How NVIDIA engineers and researchers build with Codex

NVIDIA engineers are rapidly adopting Codex, powered by GPT-5.5, as a primary tool for complex engineering tasks and end-to-end machine learning experiments. Running on NVIDIA GB200 and GB300 infrastructure, Codex offers extended autonomous sessions, identifying issues and generating ideas beyond initial prompts. This capability has led to significant improvements in scalability and reliability for projects such as evolving internal platforms into production-ready systems.

The NVIDIA coding agents team actively assists engineers in integrating AI tools like Codex into their development workflows. According to senior software engineer Dennis Hannusch, Codex with GPT-5.5 is remarkably autonomous, requiring minimal supervision. It maintains accuracy over long sessions, manages context effectively, and adeptly selects appropriate tools and skills for various tasks.

One notable achievement with Codex includes the rapid development of an internal podcast recording application, similar to Riverside, in mere hours. This was particularly impactful given privacy constraints that would have typically prolonged software procurement for weeks. The Codex desktop app, with its computer interaction capabilities, also facilitated autonomous testing of video and audio recording functions during development.

For NVIDIA's research teams, Codex is transforming the research pipeline. It automates critical steps from identifying research areas and writing scripts for machine learning experiments to running them on remote machines. AI researcher Shaunak Joshi highlights GPT-5.5's role as a creative partner, especially in knowledge-intensive tasks.

Codex functions as a research agent, processing vast corpora of papers in fields like reinforcement learning. Joshi notes its superior creativity and ability to trace evidence and conceptual links, aiding in visualizing complex ideas. With SSH support, Codex allows researchers to effortlessly run large machine learning workloads from their laptops, eliminating concerns about remote host login and setup.

Overall, Codex is accelerating NVIDIA's work across engineering and research, seamlessly integrating idea conception, execution, and testing into a single, efficient workflow. The company anticipates unlocking even greater potential as they continue to build and innovate with this cutting-edge technology.