AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

The Future of Coding with AI - OpenAI Podcast Episode 6

· OpenAI Podcast

Greg Brockman (OpenAI co-founder) and Thibault Sottiaux (Codex engineering lead) discuss the evolution of OpenAI's Codex from its early days of simple code completion in GPT-3 to the launch of GPT-5 Codex. They frame coding as an 'exception' to OpenAI's general AGI focus, requiring dedicated data, metrics, and tooling investment. A central theme is the 'harness'—the integration of the model with tools, environments, and agent loops that turns raw intelligence into a usable coding collaborator, comparable to the body's relationship to the brain. The conversation explores how Codex is deployed across multiple form factors: an async cloud agent accessible via GitHub mentions, a CLI for terminal-based workflows, and an IDE extension. They discuss internal tools like '10x' that were eventually productized, the AGENTS.md convention for giving agents persistent project context, and an internal code review tool that proved so valuable engineers were upset when it broke. GPT-5 Codex is highlighted for its ability to work autonomously on complex refactoring tasks for up to seven hours while remaining fast on simple queries. Looking toward 2030, Brockman predicts a world of material abundance constrained by compute scarcity, with millions of agents supervised by humans across organizations. They emphasize safety, sandboxing, scalable oversight, and the potential for AI to enable formal verification and novel scientific discoveries. On 'learn to code,' both guests enthusiastically endorse it—describing Codex as a powerful learning accelerator—but stress that pairing AI assistance with strong software engineering fundamentals yields the best results. Usage of Codex has reportedly grown more than 10x since launch.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.