An article on Hacker News discusses whether Large Language Models (LLMs) should function as compilers. The author argues against this, emphasizing the inherent challenges in relying on LLMs for code generation. The core issue revolves around the underspecification of requirements when using natural language as a programming interface.

The central argument is that specifying systems is inherently difficult. While LLMs excel at generating code from prompts, they lack the precision of traditional compilers. This can lead to unpredictable behavior and make it hard to guarantee the correctness of the generated code. The author draws parallels to the abstraction layers in existing programming languages.

Developing reliable software requires precise semantics and rigorous testing, which are difficult to achieve with LLMs. The article cautions against outsourcing functional precision to a generator and highlights the risks of relying on vague prompts. The potential for unexpected outcomes and the erosion of control are significant concerns.

Ultimately, the author believes that LLMs, while promising, aren't ready to replace compilers. The lack of precise semantics in natural language and the resulting underspecification pose significant challenges for software engineering. The future likely involves hybrid approaches, where LLMs assist but don't fully control the compilation process.