A developer reports a fast Rust implementation, but compatibility and total project cost need careful reading.
The Register reports that developer Theo Browne used AI agents to port Microsoft’s TypeScript implementation to Rust. Browne claims the result runs twice as fast as the official Go rewrite and 12.5 times as fast as the previous TypeScript version. Those are his reported comparisons, not independently reproduced results.
The cost needs context. The article describes roughly $24,000 in API-equivalent tokens for the later Claude-powered attempt, accessed through subscriptions, after earlier attempts consumed more than $400,000 in API-equivalent tokens. The successful approach translated the Go implementation rather than starting from the original JavaScript. Browne says he had not inspected the code himself.
The report says the implementation reached nearly complete conformance on TypeScript’s tests. That makes the experiment concrete enough to investigate, while leaving independent compatibility, security and maintenance questions open.
In plain English
A “compiler” translates code into something a computer can run. “Rust” and “Go” are programming languages. A “port” is a rewrite of existing software in another language.
Why it matters
It is an unusually large example of agents doing a big rewrite largely unsupervised, though the unreviewed code and self-reported speed mean the result is a claim, not a finished product.
A compiler has a particularly demanding definition of success. It must understand valid programs, reject invalid ones consistently and preserve the behavior developers expect. Speed is valuable only after that compatibility work is done. A fast result on a benchmark cannot establish that every unusual codebase or editor integration will behave correctly.
For teams considering the experiment, the useful next step is reproducibility. Run the same projects through the implementations with the same machine, configuration and warm-up conditions. Compare diagnostics and generated output as well as elapsed time. A difference in what is being checked can look like a performance improvement without representing equivalent work.
The cost also needs careful interpretation. The reported dollar figure describes API-equivalent token usage for a particular stage of the experiment, not a complete price for delivering and maintaining a production compiler. Earlier attempts, subscriptions, review and future maintenance are different parts of the accounting.
The interesting signal is the scale of the software translation attempted with agents. It creates something concrete that other developers can inspect and test. The remaining questions concern correctness, maintainability and whether the claimed performance advantages hold across independent workloads, not simply whether an agent can produce a large quantity of code.