How NVIDIA Engineers and Researchers Use Codex

NVIDIA uses Codex and GPT-5.5 as core tools for complex engineering tasks and end-to-end machine learning experiments. This combination runs in production on GB200 and GB300 infrastructure, and with longer autonomous sessions it even uncovers bugs and ideas that weren't in the original prompt.

The engineering team took an internal platform from MVP to a production system, and also built an in-house podcast recording app similar to Riverside in just a few hours. Using the Codex desktop app's computer interaction feature, they even automated audio/video recording tests, significantly cutting the time it could otherwise take to procure external software.

The research team has nearly automated the research loop, from exploring papers and organizing hypotheses to writing experiment scripts and running them on remote machines. GPT-5.5 proposed a knowledge graph by connecting clues from papers, and one researcher said their experiment speed increased by 10x. There was even a case of rewriting old Python code in Rust, achieving up to 20X efficiency, and about 40k NVIDIA members now have access to Codex.