An Interview with an xAI Engineer

We called at midnight on a Sunday. The office behind him was lit like noon and full of voices.

Based on an interview with an AI engineer at xAI, who asked not to be named.

Leaving big tech for the underdog

He joined a large US tech company in 2020, early in its AI push, and spent about three years on agents and tool calling, training models and building agent loops. Then xAI called.

What decided it was talent density. His prospective manager was a well known deep learning researcher who built one of the optimizers now used in nearly every large model training run. The offer beat the other labs. There was also that specific energy of a challenger moving fast.

The density turned out to be real. His previous company was full of people from Stanford, MIT and Harvard, yet he says the intensity there didn't come close to what he found in San Francisco.

One surprise was that age carried no weight at all. People much younger than him were leading, and they were genuinely good, which convinced him that leaning on seniority had been a mistake. He sees a broader pattern in it. Senior engineers at strong companies now take non-senior roles at companies closer to the frontier, on the theory that proximity to the work beats the title.

What the hours actually look like

We talked past midnight on a Sunday, his time. The office behind him was lit like noon and full of conversation.

China has 996, nine to nine, six days a week. He says the Valley right now runs harder than that. Weekend work is assumed. Team meetings were on the calendar for 3pm on both Saturday and Sunday.

"Everything except breathing is work time."

San Francisco is expensive enough that losing a job there threatens your ability to live there at all, which compounds the pace. Underneath it sits a belief shared across every lab. Fall behind in AI now and there's no path back.

The weekly meeting

Nearly every team at xAI meets Elon Musk in person once a week. During the stretch when he was focused on xAI, some teams briefed him three times a week.

I asked what he's like up close. Scarier than he looks in the press, was the short version: less consensus-building executive, more set-the-direction-and-push. He didn't frame that as purely bad, since small fast organizations often move better on strong direction than on democratic agreement, and the method did produce two very hard companies.

The message pushed hardest across the org is usefulness. If the AI isn't useful it doesn't matter. Not novelty, not surprise, but something that actually helps in someone's work or life. Anything not connected to that gets cut. That became the judgment criterion for every team.

His technical depth is real. He enjoys engineering conversations and digs into anything that sounds soft. Report that a training run takes a hundred hours and this comes back:

"Why does it have to be a hundred hours? There is an improvement somewhere. Get it to twenty."

There were costs too. AI isn't like ordinary engineering, because research needs slack. Building models, cleaning data and running training all take patience, which was in short supply.

A cofounder who worked closely with him left a line that stuck. Musk cares less about whether a decision is right today and more about eventually. The things that look wrong in the moment have a track record of arriving. The irony is that the cofounder left the company not long after saying it.

Macrohard

The project announced earlier this year had actually started the previous September as xAI's first. The name was a swing at Microsoft, and the goal was to control a computer with AI.

The instruction was specific. Build it on the same model architecture as the self-driving stack. In his view operating a computer was the same problem as driving, only easier, since a mistake behind the wheel can kill someone and a mistake on a screen can't.

The team disagreed. Roads are standardized: follow the lines, stop on red, go on green, and there isn't much planning or reasoning involved. Operating a computer is a different animal, because pressing the button isn't the hard part. Knowing why that button and what follows from it is the hard part. They argued it was a fundamentally different problem.

After several months of work, Macrohard moved over to Tesla.

A lab losing people

xAI's position isn't comfortable. Most cofounders are gone, many lead researchers have left, and departures continue. AI is a field where you can't compete without people.

He gave two reasons the hiring still works. The intensity doesn't register from the outside no matter how you describe it. And working face to face with someone like that is rare enough to be worth something on its own.

Join Bloom

Bloom builds offline rooms where people and technology meet. We run them in Seoul, and now beyond it.

• 💬 Discord community, where events are announced first
• ▶️ YouTube, full talks from past events
• 📸 Instagram, photos from the floor
• 🎟️ Upcoming events