Not My Problem Anymore

Andrew Kelley published his thoughts on the Bun Rust rewrite two days ago. The photo of him holding a mug says it all: “It Tastes Like It’s Not My Problem Anymore.” I agree with almost everything he wrote.

Kelley is right: Bun’s Zig codebase was “hacks on top of hacks.” Jarred Sumner was writing slop before LLMs existed, pushing features without time for reflection, turning the codebase into a snowball that no one could audit. Kelley is right that ZSF breathed a sigh of relief when Anthropic bought Bun. And he is right that bugs are eliminated by dedicating engineering resources, not by choosing a language.

But there is a hole in his argument.

Kelley frames the rewrite as if the choice was between “well written Zig code” and “badly written Rust code.” That was never the real choice. The real choice was between letting Bun die slowly – a promising runtime, bleeding crash reports, unmaintainable – or using AI plus Rust plus tests to salvage what could be saved. The Zig code was not going to magically improve. ZSF was not going to rewrite Bun. Jarred, with or without AI, was not going to stop producing slop. He was only going to keep producing slop in Zig.

This is where the language choice matters. And this is where my thesis about model alignment finds its most direct confirmation.

Someone needs to ask the question no one is asking: if Jarred had given Claude the same treatment – 64 parallel instances, 165 thousand dollars in tokens, the full harness of implementer, reviewer and integrator agents – and asked “fix this Zig code,” would it have worked?

I do not think so. And I think Jarred knows it would not have, because he chose Rust.

Zig is a language that trusts you. The compiler assumes you are the expert, that you will remember to call defer at the right time, that you know where the memory ends. For an experienced human, that freedom is liberating. For a language model, it is an invitation to disaster. Claude has no intuition about manual memory management. It has probabilities. Without strong constraints, it will hallucinate use-after-frees, leaks, double-frees – exactly the bugs that plagued Bun in Zig.

The borrow checker is not just a guardrail for humans. It is a logical anchor for AI. Every compilation error is a deterministic signal: this does not pass. The model iterates on an objective error message, not on a fuzzy interpretation of “will this leak memory?” Rust’s strictness is what allows AI to work at scale. The stricter the language, the better the model behaves. This is not an argument about human productivity. It is an argument about model alignment.

The result is that the rewrite succeeded. More than succeeded. 535 thousand lines of Zig became over 1 million lines of Rust in 11 days. 99.8 percent of tests passed. 128 bugs fixed. Memory usage dropped from 6.7GB to 609MB. The binary shrank by 20 percent. Jarred accomplished what seemed impossible: turning a rotten codebase into something apparently healthy without stopping development for a year.

That success is the problem. The more the rewrite proves its technical value, the more it validates the mode of production that made it possible. And that mode of production is not available to everyone. Anthropic spent 165 thousand dollars in tokens to save Bun. 165 thousand dollars is the salary of three senior engineers for a year in Brazil. It is the annual IT budget of a medium sized business. For an AI lab, it is a rounding error.

But money is only the first layer of the gap. The model Jarred used for the rewrite — Claude Fable 5 — was in pre-release since April, available only to select organizations. In June, the US government suspended access to the model on national security grounds, and for weeks no one knew if it would return. Jarred’s post came out in July, but Fable 5 had only been released to the general public a week earlier — and even then, with restrictions. Jarred spent two months waiting to be able to tell the world he used a tool the rest of the world had barely begun to touch. Privileged access operates on two levels: access to the capital to pay for the tokens and access to the model itself.

It was not just Fable 5. OpenAI went through the same script with GPT-5.6 a month later — a limited release to a select group of partners at the request of the US government. The turnaround was faster, but the pattern is the same: the most powerful models go through a gate that nobody asked for before they reach the market.

This leads me to a discomfort that might be conspiracy theory, but I will say it anyway. The US government has already shown it can suspend an entire model (Fable 5). It has already asked OpenAI to restrict its GPT-5.6 launch. Access to frontier models is being decided in Washington, not in the market. It is not nationalization yet, but the path is worrying. If the most capable AI becomes a resource controlled by the US state, the gap between who can and who cannot use frontier AI stops being economic and becomes political.

Kelley is celebrating because the problem is no longer his. He is right not to care. But the problem did not disappear. It was transferred. It now belongs to those who do not have access to 64 parallel instances of a model no one else can use. It belongs to those who will compete in a market where the difference between a project that survives and one that dies is access to tokens that cost the GDP of a small country.

The Bun rewrite is a technical milestone. It is also the clearest demonstration so far that programming as we knew it is being replaced by an industrial operation. The artisan died. The architect was born. And the architect, unlike the artisan, does not work for free.