Episode Summary
Executive Summary: Thibaut traced Codex from internal research tooling into a product: why OpenAI built it in Rust and open source, how harnesses now compensate for model limits, why model/harness separation matters, and how Codex merged into ChatGPT. He argued that AI is collapsing maintenance and re-architecture costs, changing code review, SDLC, and team workflows toward faster, more automated, agent-assisted shipping.
Main Topics: Origins of Thibaut’s engineering path (Priority: 3/5): He described how growing up in rural Belgium led him into computers, then into applied math, optimization, startups, Google, DeepMind, and finally OpenAI—consistently gravitating toward building tools that help others work faster. Why Codex was built in Rust and as a separate agent core (Priority: 5/5): Codex was intentionally split into a robust, secure agent core and product layer. Rust was chosen for correctness, performance, compile-time guarantees, and future scalability, even though other languages could have worked initially. Open source and model optionality (Priority: 5/5): The team made the CLI, SDK, and app server open source to build in public, invite contributions, avoid forks, and stay connected to the broader coding ecosystem. They also allowed non-OpenAI models to run in Codex to preserve user choice and competition. Harness vs. model responsibilities (Priority: 5/5): Thibaut framed the harness as a set of guardrails and crutches that make the model reliable, controllable, and efficient. Over time, model improvements reduce the need for those harness interventions, shrinking developer messages and special logic. Code review, maintenance, and re-architecture are being automated (Priority: 5/5): He argued that AI models are becoming superhuman at code review, security checks, dependency reasoning, and maintenance tasks. This shifts humans toward intent, architecture, and product judgment rather than line-by-line verification. Merge of Codex into ChatGPT and future of cloud execution (Priority: 4/5): The merge was hard because Codex began as local-first while ChatGPT is cloud-managed. The team unified stacks, moved toward cloud execution, and sees a future where agents orchestrate setups, run long tasks, and free users from keeping laptops open. How to succeed as an AI engineer (Priority: 4/5): His advice: cultivate deep curiosity, learn systems quickly, ask good questions, and stay close to the community or user base. Taste and clarity of intent matter as much as technical skill.
Key Arguments: Codex succeeded because the team separated the agent core from the product surface, enabling correctness, security, and future rewrites without coupling everything together. Rust was chosen not because the team expected the model to stay weak at Rust forever, but because compile-time guarantees and performance mattered for an agent that could scale massively. Open source improved onboarding, community trust, and iteration speed, even though it also created copycats, noisy contributions, and repo-boundary friction. Allowing other models in Codex was a principled choice: if the harness is the best part, it should not be locked to one model provider, and optionality benefits users and OpenAI alike. The harness is always a step ahead of the model: it supplies the missing crutches, guardrails, and developer instructions until the model learns the behavior directly. AI code review is already strong enough to block risky OpenAI pull requests automatically, especially for security issues, reducing human burden on routine verification. Maintenance and re-architecture are becoming cheaper because agents can automate dependency upgrades, refactors, and verification that used to be deferred due to effort. The merge of Codex into ChatGPT required unifying local and cloud execution models, but it enabled broader distribution and a single product experience. Future developer workflows will increasingly center on high-level intent, architecture, and invariants, while agents handle more of the mechanical implementation and review work. The best engineers in this new era will be those who can rapidly understand systems and remain aligned with the real needs of users and teams.
Data Points: OpenAI ChatGPT team size mentioned: 20 people - Thibaut cited this as evidence of how small and empowering the ChatGPT effort felt when he was deciding to join OpenAI. Timeline at the canceled Google project: Roughly 2 years - He worked on a web-speed/mobile initiative at Google before it was canceled. Time at Google Maps before DeepMind: Roughly 1 year - He moved from the canceled project to Google Maps, then later to DeepMind. Codex local-first runtime: More than a year - He said Codex has run entirely on the local machine by default for over a year, before cloud options expanded. OpenAI pull request policy: Automatic blocking for security issues - AI-assisted code review now blocks OpenAI PRs when a security vulnerability is flagged. ChatGPT Work / cloud machine scale: Runs in a managed VM / CADA container - He described the cloud execution environment used by ChatGPT Work and the merged Codex experience. ChatGPT user scale referenced: A billion active users and growing - Used to emphasize that shipping code into ChatGPT affects a very large audience. Codex user scale referenced: 20 million active users - He referenced a milestone for Codex usage after the ChatGPT merge. Cloud push throughput cited for Entire sponsor: 418 pushes per second - Mentioned in the sponsorship read, not part of the interview content. Claimed relative speed vs competitors for Entire: Up to 89x faster - Mentioned in the sponsorship read, not part of the interview content.
Pivotal Quotes: "The harness is always a step ahead of the model." — Tebo: He used this to explain how Codex compensates for model limitations with guardrails, tooling, and developer messages. "If you have that, you’re obviously going to point it at itself." — Tebo: He explained why Codex was made open source: a coding agent should be able to improve itself and be part of the community. "What happens inside the box is, you know, it could be literally anything." — Tebo: He argued that teams should focus on invariants and interfaces, while AI handles more of the internal implementation details.
Implications: AI coding tools are moving from helper utilities to agentic systems that reshape reviews, maintenance, and architecture. Teams that design for modularity, optionality, and user trust will adapt fastest as agents take over more routine engineering work.