Andrey Breslav built Kotlin at JetBrains and watched it become the default language for Android development. Now he's betting that the next shift in how engineers write software is bigger than a new language — it's replacing source code entirely with plain-text specifications that an LLM compiles into production-ready code.
That's the premise of CodeSpeak, which Breslav is developing in alpha. The pitch isn't aimed at solo developers experimenting with AI-assisted prototyping. It's for professional engineering teams maintaining large, long-lived systems — the kind of teams that spend more time reading and debugging code than writing it. Maintain a spec instead of the implementation, the argument goes, and your codebase shrinks by a factor of five to ten.
Breslav has been testing that claim against real open-source Python projects, and the numbers are hard to argue with. yt-dlp's WebVTT subtitle parser went from 255 lines to 38. Faker's Italian social security number generator dropped from 165 to 21. beautifulsoup4's encoding detection module shrank from 826 lines to 141. Most dramatically, MarkItDown's EML-to-markdown converter collapsed from 139 lines to just 14 — a near 10x reduction. All original tests passed after generation, and new tests were added in each case, pointing to genuine functional equivalence rather than compression that cuts corners.
CodeSpeak is built for incremental adoption. Manually written code and LLM-generated code can coexist in the same project, which removes the all-or-nothing barrier that has killed previous attempts to introduce new languages into existing codebases. The toolchain ships via Python's uv package manager. The project blog has been documenting features like spec dependencies and managed files, suggesting the architecture is moving toward something systematic rather than staying a demo.
The underlying bet is that writing precise, unambiguous specifications is now the harder and more valuable skill, and that LLMs have become reliable enough to handle the translation into working code. If that holds, it reframes what version control is actually storing, what code review is actually checking, and how engineering teams should be organized. Breslav has made large bets on language design before and been right. Whether CodeSpeak can get from alpha to something teams actually ship with is the question the industry will be watching him answer.