Former OpenAI engineer reveals how 17 people built Codex in 7 weeks

Ever wonder what it is actually like inside the most hyped AI company on the planet?

Calvin French-Owen, who sold Segment to Twilio for $3.2 billion back in 2020, just left OpenAI after a one-year stint. He resigned three weeks ago not out of workplace drama, but to build another startup. He wrote a candid tell-all about the reality behind the curtain.

Fast scale, duplicate code, and a backend dumping ground

During his single year at OpenAI, the headcount skyrocketed from 1,000 to roughly 3,000 employees. With ChatGPT topping 500 million active users in March, the hiring rush was unavoidable, but the internal plumbing struggled to keep up.

Communication broke down, reporting lines shifted, and red tape remained nonexistent. While that gave small teams massive autonomy, it also triggered chaos. French-Owen noted seeing at least half a dozen duplicate libraries built for simple things like queue management and agent loops.

Code quality swings wildly across the floor. You have veteran Google engineers who can design infra for a billion people working alongside freshly minted PhDs with little production discipline. Combined with Python's flexibility, the central backend monolith effectively became a bit of a dumping ground where things break constantly.

Seven sleepless weeks and Twitter vibes

Despite its valuation and size, OpenAI still operates like an early Facebook startup packed with Meta alumni. The company basically runs on Slack and an intense move-fast ethos.

Case in point: French-Owen was part of the crew that built Codex, their new coding agent taking on Cursor and Claude Code. A core squad of eight engineers, four researchers, two designers, two go-to-market staff, and one PM shipped it in just seven weeks on almost zero sleep.

Once it was live, the ChatGPT effect took over. French-Owen admitted he had never seen a product explode in adoption simply by being placed in a left-hand sidebar.

Culturally, OpenAI is a secretive fishbowl terrified of leaks, yet obsessed with public perception. Leadership constantly watches X to see what is trending, prompting one coworker to joke that the whole company runs on Twitter vibes. As for safety, he pushed back against critics: while theoretical existential doom gets less traction internally, teams obsess over practical safeguards against prompt injections, bioweapons, hate speech, and political manipulation.

Why it matters

This peek behind the curtain strips away the mystique. OpenAI is not an infallible machine operating years ahead of everyone else. It is dealing with standard hypergrowth growing pains: brutal technical debt, redundant engineering, and sleep-deprived sprints.

For startup founders, the real takeaway is distribution power. A team of seventeen people hacked together a direct Cursor competitor in under two months, and the moment it hit the ChatGPT sidebar, it printed instant scale. Competing against that level of default real estate is brutal.

We think OpenAI's pirate-ship culture and move-fast mentality are precisely what let it outrun legacy tech giants so far. But running a mission-critical platform serving hundreds of millions on an unruly Python monolith and Slack vibes is playing with fire. Sooner or later, the company will have to slow down and pay off its staggering engineering debt.