Claude Fable wrote the Goose compiler, 30,000 lines of C++

A language designer wanted to finish a project he had put aside for years. He gave the design to Claude Fable and let the model write the compiler.

Wouter van Oortmerssen published the Goose programming language this week. Goose is a memory-safe systems language. It has no heap, no garbage collector, and no lifetime annotations. The home page states it clearly. "This repo was produced almost entirely with AI, from a human language design." The Hacker News post has 172 points.

What the two authors did

The project page lists the split of work. Van Oortmerssen did the language design, the compiler design, and the coding standards. Claude Fable did the compiler implementation, the benchmarking, and the samples and documentation.

Van Oortmerssen designed Goose years ago. He also wrote Lobster, another programming language used by his game company. He says he had never expected a model to help with this work before Fable came out. He asked Fable to clone the style of the Lobster compiler. He writes that there was a long exchange with the model. He says it forced him to decide on parts of the design he had left open.

What the compiler is

The result is a whole-program compiler of about thirty thousand lines of C++. It compiles to one C file. So it runs anywhere a C compiler does. A bundled TinyCC backend can also run a program in the same process. The project has a specification, a standard library, and thirty-one sample programs. A test suite runs on Windows, macOS, and Linux. The GitHub repository was created on 30 August and has 1,854 stars.

The benchmark numbers the project cites

Over sixteen benchmarks on the MSVC backend, the project reports three results. Goose runs at 3.51 times the speed of idiomatic C++. It runs at 1.24 times the speed of hand-optimised C++. And 1.13 times the speed of the best safe Rust. Peak memory use is 1.95, 1.32, and 1.25 times less against the same three. These are the project's own numbers.

The gains come from the memory model. Data lives inline on stacks that the compiler manages. References stay valid while an array grows. Freeing a nested value is a single store, however deep it goes.

The teams most likely to try this method are the ones building their own small language or tool. That work is hard to get right by hand. The code is small enough that one person can review every line the model writes. The Goose repository is one real example of this method: a working compiler, benchmark results, and a human author who read all of it.

Source

The Goose Programming Language, by Wouter van Oortmerssen. Discussion: Hacker News.