Making the JVM Faster with AI: Promises and Pitfalls

AI is radically changing the art of software development. The most common use cases involve enhanced IDE autocompletion, tools like Claude Code and even the idea of Vibe coding entire applications. This is, of course, targeted at traditional end-user or B2B type applications. To date, AI has not been used much in lower-level systems programming and core infrastructure; in fact, OpenJDK explicitly bans the submission of AI-generated code.

At Azul, we have developed a high-performance JVM, based on OpenJDK, but with significant internal changes to deliver both lower latency and higher throughput. Part of this is replacing the legacy C2 JIT compiler with one based on the open-source LLVM project.

In this session, we’ll explore how we’ve been investigating (and making use of) optimisations identified and developed with AI tools such as Claude. Although this has provided some interesting enhancements and overall speed improvements, we’ll also discuss AI hallucinations claiming improvements that don’t exist.

At the end, you’ll have a clearer picture of how AI can be applied to less conventional areas of software development.