The ‘Meta AI mafia’ brain drain continues with at least 3 more high-level departures

At least three high-level Meta AI employees have departed the company in the past few weeks to strike out on their own. These include computer scientist and senior director of engineering Erik Meijer, who in an exclusive interview with Fortune said he had declined “stellar offers” from several of the “most sought-out” AI companies.

Meijer, who announced his departure from Meta on X last week, said he is not looking for a new role, but has a startup in the works: He is building an AI assistant platform for enterprises called Automind, and an end-user scripting language called Universalis.

In addition, Devi Parikh, Meta’s senior director of generative AI, also left last week, saying on X she was “nervous, because who in their right mind walks away from the job I had in times like these (leading research efforts in generative media and multimodal LLMs)?!” And research scientist Abhishek Das said working at Meta FAIR (Fundamental AI Research) had been “one of my most cherished career journeys,” but that he is “excited to build something new.” 

Mark Zuckerberg is wooing top AI talent

The race is on in Big Tech to court top AI talent, and Meta CEO Mark Zuckerberg has taken to penning personal emails to Google DeepMind researchers, as well as offering sky-high pay packages, to try to persuade them to come to Meta, according to a recent report from The Information.

But that didn’t stop Meijer and the others from jumping ship: He had previously led Meta’s Probability team, a 50-person research group focused on machine learning infrastructure. The entire team was laid off in November 2022, but Meijer stayed on at Meta working in the organization running PyTorch, an open source machine learning framework that helps speed research prototyping to production deployment.

In explaining his new, as-yet-unnamed startup, he said that knowledge workers have to interact with so many applications specialized for specific problems, which requires mastering different systems and constantly switching between apps. 

“What you want is a single, fully personalized application that can help you solve any problem,” he said. “People have been thinking about this for decades or even centuries. With generative AI, this finally becomes possible.” 

While many startups are looking to build this kind of “intelligent agent,” he said few have the kind of enterprise focus that tackles security and privacy and is reliable and scalable. 

“I often joke that the ‘chat with your PDF in five-minute’ solutions are like the old ‘get fit quick’ ads on the back of comic books,” he said. “There are no shortcuts.” 

Better to build on LLMs outside Big Tech

Meijer said he believes that if you want to build things on top of LLMs, it’s better to do so outside the big companies. Competitive pressure is leading to ever-more-powerful models, longer context windows (which determine how much data you can feed an AI model at one time), and lower prices, he explained. Thousands of developers are now fine-tuning, optimizing, and shrinking models like Meta’s Llama (which he called a “brilliant move” by the company). It all makes for what Meijer called a unique era in the tech industry “where the pace of research and innovation is absolutely breathtaking. His aim, he added, is “to surf that wave as far and hard as I can.”

That said, Meijer emphasized that if one is interested in creating and training new general-purpose models, there is no choice but to join Big Tech, because it costs billions to train LLMs. But “for what I want to do I do not need access to compute,” he explained, since he will simply be using an API. As for doubts about the success of LLM “wrapper” startups, he pointed out that for complex wrappers like what he has in mind, companies like Microsoft, OpenAI, or Anthropic are unlikely to tackle them. 

“I mean, you can say that Facebook is a wrapper on top of MySQL, but that does not imply that it is easy for Oracle to replicate or replace it,” he said, referring to an open-source database management system that is widely used in the industry and which was originally developed by Oracle.

Timing of recent Meta departures might be cyclical

Still, many have wondered about the timing of the recent Meta departures. But Meijer posited it might be a result of the yearly performance review rhythm. “Early in the current AI wave people left Google to found OpenAI, Cohere, and then Anthropic,” he said. “Even in 2022 people were talking about the ‘Meta AI mafia,’ and conversely, lots of people that try startups and fail boomerang back to the Big Tech companies.” 

The ebb and flow of leaving established companies to venture out on your own and returning is “the essence of Silicon Valley,” he noted. “That said, this is probably the best time in tech history to strike out on your own.” 

Anshumali Shrivastava, an associate professor in the department of computer science at Rice University, agreed, telling Fortune: “This is arguably the inflection point in technology with AI,” and that it is hard for established companies to move at a pace innovators want. “The upper management at large companies simply cannot move at a disruptive pace to risk their existing customer base and market,” he explained. 

Still, he said, there is a limited window of opportunity for those willing to jump ship. “For all the great talent in AI, if … they had a dream of being the driving force into the future, this seems like the optimal time,” he added. “Technology is there, the market is bubbling, and VCs have opened the banks for risk takers. Who will not do it?”

But Shrivastava also pointed out that in Silicon Valley, those leaving Big Tech to start companies building solutions to targeted problems may one day end up having their companies acquired by their previous employer—or a close competitor with the same problem. “Note that the Silicon Valley VC ecosystem is run by past executives of successful Big Tech companies, most of whom have seen all this firsthand, and they have a niche in identifying and supporting such opportunities,” he explained.

Arvind Narayanan, professor of computer science at Princeton University, agreed that once the fog settles and it becomes clear which apps are most useful, incumbents often start to dominate again through a combination of building and acquiring. “Every time there’s a new platform or layer of the technology stack, there is an opportunity for startups to build apps on top of it,” he said. “We saw this with the PC, the web, mobile app stores, and now we’re seeing it with generative AI.” With departures from Meta and other Big Tech companies, the action is shifting to a new layer, he explained, “and the cycle repeats.”

Still ‘bullish’ on Meta’s AI efforts

Meijer, however, says he continues to be “bullish” on Meta “as long as Mark is in charge and keeps investing in AI.” 

And even though he and his team were laid off from Meta’s Probability team in 2022, that pivot turned into an AI advantage. 

“Right around the time of the layoffs, ChatGPT came out, and I immediately realized that much of what we aimed for in Probability that was hard or impossible was now within reach,” he explained. “Having no team to support anymore allowed me to jump headfirst into the new world of generative AI.” Much of Probability’s attention was aimed at developer productivity, but after playing with ChatGPT it was obvious that the problem of developer productivity was essentially solved. 

“So I changed my attention to AI-based productivity for general knowledge workers, which is a much larger segment than developers,” he said. “As weird as it sounds, I am actually rather grateful that I was more or less forced to pivot from developing custom models to working on top of foundation models due to the layoffs. And during my tenure at the company I got to support some amazing teams and pioneer the field of AI for productivity when everyone thought that was a bat-crazy idea. So I only look back with fond memories and a sense of pride for what my people accomplished.”