Inside OpenAI: A Former Employee's Personal Perspective on Working at the Company

Calvin French-Owen, ex-member of Technical Staff OpenAI, Linkedin

The departure was not related to drama: on the contrary, the decision was not easy. After the experience of launching your own project, it is difficult to become part of a huge team of 3,000 people again. It was time for a new stage.

Nevertheless, he does not rule out a return — the mission of creating AGI is so large-scale and significant. Codex, the launch of which he witnessed, is only a small part of the technological breakthrough of the decade. This is not the company's position, but a personal view of one of the most exciting organizations of our time.

What is really happening inside OpenAI

The first thing that strikes you about OpenAI is the speed of growth. When he joined, the company had just over 1,000 employees. A year later — already more than 3,000, and he found himself among the most "senior" in terms of tenure. With such scaling, processes begin to fail: communications, management, hiring, product releases. Teams work differently — some are pushing themselves to the limit, others are stable, and a single "OpenAI experience" simply does not exist.

The uniqueness of the culture is that everything happens in Slack. There is almost no email. Such a flow is distracting if you don't know how to filter channels. But with good configuration — it works.

In the research department, there is a bottom-up culture: there is no rigid planning, progress appears iteratively, and good ideas can come from anywhere. This creates a meritocratic environment where results are valued, not status. Historically, leaders were those who brought ideas to implementation — even if they did not shine at meetings or in office politics.

The "do it yourself" principle is common in OpenAI. Often, teams independently launched similar projects, such as with Codex or ChatGPT Connectors. It could all start with a couple of people — and in fact, it became a full-fledged direction.

Researchers are perceived as mini-leaders. The main thing is to engage them with the task. Boring or solved topics do not generate interest. Therefore, good managers in research are rare and of enormous value. They know how to connect individual ideas and launch large-scale initiatives.

How work is organized at OpenAI

OpenAI is flexibility, scale, and ambition. Teams work differently, but what unites them is a "bottom-up" culture — ideas are valued, not status. A researcher can start a project on their own, and if it works, a team forms around it. Managers are more likely to help by not interfering than by controlling.

The main work tool is Slack. There is almost no email, access is strictly restricted. 

OpenAI is not a monolith. For some, it is a laboratory, for others — a product company or a mission. But what unites everyone is a serious approach: the work affects millions of people, and the race for AGI is at stake.

The company focuses on real risks: from abuse and manipulation to bioterrorism. Theoretical safety is also important, but practical threats are a priority. Much of what is done is not published.

Despite the criticism, those who really want to make the world a better place work at OpenAI. This is evident in the approach: models are available to everyone, not just corporations. Even ChatGPT can be used without logging in.

Merc is rare here, but GPUs are a huge expense item. One feature could cost as much as the entire infrastructure in another company. And yet, here they are not afraid to dream big — from API to code agents and hardware.

OpenAI is sensitive to the information field — even a viral tweet is noticed. The atmosphere, ideas, and speed of implementation make it unique.

Code and infrastructure

OpenAI uses a monorepo in Python, APIs are written in FastAPI with Pydantic. There are no strict code standards — style depends on the team. Everything runs on Azure, but only AKS, CosmosDB, and BlobStore are reliable. Due to platform limitations, much is created in-house.

Many engineers come from Meta, which is reflected in the "Meta-like" infrastructure: from internal TAO to identity systems.

The basis of the architecture is ChatGPT — the entire codebase is built around the concepts of messages and dialogues. Codex used a different model, but relied on these same developments.

Decisions are made by the teams themselves, which speeds up launches.

Launching OpenAI Codex

For the last three months, Calvin worked on the launch of OpenAI Codex — one of the key projects at OpenAI. In November 2024, the team set a goal — to create a code agent in 2025. By February, there were powerful internal prototypes, and it became clear: the models were ready.

In seven weeks, a lot was accomplished: the execution environment was assembled, repository loading was optimized, the model was fine-tuned for editing, git operations were implemented, a new interface was created, and the internet was connected. OpenAI Codex became a full-fledged, convenient product — and this was undoubtedly one of the brightest stages in his career.

The success of OpenAI Codex was made possible by a strong team: eight senior engineers, four researchers, two designers, two go-to-market specialists, and one project manager. Each knew exactly what to do, but coordination was still key. On the eve of the launch, five participants sat until 04:00, deploying the monolith, and by 08:00 they were on a live stream — and immediately saw a avalanche of traffic. Simply by adding OpenAI Codex to the ChatGPT sidebar, the product was immediately actively used — this is the power of the platform.

OpenAI Codex was made asynchronous: the user sends a task, the agent works on its own in a separate environment and returns with a pull request. This is a risky bet — models are not yet perfect, but the idea is that in the future, code agents will become full-fledged colleagues.

OpenAI Codex can work with large codebases, run multiple tasks in parallel, and compare results. Just from open data, it generated 630,000 pull requests in 53 days — an average of 78,000 per engineer. For the author, this became the largest project in his life.

Conclusion

Calvin French-Owen

At first, I had doubts about whether I should join OpenAI. I wasn't sure what it was like — to give up freedom, have a boss, become a tiny part of a giant mechanism. I didn't even tell anyone much about getting a job there — in case it didn't work out.

Looking back, there is a thought that this was one of the best decisions of my life. It is hard to imagine a place where one could learn more.

If you are a founder and feel that your startup is not going anywhere, you should either rebuild your strategy and try shooting more often or go work at one of the large AI laboratories.

Now is an incredible time to build. But it is also a unique time. 

Source: https://calv.info/openai-reflections