropic’s tool now runs on it.

The Hacker News thread lit up. The Register ran a piece. Andrew Kelley, Zig’s creator, called the whole thing “unreviewed slop.” The debate that followed was entirely predictable. Is AI-generated code good? Is it safe? Is this the future of software engineering or a cautionary tale?

All of that misses the real story.

The Safety Company That Forgot to Ask

Anthropic was founded on a premise: AI development is dangerous, and someone needs to do it carefully. The company’s entire brand is built on caution — constitutional AI, responsible scaling policies, the long, careful path to powerful systems. It is, in its own telling, the grown-up in the room.

So what do we call it when that same company takes 535,000 lines of machine-generated code — produced by 64 parallel agents over 11 days — and ships it to millions of devices without, by its own admission, a line-by-line human review?

We call it a Tuesday, apparently.

The Bun rewrite is not a small side project. Bun is a JavaScript runtime and toolchain that underpins Claude Code, which is now one of Anthropic’s flagship products. When you install Claude Code today, you are running on infrastructure that was written, in large part, by Claude itself. This is not dogfooding. Dogfooding is when you use your own product. This is something closer to a snake eating its own tail and calling it lunch.

The Review That Wasn’t

Sumner’s postmortem describes the workflow in detail: Claude agents translated Zig files to Rust, an adversarial reviewer agent checked the output, and the existing Bun test suite caught regressions. The CI pipeline ran. Tests passed. Ship it.

This is, on paper, a reasonable process. It is also a process in which every gatekeeper — the writer, the reviewer, the test suite — was either an AI or a test harness designed for the old codebase. No human being read most of those 535,000 lines. No human being could have. At 11 days, that works out to roughly 48,000 lines per day. Even if Sumner did nothing but review code for 16 hours a day, he would have had to evaluate 3,000 lines per hour — roughly one line every 1.2 seconds — to cover the whole thing. He didn’t. He couldn’t.

One engineer at a competing AI lab put it to me in a Signal thread after Kelley’s response dropped: “We spend millions on alignment research, red-teaming, and staged rollouts. They just YOLO’d half a million lines into prod. In eleven days. And their safety team signed off on it.” The engineer asked not to be named because their employer has a policy against commenting on competitors. The policy exists for a reason. The reason is moments like this.

The Conversation Nobody Is Having

The AI safety debate has spent years focused on existential risk — the paperclip maximizer, the misaligned superintelligence, the godlike system that optimizes for something we didn’t intend. These are important questions. They are also, conveniently, questions that cannot be falsified today. They require no immediate action beyond funding more research. They let everyone sound serious while changing nothing about how they ship code on a Tuesday afternoon.

The Bun rewrite is a different kind of safety question — the boring kind. It asks: what happens when an AI system produces code that passes all the tests but contains a subtle, latent bug that no human has ever looked at? What happens when that code runs on millions of machines? What happens when the company that shipped it is the same company that tells governments it should be trusted to develop AI responsibly?

The answer, so far, is that nothing has gone wrong. Bun v1.4 shipped. Claude Code runs on it. The tests pass. The benchmarks look fine. But “nothing has gone wrong yet” is not the same as “this was a good idea.” It is the same logic that justifies driving without a seatbelt because you haven’t crashed.

The Real Cost of Speed

There is a version of this story where Anthropic’s move is genuinely impressive — a demonstration that AI-assisted development has crossed a threshold. Eleven days for a 535,000-line port is, by any historical standard, remarkable. If the code holds up, Sumner and his team will have pulled off something that would have taken a traditional engineering organization the better part of a year.

But the speed itself is the problem, not the achievement. The reason software engineering developed practices like code review, staged rollouts, and human sign-off was not because engineers are slow. It was because we learned, over decades and at enormous cost, that speed without verification produces systems that fail in ways nobody predicted. The Bun rewrite short-circuits every one of those lessons in the name of getting to market faster.

Anthropic will point to its test suite and its adversarial reviewer agent and say the verification happened. But verification by the same class of system that wrote the code is not verification. It is an echo. And echoes, in engineering, are how you miss the thing that actually breaks.

The irony is that Anthropic knows this. Its own research papers warn about the risks of over-reliance on AI systems. Its own safety commitments promise careful, staged deployment. And yet, when the opportunity arose to ship faster than any competitor could match, the safety framework turned out to be optional.

That is the real story of the Bun rewrite. Not whether the code is good. Not whether Rust is better than Zig. Not whether Andrew Kelley was right to call it slop. The story is that the company most associated with AI caution just demonstrated that, when speed and safety conflict, speed wins. Every time.

If you were waiting for the grown-ups to slow things down, you can stop waiting now.

Sources