Training Data OpenAI Codex Team: From Coding Autocomplete to Asynchronous Autonomous Agents
325 snips
Jun 10, 2025 
Hanson Wang, a researcher at OpenAI's Codex team, and Alexander Embiricos, product lead focused on agentic coding, delve into revolutionary coding agents that autonomously generate pull requests from simple prompts. They discuss training these AI models for real-world programming, the shift from AI as a pair to an independent agent, and envisioning a future where AI writes most code. Challenges around long-running inference and user task articulation are also explored, all while humorously comparing interactions with Codex to social media engagements.
AI Snips
Chapters
Transcript
Episode notes
Codex Trained for Real-World Coding
- Codex was trained beyond competitive programming to align with professional software engineers' style and expectations.
- This included learning to produce mergeable code, good pull request descriptions, and proper testing practices.
Codex Fixes Last-Minute Bug
- A tough bug blocking launch was fixed at 1 a.m. by running Codex multiple times with the bug description.
- One of four attempts produced a successful fix that landed and included animation for launch.
Try Many Codex Tasks in Parallel
- Use Codex with an abundance mindset: try many tasks in parallel to find successful results.
- Running dozens of tasks in a short time shows good understanding and leads to better outcomes.