TL;DR:
- Former OpenAI engineer Calvin French-Owen shares his experience working on the Codex coding agent.
- OpenAI grew rapidly from 1,000 to 3,000 employees in one year.
- The company culture is fast-paced but chaotic, with duplicated work and varying coding quality.
- Despite its size, OpenAI still runs largely on Slack, resembling a startup mindset.
- OpenAI focuses heavily on practical AI safety concerns amid high scrutiny.
Rapid Growth and Chaos
Calvin French-Owen, an engineer who helped build OpenAI’s Codex coding agent, recently resigned after a year at the company. French-Owen, co-founder of Segment (acquired by Twilio for $3.2 billion in 2020), shared insights into OpenAI’s culture in a blog post.
During his tenure, OpenAI’s headcount skyrocketed from 1,000 to 3,000 employees, reflecting its rapid scaling amid booming demand for AI products. For context, ChatGPT, OpenAI’s flagship product, crossed 500 million active users as of March 2025. (Source: OpenAI user metrics)
However, rapid growth also created organizational challenges:
- Communication breakdowns
- Ambiguous reporting structures
- Overlapping projects and duplicated efforts
- Variable coding standards across teams
Startup Spirit in a Giant Company
Despite its size, OpenAI retains a startup mentality. Employees are empowered to act autonomously with minimal red tape. French-Owen noted teams frequently build multiple versions of similar infrastructure components, such as queue management libraries.
Coding expertise varies widely—from veteran engineers with Google backgrounds writing scalable code to newer PhDs learning on the job. The core codebase, called the “back-end monolith,” suffers from technical debt and frequent breakages, though leadership is aware and working on improvements.
Remarkably, OpenAI’s workflows still rely heavily on Slack for communication and coordination, contributing to an informal and chaotic atmosphere reminiscent of early Meta/Facebook days. (Meta’s early culture)
The Codex Sprint
One highlight from French-Owen’s account is the intense seven-week push his team made to develop and launch Codex—OpenAI’s AI assistant for programming that competes with tools like Cursor and Anthropic’s Claude Code.
Despite near sleeplessness, the team achieved a seamless launch, with instant user adoption simply by integrating Codex into the ChatGPT sidebar.
“I’ve never seen a product get so much immediate uptake just from appearing in a left-hand sidebar, but that’s the power of ChatGPT,” French-Owen said.
Secrecy and External Scrutiny
OpenAI operates in a highly scrutinized environment, which breeds a culture of secrecy aimed at preventing leaks. The company closely monitors discussions on X (formerly Twitter), often responding to viral posts and public feedback.
“A friend of mine joked, ‘this company runs on Twitter vibes,’” French-Owen observed.
AI Safety: Practical Focus Over Theoretical Fears
French-Owen addressed what he considers the biggest misconception about OpenAI—that the company neglects AI safety.
While some critics emphasize existential risks to humanity, OpenAI’s current efforts focus heavily on practical issues, including:
- Preventing hate speech and abuse
- Mitigating political bias manipulation
- Avoiding facilitation of bio-weapons or self-harm
- Defending against prompt injection attacks
OpenAI remains aware of its broad impact, with hundreds of millions of users relying on its language models for tasks ranging from medical advice to therapy.
The company also watches regulators and competitors closely, recognizing that the stakes are extremely high in the evolving AI landscape.
OpenAI Key Metrics & Culture Insights
| Aspect | Details | Source & Link |
| Employee Growth | 1,000 to 3,000 employees in one year | OpenAI Blog Post – French-Owen |
| ChatGPT Users | 500 million active users as of March 2025 | OpenAI ChatGPT Stats |
| Codex Development | 7-week sprint, integrated into ChatGPT sidebar | OpenAI Codex Launch |
| Communication Tools | Primarily Slack | Meta Early Culture |
| AI Safety Focus | Practical safety: hate speech, bias, prompt injection | French-Owen blog |
