LLMs Will Not Replace AI Compilers. They Will Call Them. (aicompilers.github.io)

A recent discussion highlighted the belief that large language models (LLMs) will replace AI compilers in the compilation process. However, the article argues that LLMs are more likely to serve as orchestration tools that enhance existing compilation methods rather than as direct replacements. The inefficiency and complexity inherent in having an LLM perform compilation tasks—where performance optimizations and code generation could consume significant resources—underscores the necessity of established compilers, which rely on robust algorithms developed over decades. Significantly, LLMs can improve compiler efficiency by guiding the compilation process, making informed decisions based on insights gained from previous compilations. They can automate low-level code generation, like writing kernels for new hardware or specific models, yet this capability necessitates the existing compiler infrastructure for correct execution and performance validation. Thus, while LLMs won't render AI compilers obsolete, their integration promises to reshape the role of compiler engineers, shifting the focus to system design and verification processes, ultimately enriching the compiler landscape with greater rigor and adaptability.