Inside Meta’s Secret Plan to Build Superintelligent AI Systems

Every technology company has a point in its history when ambition surpasses the product roadmap and becomes more difficult to identify. Meta seems to have arrived at that point. In the last few months, Mark Zuckerberg has been quietly assembling what could be the most costly and aggressive AI research operation ever created inside a consumer technology company.

He has recruited researchers from OpenAI, Google DeepMind, Anthropic, and GitHub, offering signing bonuses that have reportedly reached $100 million. not pay. bonuses for signing.

It’s worth pondering that figure for a moment. It captures a unique aspect of this specific Silicon Valley moment: a sense of controlled panic among the major players, a sense that the window for positioning in AI is closing, and the belief that whoever secures the right people within the next eighteen months will have a structural advantage that money alone cannot later fix.

In a June television interview, Andrew Bosworth, the head of technology at Meta, stated bluntly: “The market is setting a rate here for a level of talent which is really incredible and kind of unprecedented in my 20-year career as a technology executive.”

Zuckerberg made the official announcement by circulating an internal memo that described the establishment of Meta Superintelligence Labs, or MSL. This memo was swiftly obtained and extensively read throughout the industry. The new division unites Meta’s Fundamental AI Research division, product AI initiatives, and foundation model teams under one roof.

Alexandr Wang, who turned Scale AI into one of the world’s most significant data infrastructure firms, and Nat Friedman, the former CEO of GitHub, who quietly advised Meta for a year before taking on a more official position, will lead it. These ceremonial hires are not safe. These individuals have strong beliefs about what AI ought to develop into and how quickly.

To his credit, Zuckerberg’s declared vision is clear and philosophically different from what his rivals are publicly stating. Automating economic output is not his main interest. His portrayal is intimate: a superintelligence that is familiar with you, comprehends your objectives, and supports you in achieving them. something that will enable you to “be a better friend to those you care about, and grow to become the person you aspire to be.”

Although “personal empowerment” is a more palatable public narrative than “we are building a system that will outthink every human who has ever lived,” there is a sense that Zuckerberg truly believes this—that this is not just messaging.

Apart from the strategic maneuvering, the team Meta has put together merits consideration on its own terms. Shengjia Zhao, who contributed to the development of GPT-4 and ChatGPT, is currently employed at Meta. Hongyu Ren, who co-developed multiple iterations of OpenAI’s most potent reasoning models, is likewise. At DeepMind, Jack Rae oversaw Gemini’s pre-training.

GPT-4.1 and o3 were co-created by Jiahui Yu. These are not mid-level contributors who are being stolen for their resumes. These individuals created the systems that established the boundaries of what the industry believed to be feasible. You wouldn’t be able to tell from the outside that Meta’s Menlo Park campus houses, in a way, a concentrated history of the last five years of AI research.

It’s still unclear if bringing together so much talent in one location results in something better than the sum of its parts or if the social dynamics of bringing together so many determined researchers cause conflict of their own. At this level, developing AI systems is more than just a computational challenge. It necessitates a level of institutional coherence—a common understanding of what you are optimizing for—that is actually difficult to produce rapidly.

Meta has not had an easy journey thus far. The Llama 4 model was criticized for what some researchers called inflated benchmark results; this accusation caused unease in the technical community. The company’s biggest and most anticipated model, the Llama 4 Behemoth, has experienced delays that have sparked internal and external concerns. At this level of investment, these are not small mistakes. These are the kinds of setbacks that would cause a smaller business to seriously reevaluate.

However, Meta is not a smaller business. In 2025 alone, it will spend between $64 billion and $72 billion on AI infrastructure, which is more than twice as much as it did just two years earlier. The data centers under construction are not small-scale improvements. They represent a tangible commitment to a vision of the upcoming decade in steel and power cables. It’s difficult to ignore the confidence and almost impatient nature of that type of capital allocation.

Here, too, the larger competition is important. Google, Anthropic, and OpenAI are all constantly evolving, and each has a different philosophical stance on what superintelligence should do in the end and who should be in charge of it. Zuckerberg’s memo takes a subtle but significant stance against the notion that superintelligence is “directed centrally”—a term that serves as a criticism of specific rivals without naming them. At least formally, Meta believes that personal agency is important and that people should control the tool instead of having it operate from a distance.

It is genuinely unclear if that philosophy will endure interaction with the real technology. If superintelligence develops, it won’t be an app. Without much planning, the industry, governments, regulators, and common people will have to engage in real-time negotiations.

As Zuckerberg himself admitted, the remainder of this decade will probably determine how this develops. Meta has placed its wager. Now, the question is whether the execution and the ambition can come together in due course.