OpenAI's Codex product lead Tibo revealed in a podcast that the 'reset button' he uses to restore user quotas is actually a physical object he can press anytime. The article also compiles his insights on OpenAI's culture, the next generation of AI agents, and changes in development workflows.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
OpenAI Codex product lead Tibo confirmed for the first time on a podcast that the "reset button" is a physical object, and delved into OpenAI's bottom-up culture, the shape of next-generation agents, the current state of recursive self-improvement, and the real considerations behind pausing frontier RL training.
On the podcast, Tibo confirmed that Codex's "reset button" is a physical object, "very, very cool," which he can press at any time he deems appropriate. This operation was not originally a marketing ploy but a direct user compensation mechanism: when there are configuration errors, brief outages, or degraded experiences, OpenAI compensates users by restoring credits, rather than just issuing an apology notice.
Over time, the credit reset evolved into a community ritual, also used for celebrations when major features launch. Tibo emphasized that this doesn't require joint planning with marketing or finance; he has the authority to do it directly. However, this practice requires ample compute and long-term capacity planning; OpenAI's upfront investment in compute resources partly supports this generosity.
Tibo describes OpenAI as a highly bottom-up organization where employees can propose ideas, form ad-hoc teams, and quickly build products, with little internal "stopping energy" that blocks new projects. But he stresses that sheer momentum alone would result in a hodgepodge; it must be countered with simplicity, quality, and an overall sense of direction. He believes ChatGPT's iOS app is one of the best apps on the market, and he continues to invest heavily in invisible aspects like delight, performance, efficiency, and simplicity.
His advice to entrepreneurs: have conviction, iterate quickly based on user feedback, and dare to disrupt yourself. The third is the hardest because big companies are often constrained by cash cows. Tibo also revealed that OpenAI itself only discovers what a model can do after training; benchmarks are limited, and teams must repeatedly "play" with the model to uncover new capabilities.
This article's information primarily comes from a systematic compilation by Machine Intelligence of Tibo's podcast interview, making it second-hand reporting. Tibo's statements are source claims and have not been independently confirmed by OpenAI. Specific data mentioned (such as a 60% inference speed improvement and an 80% price drop for Luna) are from the interview and should be treated as source claims.
Primary report
Primary source