Three weeks ago, Calvin French-Owen, an engineer who had participated in the development of one of OpenAI's key products, chose to leave this prominent AI giant. Recently, he published an engaging blog post detailing his experiences working at OpenAI for a year, including the intense efforts to develop a new coding agent called Codex to compete with Cursor and Anthropic's Claude Code.

French-Owen's departure was not due to any "dramatic incident," but rather a desire to return to his role as a startup founder. He was a co-founder of Segment, a customer data startup that was acquired by Twilio for $3.2 billion in 2020. The cultural insights he shared about OpenAI in his blog, while some were expected, also challenged some common misunderstandings about the company.

OpenAI

Challenges of Rapid Expansion and Internal Operations

French-Owen pointed out that during his one year at OpenAI, the company's employee count increased from 1,000 to 3,000. There are solid reasons for this rapid growth: ChatGPT is the fastest-growing consumer product in history, and its competitors are also growing quickly. According to the company's disclosure earlier this year, ChatGPT has over 500 million active users and continues to grow rapidly.

However, this rapid expansion also brought chaos. French-Owen wrote, "When you expand quickly, everything falls apart: how the company communicates, reporting structures, how products are delivered, how people are managed and organized, hiring processes, etc."

Although OpenAI employees still enjoy the autonomy to put their ideas into practice, similar to a small startup, this has also led to redundant work between different teams. "I definitely saw six libraries for queue management or agent loops, for example," he said. In addition, the programming skills within the company vary greatly, ranging from experienced Google engineers who can handle code for a billion users to recent PhD graduates with limited coding experience. This situation, combined with the flexibility of the Python language, has made the central codebase—described by French-Owen as a "backend monolith"—"a bit of a mess." He mentioned issues such as equipment failures or long running times, but also noted that senior engineering managers have recognized these problems and are working to improve them.

Maintaining the 'Startup Spirit' and the Mysterious 'Fishbowl' Culture

It seems OpenAI has not fully realized it has become a tech giant, and it still heavily relies on Slack for operations. French-Owen observed that this is similar to Facebook's early "move fast and break things" culture at Meta. Notably, OpenAI has also recruited many talents from Meta.

He described in detail how his senior team (composed of approximately 8 engineers, 4 researchers, 2 designers, 2 marketing personnel, and 1 product manager) built and launched Codex in just seven weeks, almost without sleep. The product launch process was described as "too amazing." "As soon as you open it, they get users. I've never seen a product achieve such instant growth just by appearing in the sidebar, but that's the power of ChatGPT," he exclaimed.

However, as a company under strict scrutiny, OpenAI has developed a culture of confidentiality to prevent information leaks to the public. At the same time, the company closely monitors social media platform X. If a post goes viral there, OpenAI will see it and may respond. French-Owen quoted a friend's joke saying, "This company runs like Twitter."

The Biggest Misunderstanding About Safety

French-Owen hinted that the biggest misunderstanding about OpenAI is that it does not place enough emphasis on safety. Although many AI safety experts, including former OpenAI employees, have criticized its processes, he emphasized that while some "doomsters" worry about theoretical risks to humanity, the company focuses more on practical safety, such as "hate speech, insults, manipulating political biases, manufacturing bioweapons, self-harm, and immediate injection" issues.