Originally reported by InfoQ on November 16, 2026. This article is based on an edited account of a talk by Bryan Cantrill, DTrace co-creator and Oxide co-founder.
Rust’s ownership rules shift work toward development and help prevent memory-safety errors. But they cannot tell a developer whether an AI-generated system has become unnecessarily complicated. That was Cantrill’s central warning about using large language models to write software.

Cantrill argued that conveniences such as garbage collection can impose costs on people operating infrastructure software, including unpredictable latency. He qualified that criticism in the talk’s Q&A: garbage collection has legitimate uses, but he considers it a poor fit for the kinds of software he builds.

In his view, Rust makes a different trade-off. Its compiler requires developers to account for ownership and memory relationships before a program runs. Compiler feedback can be especially useful when an LLM is iterating on code, he said. Passing compilation, however, does not establish that a program is well designed—or that it will work correctly.

Cantrill pointed to Gas Town, Steve Yegge’s Go-based software-development tool, as a cautionary example. According to Yegge’s account quoted in the article, the tool worked well with Opus 4.6 and earlier, but with Opus 4.7 it kept revising Gas Town itself instead of settling into the work it was meant to do. Yegge said that behavior was the final blow, while acknowledging other problems. It is his account of the project’s failure, not a demonstrated outcome for AI coding tools generally.

Cantrill’s practical advice was to give LLMs clearly bounded tasks, inspect what they produce and learn to understand the code rather than depend on a model to explain every change. He said LLMs can be particularly useful for answering questions about an existing codebase, where answers can be checked. For Rust users, the distinction remains: compiler checks can help catch certain errors, but deciding when a system is simple enough is still an engineering judgment.

Source: infoq.cn (originally published 2026-11-16 14:28 UTC+8)
