As Mark Zuckerberg sought to play catch-up in the generative AI race, he reached out a few months ago to OpenAI’s former chief technology officer, Mira Murati, and offered to buy her fledgling startup, Thinking Machines Lab.

When she said no, the Meta chief executive responded by launching a full-scale raid. In the following weeks he approached more than a dozen of Murati’s roughly 50 employees to sound them out about jumping ship. His chief target: Andrew Tulloch, a leading researcher and co-founder at the startup.

To peel him off, Zuckerberg dangled a billion-dollar package that could, with top bonuses and extraordinary stock performance, have been worth as much as $1.5 billion over at least six years, according to people familiar with the matter.

Tulloch said no. None of his colleagues left either.

Meta spokesman Andy Stone called the description of the offer “inaccurate and ridiculous” and said that any compensation package is predicated on a stock rising. He added that Meta is not interested in acquiring Thinking Machines.

Even in Silicon Valley, where star engineers have long wielded outsize economic power, turning down nine-figure pay packages is rare. But as the battle royal for AI talent escalates, the companies with the biggest war chests are finding the cash cannon only gets them so far.

While some AI researchers act like free agents, bouncing between labs in pursuit of more pay and power, quite a few display an unwavering allegiance to their chosen leaders, larger-than-life figures who, in the tech industry, carry the single-name cachet of rock stars. The idiosyncratic cultures of the different startups bind employees to one another. Meanwhile, after years of poaching en ronde, the companies involved are getting savvier about playing defense.

OpenAI and startups launched by its alumni like Murati are the frequent targets of Zuckerberg’s recruiting blitz. From the earliest days of the AI race, pioneers lured researchers with the promise of getting to work for the historic mission of creating artificial general intelligence, or a system that is smarter than a human at most tasks. It was OpenAI, whose co-founders include Sam Altman, Elon Musk and Ilya Sutskever, that turned that mission into a quasi-religious quest, with its nonprofit charter to ensure that the work benefits humanity.

Sam Altman, chief executive officer of OpenAI, and Mira Murati, who was OpenAI chief technology officer and left to found Thinking Machines Lab, both favor flat reporting hierarchies.

Meta has reached out to more than 100 of OpenAI’s employees. It has hired at least 10. On July 25, Zuckerberg picked Shengjia Zhao, a Chinese researcher who spent three years at OpenAI, to lead Meta’s new superintelligence team.

The OpenAI researchers who have so far rebuffed Meta’s advances chose to remain because they believed OpenAI was the closest to reaching artificial general intelligence, wanted to work at a smaller company and were wary of having the fruits of their labors go toward a product that was primarily driven by advertising, according to people familiar with the matter.

Zuckerberg has recruited even fewer researchers from Anthropic, the $170 billion startup led by Dario Amodei, who took some of OpenAI’s top talent with him when he left five years ago to start it.

All of Anthropic’s seven co-founders are still at the company. Many of them met Amodei through the tightknit world of effective altruism over a decade ago, a social movement that grew popular among researchers for its early concerns that AI could go rogue and destroy humanity. Some of them lived in a group house in San Francisco, debating the most effective way to give away their wealth and the risk that AI posed.

Zuckerberg has hired at least two employees from the startup—Joel Pobar and Anton Bakhtin—both of whom previously spent multiple years working at Meta.

Ilya Sutskever has taken steps to defend his company’s talent from potential poaching.

Sutskever has built the startup he co-founded last year, Safe Superintelligence, or SSI, in a way that makes it relatively poach-proof. Unlike Amodei, Sutskever didn’t pick off a large group of OpenAI researchers to join him.

Most of Sutskever’s staffers aren’t well-known in Silicon Valley, in part because the company is looking for promising technologists with new ideas whom Sutskever can mentor. They are discouraged from mentioning SSI on their LinkedIn profiles, in part to prevent other companies from trying to snatch them away. Earlier this year, Sutskever rebuffed an offer from Zuckerberg to buy SSI.

Murati, who spent six years at OpenAI before leaving last September, has her own band of apostles. Originally from Albania, she joined OpenAI when it was a small research lab, helped launch its first product and came to run almost every aspect of the company as CTO.

At OpenAI, she was known for her emotional intelligence and lack of ego, which earned her the loyalty of the research and engineering staff. At Thinking Machines, she has instituted a similarly egoless reporting hierarchy as OpenAI and SSI, where even senior researchers are listed as a “Member of Technical Staff” in a nod to the flat, collegial culture of Bell Labs, one of the inspirations behind OpenAI.

When she started Thinking Machines in February, more than 20 OpenAI colleagues followed her, including co-founder John Schulman, one of the key minds behind ChatGPT who had decamped to Anthropic just a few months prior. Many of Murati’s researchers came from OpenAI’s post-training team, the research division that built ChatGPT and was in charge of teaching AI models how to communicate with humans.

Meta has been recruiting AI talent to its offices in Menlo Park, Calif.

What exactly Murati is building remains a closely guarded secret even to some of her investors, who recently poured $2 billion into the company. Its stated mission is making “AI systems more widely understood, customizable and generally capable.” Murati recently said the company was “building multimodal AI that works with how you naturally interact with the world,” and would share its first product in the “next couple months.” The startup has leased an office in one of the quieter parts of San Francisco’s Mission District, just a couple blocks away from OpenAI.

Wired has previously reported some of the details of Meta’s outreach to Thinking Machines’ talent.

Tulloch, listed as a co-founder at Thinking Machines, was the subject of a spirited recruitment campaign from Zuckerberg and Alexandr Wang, the newly appointed head of Meta’s superintelligence lab, both of whom peppered him with messages asking him to join.

Originally from Australia, Tulloch graduated from the University of Sydney, where he had the highest GPA among science students. He worked on machine learning at Facebook for 18 months before starting graduate school at Cambridge. He then moved to California to work at Facebook’s AI Research Group, eventually becoming a distinguished engineer, one of the highest technical roles at the company. “He was definitely known as an extreme genius,” said Mike Vernal, a former Facebook executive who overlapped with Tulloch.

In 2016, a few years after Tulloch joined Facebook, OpenAI’s president Greg Brockman tried to hire him as one of the organization’s first employees, writing in an email to Elon Musk that he was getting $800,000 a year at Facebook and would likely try to negotiate up. At the time, OpenAI was offering new recruits a $175,000 annual salary with a $125,000 annual bonus.

“Andrew is very close to saying yes. However, he’s concerned about taking such a large paycut,” Brockman wrote on Feb. 21.

He didn’t join OpenAI then.

Tulloch eventually came aboard seven years later, when ChatGPT was already a viral sensation and the startup’s valuation soared.