The Rust Rewrite and the Explicit Pay to Win Era

I wrote a post just a few days ago calling the Rust rewrite of Bun an experiment. Today, PR #30412 was merged. One million lines of Zig code were ported to Rust in six days. While the community argues about whether this is a betrayal of Zig or a massive technical flex, they are missing the most important detail. This rewrite was an industrial manifestation powered by infinite tokens and a very strict cop, not a human effort.

We need to stop pretending that Bun was an artisan, hand-crafted project. Jarred Sumner never really hid the fact that he was already using AI agents to write 80 to 90% of the Zig codebase. The memory leaks, the segfaults, and the constant stability issues that plagued Bun were the natural result of giving an AI agent a sharp knife like Zig with zero safety guards. When you let an agent vibe its way through manual memory management, you get a mess that no human can audit. Jarred himself addressed the bad news sentiment on X, noting that most feedback involved crash reports that a borrow checker would have simply prevented.

The move to Rust is a pivot toward Model Alignment. We used to choose languages based on how they felt to humans or how fast they ran on hardware. Now, we have to choose them based on how well an AI can reason about them. Rust is the perfect language for an agent because the borrow checker acts as a formal logic engine. If the code does not compile, the agent has a deterministic error message to iterate on. Rust is the cop in the room that prevents the AI from hallucinating a use-after-free bug. Until agents are 100 percent capable of perfect code, the stricter the language and the tools, the better.

High quality tests are not the source code, but they have become the most valuable constraint we have. If you want a model to reproduce a codebase or port it between languages, the tests are the only way to stop the AI from wandering around. You need a test for every single characteristic you want to replicate. Without that ground truth, the model has nothing to aim at. The 99.8 percent pass rate on the Bun test suite is the only reason this port succeeded. The tests are the mandatory guardrails for a world where AI does the heavy lifting.

Life has always been pay to win, and software is no exception. We are just seeing that reality become much more explicit. Jarred has what most of us do not: infinite tokens and access to models that the rest of the world cannot touch. We are entering a two tier world of software engineering, a sentiment echoed in the heated debates on Hacker News. There are the Architects, who sit at the top of companies like Anthropic and use god tier models to manifest entire systems overnight. Then there are the rest of us, the dev peasants, who have to toil away with cheaper, dumber models while life continues to favor those with the most compute.

The merge of PR #30412 is the finish line for the era of manual systems programming. The soul of the project lives in the scale of the compute and the quality of the constraints used to generate it rather than the craftsmanship of the bytes. I am still watching to see if this code gets reverted or how it will actually perform in production, but the message is clear. The era of the artisan is over, and the era of the machine generated Architect has begun.