Codex Head Tibo: Codex is still a transitional form, with a physical 'reset' button to reset user quotas at any time
Thibault Sottiaux (Tibo), head of OpenAI Codex, stated in an interview that Codex is still just a transitional form, with real competition expanding from individual models to entire systems.
He revealed that about a year before the emergence of ChatGPT, DeepMind had an internal chat product called LMChat, but Google chose to shelve it due to concerns about its impact on search business. Tibo confirmed that he has a physical 'reset' button that can reset user quotas at any time.
He believes that OpenAI's biggest difference lies in the close connection between research and product teams, and a strong culture of 'release first, iterate based on user feedback.' Large companies are often disrupted not because they fail to see new technologies, but because they are unwilling to let new technologies disrupt their old businesses.
Currently, programming agents still require users to maintain skill files, long-term memory, sub-agents, tools, workflows, and parallel tasks, which is highly complex. An ideal AI should automatically understand user identity, goals, daily and team work, proactively identify problems, suggest solutions, and execute tasks.
Driven by events, AI programming tools are evolving from model capabilities to complete agent systems, with funding and development resources flowing towards contextual understanding and proactive execution capabilities. The beneficiaries are the speed of OpenAI's product iterations, while the traditional large companies face internal resistance.
Source: Public Information
ABAB AI Insight
After joining OpenAI from DeepMind, Tibo has led Codex and is nicknamed the 'God of Reset' due to his frequent quota resets. The interview reveals that Google shelved its early chat product for business protection, contrasting with OpenAI's rapid release culture.
In terms of capital strategy, OpenAI tightly couples research with products, shifting resources from single model training to system-level agent integration (merging Codex with ChatGPT). The motivation is to quickly iterate through user feedback, seize developer workflows, and avoid the common internal business protection traps of large companies.
This is similar to the hesitation of early Google and other tech giants towards disruptive technologies; currently, we are in a phase where programming agents are transitioning from tools to autonomous systems.
Essentially, this is a technological replacement. The evolution from single models to full system agents aims to eliminate the friction of users manually managing systems, upgrading AI from passive response to proactive collaboration, and reshaping the interaction between development and knowledge work.
ABAB News · Cognitive Laws
Large companies often shelve products due to old business protection
Releasing first and iterating is better than perfecting before launch
A true agent should proactively understand and execute