Inside OpenAI: Unfiltered Lessons on AI, Culture, and Rapid Product Launches

01 About OpenAI

I joined OpenAI in May 2023, just as the company exploded from about a thousand employees to three thousand, and helped move the Codex project from prototype to production.

OpenAI feels less like a single, tightly coordinated team and more like a cluster of many small teams working in parallel without a unified roadmap; research directions emerge from individual curiosity rather than top‑down mandates.

02 Work Culture

Communication happens almost entirely on Slack—email is virtually nonexistent. The rapid expansion created challenges in communication, reporting structures, product release processes, and hiring practices.

Different teams operate at wildly different tempos; there is no single "OpenAI experience" across research, applied, and market teams.

Leadership is highly visible on Slack, and decisions are made quickly, often without extensive planning.

03 Code

OpenAI uses a massive monorepo primarily written in Python, with growing Rust services and some Go for networking. Code style varies widely, from production‑grade Google‑style libraries to ad‑hoc Jupyter notebooks.

Most services are built on FastAPI with Pydantic for validation, but there is no enforced style guide.

All services run on Azure, with Azure Kubernetes Service, CosmosDB, and BlobStore being the most reliable components.

The engineering talent pool is heavily influenced by former Meta employees, bringing a fast‑moving, product‑focused mindset.

04 Other Learnings

OpenAI’s rapid product cycles demand a strong "action bias"—teams can start projects without prior approval, and successful prototypes quickly attract resources.

Security is taken seriously, with dedicated teams focusing on misuse, bias, weaponization, and prompt injection, though many safety efforts remain unpublished.

The company’s culture emphasizes openness: the most advanced models are available via public APIs, and the product aims to benefit everyone, not just enterprise customers.

05 Codex Release Journey

The final three months of my tenure were devoted to launching Codex, a code‑generation agent. From concept to production, the sprint took only seven weeks, involving a tiny core team of engineers, researchers, designers, marketers, and a product manager.

We built a container runtime, optimized code‑repo downloads, fine‑tuned a model for code editing, handled Git operations, created a new UI, and added networking capabilities.

The launch was intense: late‑night coding, early‑morning baby calls, and a final deployment that went live at 8 am, instantly attracting user traffic via the ChatGPT sidebar.

Codex proved especially strong at handling large codebases, navigating project structures, and running multiple tasks in parallel, generating over 630 k public PRs in its first 53 days.

06 Farewell Reflections

Leaving OpenAI taught me three things: develop intuition for model training and limits, collaborate with exceptional talent, and ship truly great products.

The experience reinforced that rapid, consumer‑focused AI development can be both exhilarating and demanding, and that the competition for AGI now centers on OpenAI, Anthropic, and Google.