Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

Akshay Nathan discusses his journey from working with no-code primitives like Airtable to understand the potential of bringing complex technology to everyone through AI. Since joining OpenAI in 2023, he highlights that despite rapid product development, the company's core mission of democratizing frontier intelligence and building AGI has remained consistent. The internal culture of bottoms-up innovation continues to thrive, fostering ambitious projects that aim to make AI accessible and useful across all domains.

Akshay reflects on lessons learned from deploying AI in enterprise settings, emphasizing that there’s no one-size-fits-all solution. Different teams have vastly different use cases, and success involves meeting users where they are, teaching them how to leverage AI effectively. Although adoption has surged, many still see AI as a black box, and the challenge lies in making its workings more transparent and usable within various user environments.

The episode explores the strategic decision to unify ChatGPT and Codex into a shared agent harness, facilitating a seamless experience whether users are coding, managing projects, or doing knowledge work. This unified approach aims to leverage the power of AI across different workflows, reducing the need for users to navigate multiple products or settings. Akshay explains that the underlying harness is shared, with UX differences to cater to specific use cases.

Discussions focus on designing simple yet capable interfaces that expose enough of the AI's power without overwhelming users. Features like advanced sliders for model control, visibility into sub-agents for efficiency and cost management, and the importance of default model configurations are highlighted. Akshay advises users to experiment with default setups, try different configurations as needed, and focus on building quality 'at-bats' or iterations to drive progress.

Akshay talks about how tools like Sites are revolutionizing knowledge artifacts, enabling teams to create and share living documents that surpass traditional decks and spreadsheets in flexibility and scalability. These interactive sites facilitate collaborative work, research, and experimentation—examples include automating research dashboards, training AI on physical games, and creating detailed, functional reports.

The episode covers advancements in AI memory systems, including ChatGPT's memory V3 and experimental tools like Chronicle that learn from user activity across devices. These systems improve context retention, allowing AI to surfacing relevant past information proactively. Akshay also discusses how auto research and benchmarking tools tie into this, streamlining complex tasks like training game AIs or hyperparameter tuning.

Akshay reflects on how AI is transforming product development and team dynamics. He emphasizes that measuring productivity will shift away from proxies like code commits, toward 'quality at-bats'—the ability to generate, build, receive feedback, and validate ideas efficiently. The conversation underscores how AI enables individuals and small teams to achieve more with less, blurring traditional role boundaries and fostering broader generalist capabilities.

The discussion also touches on practical tips for power users, including experimenting with model settings, using sub-agents strategically, and the importance of context and feedback. Akshay warns against overloading systems with sub-agents, emphasizing deliberate configurations and default settings for most users. The importance of privacy and security, especially when connecting personal data sources, is also acknowledged, highlighting that managing permissions and protecting personal info remains a critical challenge.