OpenAI Co-founder Claims Codex Has Great Capabilities
Greg Brockman stated that whenever tasks are not handled using Codex, it is usually due to a lack of context, the need to write skills, or not thinking to call it, and rarely because the task exceeds the model's capabilities.
He emphasized that the current AI capability overhang is significant, suggesting that the model's potential has yet to be fully tapped, and developers should focus on addressing context management and skill-building issues to unleash more productivity.
This viewpoint is rapidly spreading in the AI engineering community, prompting more teams to prioritize investing in workflow tools rather than merely pursuing larger models, accelerating the transition from model capabilities to practical implementation.
Source: Public Information
ABAB AI Insight
Greg Brockman, as OpenAI's co-founder and former CTO, has long advocated for the development of coding tools like Codex. He has previously shared insights on agent skill systems and context management practices in public forums. This viewpoint continues his focus on engineering LLM workflows, with similar observations on capability overflow noted in earlier discussions regarding Anthropic Claude Code.
On the capital front, OpenAI is converting model capabilities into developer subscriptions and enterprise payments through products like Codex, while encouraging skill building and context engineering, creating a positive feedback loop of user-generated assets. This guides industry capital from a pure computational power race towards investments in toolchains and workflow infrastructure.
Similar to the iterative frameworks of Anthropic Skills folders and platforms like Cursor; we are currently in a transition phase where LLMs are moving from raw capability release to structured, reusable workflow control.
Essentially, this represents a technological substitution: the powerful model overhang is complemented by skills and context to replace traditional development processes, restructuring the productivity chain, and shifting pricing power from model parameter scale to platforms and developers who master skill assets, persist context, and orchestrate agents.
ABAB News · Cognitive Laws
Model capability overflow is like a mineral deposit: without mining skills and writing context, gold remains buried underground.
Not using tools is not a limitation of capability, but a limitation of habit: reflecting on "why not use it" opens the compounding switch.
Context is the new leverage: filling one gap can amplify model potential tenfold; engineering surpasses parameters.