OpenAI Podcast Episode 6 - The future of coding with AI
7 snips
Sep 15, 2025 
Greg Brockman, co-founder and president of OpenAI, and Thibault Sottiaux, Codex engineering lead at OpenAI, discuss the evolution of coding with AI. They dive into the journey from GPT-3 to the powerful GPT-5 Codex, exploring its capabilities in complex refactoring and coding assistance. The duo reveals insights on building AI harnesses, breakthroughs in code review, and the transformative potential of agentic coding. They also touch on the future landscape of AI in software development and the implications for learning to code in an automated world.
AI Snips
Chapters
Transcript
Episode notes
First Sparks Of AI Coding
- Greg Brockman recalls seeing GPT-3 produce code completions from docstrings and realizing the approach would work.
- He describes dreaming of a model writing a thousand lines of coherent code and that goal has now been surpassed.
Harness Equals Body To The Model's Brain
- The harness (tooling + loop) matters as much as model intelligence for usable coding agents.
- Greg and Thibault stress integrating execution, tools, and environment to make models act usefully.
Optimize For Latency Then Interface
- Prioritize latency for autocomplete experiences: aim for ~1500 ms or users will abandon it.
- When you need deeper intelligence, redesign the harness rather than discarding slower models.