What This Is
Simon Willison had Claude help develop a SQLite extension that converts strings such as “-$1,234.56” and “12.5%” into computable numbers. In the first round of results, the Rust extension was roughly 298–321 KB, while the C extension was only 14 KB. Yet Rust was faster in some parsing scenarios, partly because its number-conversion function ran about twice as fast as the equivalent function in C’s standard library.
After further prompting, Claude rewrote the C version and overtook Rust in performance. When processing 2 million rows, the added cost per row fell from 67 nanoseconds to 19 nanoseconds for mixed data, and from 106 nanoseconds to 18 nanoseconds for decimal data. The key lesson is not that C is inherently superior to Rust, but that the model needs explicit benchmarks and repeated testing to find an implementation that genuinely works.
Industry View
We believe this demonstrates that Claude can already complete a continuous sequence of tasks spanning solution design, coding, and reporting. AI-assisted programming is moving beyond code completion toward delivering small engineering projects. For enterprises, model generation speed is no longer the only primary bottleneck; validating code, defining test conditions, and reviewing results are becoming the new priorities.
The risks are equally clear: the first-round builds differed in size by more than twentyfold, and the Rust version showed counterintuitive speed results because of differences in their underlying implementations. This means model-generated comparisons cannot be treated as conclusions. Without shared datasets, performance benchmarks, and human review, AI may present a runnable version as a seemingly complete solution that has not been adequately validated. So-called “autonomous development” still depends on human judgment about quality and cost.
Impact on Regular People
For enterprise IT: Small utilities around databases, format conversion, and data cleaning are now well suited to having models generate initial drafts. Teams should retain testing, permission checks, and deployment approval rather than connecting model-generated code directly to production systems.
For individual careers: Being able to write code no longer means being able to deliver software. Framing problems, breaking down metrics, and validating results are becoming important parts of professional work, with these capabilities extending into process automation and data analysis.
For the consumer market: Lightweight applications for everyday users will appear faster, but they may still conceal compatibility, performance, and maintenance costs. Whether a tool is reliable should be judged by its testing and real-world performance, not merely by the demonstration.