Thibault
Head of the OpenAI Codex team, with extensive engineering experience in scaling prototype systems to data-center-level infrastructure.
Timeline (last 90 days)
Thibault believes software re-architecting has become extremely inexpensive, with architectural refactoring that might previously have taken years now being dramatically accelerated.
Codex team lead Thibault previously worked at Google Maps and DeepMind before joining OpenAI.
Thibault believes code review has historically held three kinds of value: correctness, security, and social attributes; the first two are being automated, leaving intent as the remaining aspect requiring human discussion.
When a requirement arises, the Codex team first determines whether it calls for a change to the Harness or a change to the model, and if it is a model change, they assess whether it can be achieved within one month, three months, or half a year.
Code review has historically had three-fold value: correctness, security, and the ritual of information exchange; the first two are being automated, and the remaining core is discussion around intent.
The Codex team treated the 'product interface' and the 'Agent core' as two separate parts from the very beginning.
The complexity of the Harness is an inverse function of model capability — the stronger the model, the thinner the Harness and the shorter the developer message.
OpenAI's Thibault publicly stated that Codex's reliability is approaching 100%, that its technology stack may be partially open-sourced, and claimed that OpenAI will own Astra.
All Facts (7)
Codex team lead Thibault previously worked at Google Maps and DeepMind before joining OpenAI.
50%UnverifiedThibault believes code review has historically held three kinds of value: correctness, security, and social attributes; the first two are being automated, leaving intent as the remaining aspect requiring human discussion.
50%UnverifiedThibault believes software re-architecting has become extremely inexpensive, with architectural refactoring that might previously have taken years now being dramatically accelerated.
50%UnverifiedThe Codex team treated the 'product interface' and the 'Agent core' as two separate parts from the very beginning.
50%UnverifiedCode review has historically had three-fold value: correctness, security, and the ritual of information exchange; the first two are being automated, and the remaining core is discussion around intent.
50%UnverifiedThe complexity of the Harness is an inverse function of model capability — the stronger the model, the thinner the Harness and the shorter the developer message.
50%UnverifiedWhen a requirement arises, the Codex team first determines whether it calls for a change to the Harness or a change to the model, and if it is a model change, they assess whether it can be achieved within one month, three months, or half a year.
50%