At Elon Musk’s AI startup xAI, some “employees” don’t breathe, sleep, or even exist in the real world.
Instead, they run on code.
The company is quietly testing what it calls “human emulators” — AI systems designed to mimic the behavior of real white-collar workers. According to a former xAI engineer, these AI workers are already operating inside the company, appearing on internal org charts and interacting with staff, sometimes without people realizing they are not human.
The experiment offers a rare look into how far Musk’s vision of automated work may already be going… and where it still falls short.
What xAI means by ‘AI employees’
The details came from Sulaiman Ghori, a former member of xAI’s technical staff, who spoke at length on the Relentless podcast before leaving the company days later.
Ghori said xAI is building digital workers that can do anything a human does on a computer — looking at a screen, using a keyboard and mouse, and making decisions. Internally, these systems are treated like staff.
“Multiple times I’ve gotten a ping saying: ‘Hey, this guy on the org chart reports to you. Is he not in today or something?’” Ghori said on the Relentless podcast. “It’s an AI. It’s a virtual employee.”
He explained that some teams already rely heavily on these agents. In one case, he said a single human engineer was leading a project alongside about 20 AI agents.
The bigger bet: millions of digital workers
Ghori worked on xAI’s Macrohard team — a tongue-in-cheek name Musk has used to describe efforts to automate software creation and challenge traditional enterprise tools from companies like Microsoft.
The long-term plan is ambitious. Ghori said xAI eventually wants to run up to 1 million human emulators at once. To do that cheaply, the company is exploring an unusual idea: using idle Tesla vehicles as a distributed computing network.
“There are 4 million Tesla cars in North America alone,” Ghori said on the podcast. He added that many sit unused most of the day and already have power, cooling, and networking. xAI has suggested paying owners to lease compute time from their cars instead of relying solely on traditional data centers.
Despite the bold vision, Ghori stressed that much of the work today remains hands-on and customized.
Building AI workers for customers is especially difficult, he said, because people often forget to explain all the small, routine steps that make up their jobs. As a result, automating real office work is slower and more complex than it sounds.
Even inside xAI, the human emulator rollout is happening gradually, with plenty of trial and error along the way.
The ‘carnival’ strategy for speed
Speed is the primary currency at xAI, and the company has taken some unorthodox shortcuts to maintain its pace. The massive Colossus data center, for instance, was reportedly stood up in just 122 days.
Ghori explained that this was made possible by leveraging a legal loophole usually reserved for traveling shows. “It was the fastest way to get the permitting through and actually start building things,” Ghori told the Relentless podcast. When the host suggested this technically made xAI a carnival company, Ghori agreed, stating, “It’s a carnival company.”
This “move fast and break things” mentality extends to the hardware itself. To keep the training runs from crashing the local power grid, the company utilizes roughly 80 mobile generators and massive battery packs to balance the load.
The environment is described as a high-stakes engineering playground where a single successful coding “commit” is estimated to be worth millions. “We did the math… right now we’re like I think at about $2.5 million per commit to the main repo,” Ghori noted during the interview.
The high cost of transparency
While Ghori’s insights provided a rare look behind the curtain of Musk’s AI ambitions, his transparency may have come at a personal price.
Just days after the unfiltered podcast interview went live, Ghori announced on X that he had left the company. While neither xAI nor Musk has officially commented on whether the departure was a direct result of the interview, the timing has sparked significant speculation within the tech community.
The culture Ghori described is one of extreme intensity, where engineers sleep in pods or bunk beds and chase “impossible” deadlines.
He recalled a moment when a colleague was offered a Cybertruck as a reward for hitting a 24-hour training goal. “Elon’s like, ‘OK, you can get a Cybertruck tonight if you can get a training run on these GPUs in 24 hours,'” Ghori said. The engineer won the bet, according to Ghori.
The interview remains a uniquely detailed snapshot of a famously secretive company.
Also read: A recent benchmark found that AI agents still struggle with real work tasks.