Introduction

Since the release of GPT-5 in August, Codex has demonstrated remarkable growth, with a 20-fold increase in users and processing trillions of tokens weekly, becoming OpenAI’s most popular programming assistant.

In a recent podcast, Alexander Embiricos, the product lead for Codex, revealed that the rapid growth of Codex is not solely due to model improvements but also because of a deeper understanding that a true intelligent agent is the result of the collaboration between the model, APIs, and frameworks.

Breakthroughs in Long-Term Task Capability

For Codex to handle long-duration tasks spanning hours or even days, the team designed a mechanism called “compression”—the model extracts key information, the API manages task chains, and the framework ensures stable operation. This three-layered approach allows Codex to tackle long programming tasks that traditional large models struggle with.

Real-World Success Stories

Andrej Karpathy shared that he was stuck on a bug for hours until he handed it over to Codex, which resolved it within an hour. The Sora team leveraged Codex to launch an Android app from scratch in just 28 days, achieving the top spot in the App Store.

Evolution of Codex

Reflecting on its journey, Embiricos noted that Codex’s path was not always clear. Initially, Codex was too advanced, using remote asynchronous interactions that were friendly to experienced engineers but challenging for most. The turning point came when the team moved Codex from the cloud to local environments, allowing it to work directly within engineers’ IDEs.

Currently, Embiricos describes Codex as a “smart but passive intern, writing code quickly.” Codex continuously self-monitors and trains, evolving over time. He hopes that in the future, Codex will participate fully in the software development process, becoming a valuable teammate for engineers.

OpenAI’s Unique Culture

Embircious also discussed OpenAI’s organizational culture, expressing admiration for its speed and ambition. He noted that unlike other organizations that “aim before they shoot,” OpenAI’s approach is to “shoot first, then aim,” releasing products and optimizing based on real user feedback. This, combined with a culture that attracts top talent, enables rapid iteration.

The Future of AGI

Regarding the timeline for AGI, Embiricos provided an interesting perspective, suggesting that the current limitations are not due to model capabilities but rather human factors—our limited input and review speeds are hindering progress. He predicts that the first wave of users experiencing a steep increase in productivity will emerge next year, leading to accelerated changes.

Key Insights from the Podcast

1. With Codex’s help, OpenAI developed the Sora Android app in just a few weeks with only a couple of engineers, achieving a top App Store ranking.
2. Even if AI models stop improving, it will take years to fully develop products that leverage their potential.
3. The key to utilizing Codex effectively is to tackle the toughest problems rather than the simplest ones.
4. The initial version of Codex was too advanced, but its growth exploded once it was integrated into engineers’ daily workflows.
5. Writing code may become the universal way for AI to accomplish tasks, suggesting all AI assistants should have coding capabilities.
6. OpenAI’s designers now write and release their own code, maintaining a fully functional prototype built with AI assistance.
7. The biggest bottleneck in AI productivity is not the AI itself but human typing speed and review processes.
8. The joy of coding is being replaced by the need to review AI-generated code.
9. New AI models can now work continuously for 24 to 60 hours on single tasks, thanks to the compression technique.
10. Understanding specific customers is more important than product development skills for starting a new company.

OpenAI’s Speed and Culture

Embircious shared insights from his experience at OpenAI, emphasizing the pace and ambition of the organization. He noted that the rapid growth of models like Codex is unprecedented, leading him to set high goals for product development.

Codex’s Position and Vision

As the head of Codex, Embiricos explained that Codex is an IDE extension for VS Code, designed to assist software engineers throughout the development lifecycle. He envisions Codex as a proactive team member, not just a coding tool.

Technical Breakthroughs and Growth Drivers

Codex has experienced explosive growth since its launch, with usage increasing over tenfold since August. The integration of product and research teams has allowed for rapid experimentation and understanding of how models and tools work together.

Future Directions for AI Agents

Embircious believes that future AI agents will transition from passive tools to proactive teammates, automatically understanding context and participating in workflows without needing explicit commands.

Conclusion

As AI continues to evolve, the role of tools like Codex will expand, fundamentally changing the landscape of software development and productivity. The future will see AI becoming an integral part of teams, enhancing creativity and execution capabilities across various domains.