Theo Ported TypeScript to Rust with AI and Never Read the Code

Описание и применение

Большой компилятор TypeScript был переписан на Rust с помощью больших языковых моделей, что потребовало $420 000 токенов, но после перехода на другую модель удалось получить полностью совместимый код за $24 000, ускорив проверку типов на 13× по сравнению с оригиналом, хотя конкурент bun check оказался быстрее.

Задача

Can a large compiler be ported by AI without human code review, and what are the cost and performance implications?

Результат

An LLM‑generated Rust port of TypeScript achieved full compatibility for $24k, delivering a 13× speedup over the original compiler, but was outperformed by Bun’s checker. The high initial cost ($420k) and lack of code review highlight risks of agent‑written code.

Рабочий процесс

• 1. Theo used OpenAI models over five months, generating 1.3M lines of Rust code and spending ~$420k in tokens.
• 2. The generated code reached ~84% compatibility but was deemed unusable.
• 3. Theo switched to the Opus model, which produced a working Rust crate in 10 hours for ~$24k.
• 4. The new tsc-rs implementation was benchmarked against tsc 6, tsc 7, and bun check.
• 5. Theo published performance results and a warning that he had not read the code.
• 6. The community compared tsc-rs to bun check and tsc 7, noting trade‑offs in speed and diagnostics.

Доказательства

The port cost over $420,000 in API tokens. Then he says you could probably have done it for $20k.

Источник

On VS Code, a full type check runs in 4.20s against tsc 6's 54.56s. That is a 13x speedup.

Источник

He spent about $400,000 in tokens over five months using OpenAI models, burning through /goal loops. They wrote over 1.3 million lines of Rust and never got past roughly 84% compatibility.

Источник

The final spend was about $24,000 over two weeks.

Источник

On VS Code, bun check runs a full check in 1.62s. That is 33.7x faster than tsc 6, on a 3.75 million line codebase.

Источник