Mark Zuckerberg Overhauls Meta AI Division Again in 2025: Superintelligence Labs Split, Internal Tensions, Alexandr Wang Leadership, and Strategic Shifts

Key Takeaways

• Zuckerberg created Meta Superintelligence Labs in June 2025, then split it into four groups by August 

• Alexandr Wang leads the new structure as Chief AI Officer after leaving Scale AI 

• Internal tensions plague the division with staff departures and lukewarm Llama 4 reception 

• Four new groups focus on research, products, infrastructure, and AI hardware 

• This marks the fourth major AI reorganization at Meta in six months 

• Billions spent on talent acquisition from OpenAI and other competitors 

• "Behemoth" model development remains secretive under the new structure

---

Article Outline

1. The Birth and Death of Superintelligence Labs

• June 2025 creation and August 2025 split
• Wang's appointment and immediate restructuring

2. Four Corners of the AI Kingdom

• TBD Lab for next-gen research
• Product teams for Meta AI assistant
• Infrastructure group scaling
• FAIR lab long-term research

3. The Talent War Gets Expensive

• OpenAI poaching with Shengjia Zhao hire
• Scale AI exodus following Wang
• Billions in recruitment costs

4. Internal Friction Points

• Staff departures and retention issues
• Llama 4 reception problems
• Competition pressure from rivals

5. Alexandr Wang's Leadership Style

• Scale AI background and approach
• Integration challenges at Meta
• Vision for superintelligence timeline

6. The Behemoth Project Mystery

• Secret model development
• Menlo Park operations
• Competition implications

7. Strategic Implications for Meta

• Market positioning against OpenAI
• Investment justification pressure
• Product pipeline changes

8. What This Means for AI Development

• Industry restructuring trends
• Superintelligence race acceleration
• Consumer impact timeline

---

The Birth and Death of Superintelligence Labs

Mark Zuckerberg created Meta Superintelligence Labs in June 2025, positioning Meta to "deliver superintelligence to the world" through their "strong business that supports building out significantly more compute than smaller labs." The grand vision lasted exactly two months.

Meta is now splitting this newly formed AI group into four distinct teams, the company's latest shake-up in a string of restructurings over the past six months. Employees got whiplash watching their org charts change faster than TikTok trends.

The man tasked with making sense of this chaos? Alexandr Wang, former Scale AI CEO, now leads the lab as Meta's chief AI officer. Wang inherited a billion-dollar science experiment that couldn't decide what it wanted to be when it grew up.

Zuckerberg's memo announcing the original lab read like a manifesto. He talked about "personal superintelligence for everyone" , the kind of AI that would make your phone smarter than your college roommate. The reality proved messier. The latest reorganization follows a series of internal challenges including recent senior staff departures and a lackluster reception for its Llama 4 model.

Wang found himself managing not just artificial intelligence, but the very human problem of keeping brilliant people from walking out the door. The irony wasn't lost on anyone , a superintelligence lab that couldn't predict its own organizational needs.

Meta's approach to AI development resembles a jazz improvisation more than a classical symphony. Players join and leave the stage. The melody changes. The audience wonders what they're hearing. But the show goes on, fueled by billions in venture capital and Zuckerberg's relentless belief that he can code his way to digital deity status.

The timing of these changes reveals something deeper about Silicon Valley's relationship with artificial intelligence. Companies announce grand visions in June, then pivot in August when reality intrudes. Progress happens in fits and starts, not smooth exponential curves that look good in PowerPoint presentations.

Four Corners of the AI Kingdom

The company's Meta Superintelligence Labs will be split into four distinct groups: a new lab tentatively called TBD Lab, a team focused on products like the Meta AI assistant, an infrastructure-focused team, and the Fundamental AI Research (FAIR) lab, which concentrates on long-term AI research.

The "TBD Lab" name tells you everything about Meta's current state of mind. They know they need a research division. They just haven't figured out what to call it yet. This lab will handle the bleeding-edge stuff , the kind of research that might work in five years or might prove to be expensive digital masturbation.

Product teams get the unglamorous job of making AI useful for actual humans. They're the ones trying to convince your grandmother that she needs an AI assistant to help her post cat photos on Facebook. These engineers wake up every day knowing their success gets measured in user engagement metrics, not citations in academic journals.

Infrastructure teams live in the basement, metaphorically speaking. They build the plumbing that keeps AI models running without catching fire. Sources said that one group will focus on AI research, another on products, a third on infrastructure, and the fourth on AI hardware. Hardware teams get even deeper into the weeds, designing chips that can handle the computational equivalent of teaching a computer to think.

FAIR , Fundamental AI Research , represents Meta's attempt at respectability in academic circles. These researchers publish papers, attend conferences, and pretend they're not working for a company that made its fortune selling targeted advertising to people who click on videos of dancing cats.

Each group operates with different timelines, different success metrics, and different ideas about what artificial intelligence should become. The TBD Lab thinks in decades. Product teams think in quarters. Infrastructure teams think in milliseconds. Hardware teams think in manufacturing cycles.

Wang's job involves orchestrating these different rhythms into something resembling coherent progress. He's conducting a symphony where the violin section wants to play jazz, the brass section prefers classical, the drums are doing their own thing, and the audience keeps changing their requests mid-performance.

The split acknowledges what everyone in Silicon Valley knows but rarely admits , building artificial intelligence isn't just a software problem. It requires hardware, infrastructure, products, and research working together. Most companies get good at one or two of these things. Meta wants to master all four simultaneously.

The Talent War Gets Expensive

Shengjia Zhao , formerly of OpenAI , is the new chief scientist at Meta's new Superintelligence Lab, as the company is poaching top AI talent from rivals. The talent acquisition resembles professional sports free agency, except the signing bonuses involve stock options and the players write code instead of throwing footballs.

Meta's recruitment strategy operates on the theory that artificial intelligence breakthroughs come from people, not just algorithms. They're hiring researchers the way collectors buy rare paintings , paying premium prices for proven talent with established reputations. The explicit aim is to build "personal superintelligence for everyone," with aggressive talent acquisition from rivals.

The hiring spree creates a domino effect across Silicon Valley. OpenAI loses a chief scientist. Scale AI loses its founder. Startups lose their best engineers. Everyone moves up one chair in this elaborate game of musical talent, except the music costs billions of dollars and the chairs are equipped with noise-canceling headphones and ergonomic keyboards.

Zhao's move from OpenAI to Meta represents more than a career change , it's a statement about where smart people think the future of AI lives. OpenAI built ChatGPT and captured the world's attention. Meta wants to build something bigger, something that lives inside your pocket and knows what you want before you ask for it.

The financial terms of these hires remain confidential, but industry insiders estimate that top AI researchers now command salaries comparable to professional athletes. The difference is that basketball players retire at 35. AI researchers work until their keyboards wear out or they start their own companies.

Wang understands the economics better than most. He built Scale AI into a billion-dollar business by recognizing that AI needs high-quality training data. Now he's applying similar logic to talent , if you want better artificial intelligence, you need better humans building it.

The recruitment arms race reflects a broader truth about Silicon Valley's relationship with innovation. Companies compete not just on products or market share, but on their ability to attract and retain the smartest people in the room. The smartest people, in turn, gravitate toward companies that give them the biggest budgets and the most interesting problems to solve.

Meta offers both. Their problems are genuinely hard , how do you build AI that understands three billion users across multiple languages and cultures? Their budgets are genuinely massive , Zuckerberg has committed to spending whatever it takes to win the AI race.

Internal Friction Points

This latest reorganization follows a series of internal challenges for Meta's AI efforts, including recent senior staff departures and a lackluster reception for its Llama 4 model, amid internal tensions over the technology. The company's AI division resembles a pressure cooker with a faulty release valve.

Staff departures hit hardest in an industry where individual contributors can make or break entire product lines. When senior researchers leave, they take their knowledge, their contacts, and their understanding of what works and what doesn't. Meta Platforms is restructuring its artificial intelligence division amid internal tensions over the technology.

The Llama 4 model reception problems reveal deeper issues about Meta's approach to AI development. The company invested heavily in open-source AI models, betting that giving away their technology would create competitive advantages through ecosystem effects. The lukewarm reception suggests that free doesn't always mean better, especially when competitors offer more polished experiences.

Internal tensions manifest in meetings where engineers argue about resource allocation, timeline expectations, and strategic priorities. Some employees want to focus on breakthrough research. Others push for products that can ship tomorrow. Still others advocate for infrastructure improvements that won't show results for years.

Wang inherited these tensions along with his new title. His background at Scale AI , a company focused on practical AI applications rather than research breakthroughs , brings a different perspective to Meta's academic-leaning culture. The culture clash plays out in decisions about which projects get funding, which timelines get extended, and which researchers get promoted.

The revolving door of organizational changes doesn't help morale. Employees who joined Meta Superintelligence Labs in June found themselves reassigned to new teams with new managers and new priorities by August. Some adapted. Others polished their LinkedIn profiles and started taking recruiting calls from competitors.

Meta's size creates additional friction points. The company employs thousands of people across dozens of AI-related projects. Coordinating these efforts requires meetings, committees, and processes that slow down individual contributors who prefer writing code to attending status updates. The bureaucracy that enables scale also limits agility.

Competition pressure adds external stress to internal tensions. Every week brings news of breakthroughs from OpenAI, Google, or Anthropic. Meta's researchers read these announcements and wonder if they're falling behind, working on the right problems, or getting the support they need to compete effectively.

Alexandr Wang's Leadership Style

Wang is leading the lab as Meta's chief AI officer after building Scale AI into one of Silicon Valley's most valuable private companies before age 30. His approach to leadership combines the urgency of a startup founder with the systematic thinking of someone who scaled complex technical operations.

At Scale AI, Wang built a business model around the unglamorous but essential work of preparing training data for machine learning models. The company's success came from recognizing that AI breakthroughs depend not just on smart algorithms, but on massive amounts of high-quality, labeled data. This perspective shapes his approach to Meta's AI challenges.

Wang's integration at Meta reveals the cultural differences between running your own company and managing inside someone else's organization. At Scale AI, he made decisions quickly and adjusted course based on market feedback. At Meta, he navigates committee structures, stakeholder management, and the complex politics of a public company with 70,000 employees.

His leadership style emphasizes practical results over theoretical elegance. Wang cares more about AI systems that work reliably than research papers that win academic awards. This focus creates productive tension with Meta's existing research culture, which values scientific rigor and peer recognition alongside commercial applications.

The four-way split of Superintelligence Labs reflects Wang's systematic approach to complex problems. Instead of trying to manage everything under one umbrella, he's creating specialized teams with clear responsibilities and success metrics. Each group can optimize for their specific objectives without getting tangled up in other groups' priorities.

Wang's timeline for superintelligence achievement remains aggressive but undefined. He talks about building AI systems that exceed human performance across multiple domains, but avoids specific dates or benchmarks. This approach protects him from over-promising while maintaining internal momentum toward ambitious goals.

His communication style differs from typical Silicon Valley executives. Wang speaks in specifics rather than abstractions. He discusses technical challenges, resource allocation, and competitive positioning with the precision of someone who understands both the engineering requirements and business implications of AI development.

The Scale AI experience taught Wang lessons about building AI companies that apply directly to his Meta role. He understands that successful AI systems require more than algorithmic breakthroughs , they need operational excellence, quality control, and the ability to scale reliably across different use cases and customer requirements.

The Behemoth Project Mystery

The group reportedly oversees development of the ultra-secret "Behemoth" model, launched in June 2025 and based in Menlo Park, California, with tech figures Alexandr Wang and Nat Friedman leading parts of the initiative. The project name suggests ambitions that extend far beyond incremental improvements to existing AI capabilities.

Details about Behemoth remain closely guarded, even by Silicon Valley standards. Meta employees working on the project sign additional non-disclosure agreements and work in secured facilities with restricted access. The security measures indicate either genuine breakthroughs or elaborate corporate theater designed to generate speculation and recruitment advantages.

Meta's decision puts it at odds with rivals like OpenAI who have taken different approaches to AI model development and deployment. While OpenAI releases models publicly and builds business models around API access, Meta appears to be developing Behemoth for internal use or highly controlled partnerships.

The Menlo Park location provides strategic advantages for a secretive AI project. The facility sits in Meta's backyard, making it easier to transfer researchers and resources between projects. The California location also provides access to the state's deep pool of AI talent and academic partnerships with Stanford and other research universities.

Nat Friedman's involvement adds credibility to the project's technical ambitions. The former GitHub CEO brings experience with developer tools and platforms that could inform how Behemoth gets integrated into Meta's product ecosystem. His presence suggests that this isn't just a research project , it's intended to become part of Meta's commercial offerings.

The secrecy around Behemoth creates challenges for Wang's leadership of the restructured AI division. Team members working on related projects don't know how their work connects to the flagship effort. This information compartmentalization protects intellectual property but limits collaboration and knowledge sharing across research groups.

Industry analysts speculate that Behemoth represents Meta's attempt to leapfrog competitors in specific AI capabilities , possibly multimodal understanding, reasoning, or real-time interaction. The project's timeline suggests that results could emerge within the next 12-18 months, putting pressure on competing firms to accelerate their own development efforts.

The mystery surrounding Behemoth also serves recruitment purposes. Top AI researchers want to work on the most advanced projects available. By creating an aura of cutting-edge development around Behemoth, Meta can attract talent who might otherwise choose positions at OpenAI, Google DeepMind, or Anthropic.

Strategic Implications for Meta

The AI division overhaul reflects Meta's recognition that artificial intelligence will determine the company's competitive position across all its products and services. Facebook, Instagram, WhatsApp, and emerging platforms like Threads all depend on AI algorithms for content recommendation, moderation, and user engagement optimization.

Meta's commitment to developing AI "superintelligence," or systems that can complete tasks as well as or even better than humans represents a bet that general artificial intelligence will arrive sooner than most experts predict. This timeline assumption drives resource allocation decisions and competitive strategy across the company.

The four-team structure positions Meta to compete simultaneously in multiple AI markets. Research teams can pursue breakthrough discoveries. Product teams can improve existing offerings. Infrastructure teams can reduce operational costs. Hardware teams can create competitive advantages through custom silicon design.

Meta's open-source approach to AI models through the Llama family creates strategic advantages that extend beyond immediate commercial returns. By giving away advanced AI technology, Meta builds ecosystems of developers, researchers, and companies that depend on their platforms and contribute improvements back to the community.

The talent acquisition strategy aims to deny competitors access to the best AI researchers while building internal capabilities that can't be easily replicated. Unlike other competitive moats, technical talent can walk out the door and join competitors. Meta's approach involves creating working conditions and research opportunities that make leaving financially and professionally unattractive.

Wang's appointment signals Meta's shift from treating AI as a supporting technology to viewing it as the foundation of all future products. His background building AI-focused businesses brings operational experience that complements Meta's existing research capabilities and engineering infrastructure.

The restructuring timing coincides with increased regulatory scrutiny of large technology companies' AI development efforts. By organizing AI work into clearly defined groups with specific responsibilities, Meta creates clearer accountability structures that can respond to government oversight and public concerns about AI safety.

Meta's strategic position in AI development benefits from the company's massive user base and data resources. Three billion people use Meta's products regularly, generating behavioral data that can inform AI training and validation processes. This data advantage becomes more valuable as AI systems require larger, more diverse training sets.

What This Means for AI Development

The Meta reorganization reflects broader trends in artificial intelligence development that extend far beyond one company's internal structure. The industry is consolidating around a small number of well-funded players who can afford the computational resources and talent necessary to compete at the highest levels.

This marks the fourth internal overhaul at Meta in six months, suggesting that even the largest tech companies struggle to find optimal organizational structures for AI development. The rapid pace of change indicates that traditional corporate management approaches may not work for artificial intelligence projects.

The split into specialized teams represents a maturing approach to AI development that acknowledges the different skills and timelines required for research breakthroughs versus product deployment. This specialization mirrors the evolution of other technology sectors where generalist approaches gave way to focused expertise.

Wang's leadership appointment demonstrates the growing importance of operational experience in AI development. Technical breakthroughs matter, but successful AI companies also need leaders who understand how to scale complex systems, manage talent, and navigate competitive dynamics.

The secretive nature of projects like Behemoth indicates that AI development increasingly resembles national security research more than traditional technology development. Companies guard their innovations with security measures that would be familiar to defense contractors or pharmaceutical researchers.

Consumer impact from these organizational changes will take months or years to become visible. AI improvements typically compound over time rather than appearing as sudden breakthroughs. Users might notice that their Facebook feed becomes more relevant or their Instagram suggestions get more accurate, but they won't directly experience the internal restructuring that enabled these improvements.

The talent war between major AI companies creates opportunities for smaller startups and academic researchers who can offer compelling alternatives to corporate employment. Some of the best AI innovations may come from individuals or teams who choose independence over joining large technology companies.

Meta's approach to AI development will influence how other companies organize their own artificial intelligence efforts. Success or failure of the four-team structure will provide evidence about optimal organizational designs for AI research and development at scale.

The timeline pressure created by competition between Meta, OpenAI, Google, and other major players accelerates overall progress in artificial intelligence capabilities. This competitive dynamic benefits consumers and society through faster innovation, but also creates risks if companies prioritize speed over safety in their development processes.

---

Frequently Asked Questions

Q: What is Meta Superintelligence Labs? A: Meta Superintelligence Labs is Meta's newly created AI division announced in June 2025, designed to deliver superintelligence through the company's substantial computational resources and experience serving billions of users.

Q: Who is Alexandr Wang and why is he important to Meta's AI efforts? A: Alexandr Wang is Meta's Chief AI Officer and former CEO of Scale AI, who now leads Meta Superintelligence Labs. His experience building AI-focused businesses brings operational expertise to Meta's research-heavy AI culture.

Q: Why did Meta split Superintelligence Labs into four groups? A: The latest reorganization follows internal challenges including senior staff departures and lukewarm reception for the Llama 4 model. The split allows specialized teams to focus on research, products, infrastructure, and hardware with clear responsibilities.

Q: What are the four groups in the new structure? A: The four groups are: TBD Lab for next-generation research, a product team for Meta AI assistant, an infrastructure team, and the FAIR lab for long-term AI research.

Q: What is the Behemoth project? A: Behemoth is an ultra-secret model under development at Meta's Menlo Park facility, reportedly led by Alexandr Wang and Nat Friedman. Details remain confidential, but the project appears central to Meta's superintelligence ambitions.

Q: How many times has Meta reorganized its AI division recently? A: This represents the fourth internal overhaul for Meta's AI efforts in six months, indicating the company's struggle to find optimal organizational structures for AI development.

Q: What does "superintelligence" mean in Meta's context? A: Superintelligence refers to AI systems that can complete tasks as well as or better than humans, with Meta's explicit goal being "personal superintelligence for everyone."

Q: Why are there internal tensions at Meta's AI division? A: Internal tensions stem from disagreements over technology direction, staff departures, and competitive pressure, exacerbated by frequent organizational changes and high-stakes competition with rivals like OpenAI.