A former OpenAI engineer describes what it’s really like to work there

Three weeks in the past, an engineer named Calvin French-Owen, who labored on one in every of OpenAI’s most promising new merchandise, resigned from the corporate. 

He simply printed an enchanting weblog submit on what it was wish to work there for a 12 months, together with the sleepless dash to construct Codex. That’s OpenAI’s new coding agent that competes with instruments like Cursor and Anthropic’s Claude Code.

French-Owen mentioned he didn’t go away due to any “drama,” however as a result of he desires to get again to being a startup founder. He was a co-founder of buyer knowledge startup Phase, which was purchased by Twilio in 2020 for $3.2 billion. 

A few of what he revealed concerning the OpenAI tradition would shock nobody, however different observations fight some misconceptions concerning the firm. (He couldn’t be instantly reached for remark.)

Quick progress: OpenAI grew from 1,000 to three,000 individuals within the 12 months he was there, he wrote. 

The LLM mannequin maker definitely has causes for such hiring. It’s the fastest-growing shopper product ever, and its rivals are additionally rising quick. In March, it mentioned that ChatGPT had over 500 million energetic customers and climbing rapidly.

Chaos: “Every little thing breaks once you scale that rapidly: the right way to talk as an organization, the reporting buildings, the right way to ship product, the right way to handle and arrange individuals, the hiring processes, and so forth.,” French-Owen wrote.

Like a small startup, individuals there are nonetheless empowered to behave on their concepts with little-to-no crimson tape. However that additionally implies that a number of groups are duplicating efforts. “I have to’ve seen half a dozen libraries for issues like queue administration or agent loops,” he supplied as examples. 

Coding talent varies, too, from seasoned Google engineers who write code that may deal with a billion customers, to newly minted PhDs who don’t. This, coupled with the versatile Python language, implies that the central code repository, aka “the back-end monolith,” is “a little bit of a dumping floor,” he described. 

Stuff often breaks or can take extreme time to run. However prime engineering managers are conscious of this and are engaged on enhancements, he wrote.

“Launching spirit”: OpenAI doesn’t appear to know but that it’s a large firm, proper right down to operating solely on Slack. It really feel very very like move-fast-and-break-things Meta in its early Fb years, he noticed. The corporate can be filled with hires from Meta.

French-Owen described how his senior crew of round eight engineers, 4 researchers, two designers, two go-to-market employees and a product supervisor constructed and launched Codex in solely seven weeks, begin to end, with virtually no sleep.

However launching it was magic. Simply by turning it on, they received customers. “I’ve by no means seen a product get a lot rapid uptick simply from showing in a left-hand sidebar, however that’s the ability of ChatGPT.” 

Secretive fishbowl: ChatGPT is a extremely scrutinized firm. This had led to a tradition of secrecy in an try to clamp down on leaks to the general public. On the similar time, the corporate watches X. If a submit goes viral there, OpenAI will see it and, probably, reply to it. “A pal of mine joked, ‘this firm runs on twitter vibes,’” he wrote.

Greatest false impression: French-Owen implied that the most important false impression about OpenAI is that it isn’t as involved about security correctly. Actually a number of AI security of us, together with former OpenAI staff, have criticized its processes. 

Whereas there are doomsayers worrying about theoretic dangers to humanity, internally there’s extra deal with sensible security like “hate speech, abuse, manipulating political biases, crafting bio-weapons, self-harm, immediate injection,” he wrote. OpenAI isn’t ignoring the long-term potential impacts, he wrote. There are researchers taking a look at them, and it’s conscious that lots of of tens of millions of individuals are utilizing its LLMs immediately for all the pieces from medical recommendation to remedy.

Governments are watching. Rivals are watching (and OpenAI is watching rivals in return). “The stakes really feel actually excessive.”