While everyone proclaims that AI is about to replace office productivity software, the ChatGPT desktop app quietly bundled a 1.7GB runtime environment onto users’ hard drives. Inside that hidden directory lies 429.7MB of LibreOffice—the open-source office suite. Beneath documents supposedly generated by AI, code written two decades ago is doing the heavy lifting.
A 1.7GB Traditional Toolchain on Your Hard Drive
On September 1, 2026, developer Simon Willison discovered a massive directory under ~/.cache/codex-runtimes on macOS. Packed inside this 1.7GB environment were Python and Node.js, bundled alongside git and the PDF rendering tool Poppler. Deep within this runtime sat a 429.7MB headless build of LibreOffice.
| Component | Size | Purpose |
|---|---|---|
| Node.js | 446.4MB | JavaScript runtime |
| Python | 440.6MB | Scripting execution environment |
| LibreOffice (Headless) | 429.7MB | Office document format conversion |
| Poppler | 187.9MB | PDF rendering |
| git | 148.1MB | Version control |
To generate documents locally, ChatGPT simply shipped an entire traditional software toolchain.
Image: Simon Willison’s disk usage analysis screenshot showing components inside codex-runtimes. Source: Simon Willison
AI Can’t Fabricate docx Files Out of Thin Air
AI models excel at generating plain text, but struggle when dealing directly with complex corporate document formats. Formats like docx and xlsx are essentially compressed zip archives embedding stylesheets and metadata. Forcing an LLM to accurately stream binary content formatted to strict layout specifications consumes immense compute and yields high output failure rates.
OpenAI engineers opted for a pragmatic route. Inside the plugins/documents directory are skill scripts instructing the AI on how to invoke these binary CLI utilities. The model only needs to produce structured text, leaving the dirty work of format conversion silently to LibreOffice in the background.
Image: LibreOffice 25 interface screenshot on Debian. Source: Wikipedia
From Alternative to Symbiotic Host
LibreOffice, originally forked from OpenOffice.org, spent the past decade as a free open-source alternative to Microsoft Office. Today, it has become foundational infrastructure for AI. At the top layer, massive neural networks run inference; at the bottom, decades-old C++ code performs document layout rendering.
There is an unavoidable irony here: while AI was expected to render office software obsolete, AI applications ended up bundling office software themselves. Word document formatting is far more intricate than it appears—a styled docx file is a zipped bundle of XML schemas and assets where any minor discrepancy corrupts the document. Asking an LLM to emit raw binary structures directly leads to high failure rates, making LibreOffice the ideal conversion engine.
The evolution of AI apps hasn’t discarded legacy ecosystems. Combining large language models for intent understanding with mature open-source toolchains for output generation represents the most reliable engineering pattern today. Trading a 1.7GB local disk footprint delivers robust backward compatibility across legacy document formats. The discovery sparked 199 points and over 100 comments on Hacker News, with developers showing keen interest in AI embedding classic office suites.
Cutting-Edge Applications Built on Open-Source Foundations
ChatGPT bundling LibreOffice exposes the reality of AI engineering: cutting-edge AI progress directly feeds on open-source projects refined over decades of bug fixes. Without relying on this heavy open-source toolchain, the AI assistant claiming to reshape document productivity couldn’t even output a properly formatted office file.
Reference Links:
- Simon Willison’s Blog
- HN Discussion
