A former employee of Meta’s AI division shared his first-hand experience in developing the Llama large model with Silicon Valley Observer, uncovering the story behind the massive layoffs at Meta AI.

In his view, it was the open-source large model Llama 4’s obvious lag behind Chinese competitors like DeepSeek that created a strong sense of crisis for Mark Zuckerberg. This prompted Zuckerberg to make a determined move—sparing no expense to poach AI leaders and elite talents from outside the company—to carry out a complete restructuring of Meta AI’s business and strategy.

Meta Lays Off Staff in Its AI Division

Competition in the AI industry has reached a fever pitch, with talent becoming the most sought-after asset as companies vie for dominance. Tech giants such as Google and Microsoft, in order to concentrate resources and compete in AI R&D, have continued to expand their AI research teams while even resorting to large-scale layoffs in other non-core departments.

However, Meta took a contradictory approach: while eagerly recruiting talent and offering sky-high salaries to poach from competitors, it swung the axe at its own AI division, laying off as many as 600 employees—including well-known senior researchers in the industry. This left many observers confused and surprised, but also presented a golden opportunity for other AI companies to recruit talent.

Last Wednesday, Meta announced a restructuring of its artificial intelligence division, cutting approximately 600 jobs. The news was delivered by Alexandr Wang, Meta’s Chief AI Officer and head of the Super Intelligence Lab, and was framed as an internal departmental adjustment.

In a memo sent to employees, Alexandr Wang wrote: “By reducing the size of the team, there will be fewer conversations needed to make decisions, and each person will take on more responsibilities with greater scope and influence.” Since joining Meta, he has been committed to streamlining the bloated organizational structure and workforce.

After the layoffs, the total number of employees in Meta AI dropped to fewer than 3,000. Affected employees were informed that their last day would be November 21, and until then, they would be in a “non-working notice period” during which their internal access rights would be revoked.

In compliance with California’s layoff regulations, laid-off employees will retain their contracts and salaries for two months, with specific compensation to be determined separately. Some employees holding H-1B work visas will need to find a new employer within two months to maintain their visas and stay in the United States.

Core Departments Remain Unscathed

To provide context, Meta’s current AI division—the Super Intelligence Lab—comprises four departments:

• TBD Lab (Model Training and Scaling Department)
• FAIR (Fundamental AI Research Department)
• Product Applications (Product Integration Department)
• MSL Infra (Infrastructure Department)

The layoffs affected three of these departments; only TBD Lab remained untouched and will even continue to expand its recruitment. Clearly, this is the AI strategic department that Zuckerberg values most. Streamlining staff in other departments is intended to better support TBD Lab, as this department bears the core responsibility of developing Meta’s large models and AI products.

TBD Lab is a brand-new AI team established by Meta in late June this year, shortly after Alexandr Wang joined the company. Wang personally oversees the lab, which focuses on developing next-generation foundational models—including iterations of the Llama series—with the goal of achieving stronger reasoning capabilities and “super intelligence.”

TBD Lab consists of dozens of key employees, including not only elite members from Meta’s original AI team but also many well-known AI talents that Zuckerberg has spent heavily to poach from competitors like Google, OpenAI, and Apple over the past few months. The average salary level at TBD Lab is significantly higher than that of the other three departments.

Just how “star-studded” is TBD Lab’s team? It has assembled the top technical talents in the AI industry. Meta has been waving its checkbook to poach star employees from other AI companies: it once recruited eight core developers from OpenAI in one go, including several key architects behind GPT-4. It also poached Ruoming Pang, the head of Apple’s AI large model team, even offering an incredible contract worth $200 million.

External Executives Join Meta

In fact, Meta’s restructuring of its AI division may have been anticipated by outsiders. When Zuckerberg brought in Alexandr Wang to take full charge of the AI division in June this year, observers predicted that Wang would make major adjustments to Meta’s AI business priorities and resource allocation—especially for the FAIR fundamental research department.

In June, Zuckerberg approved Meta’s $14.8 billion investment in startup Scale AI, acquiring a 50% stake with no voting rights. As part of the deal, Zuckerberg also brought Alexandr Wang—founder and CEO of Scale AI—to Meta to take full responsibility for AI operations.

Zuckerberg has always been generous with deals he deems strategically significant for the company. Following Meta’s investment, Scale AI’s valuation reached $30 billion, boosting Alexandr Wang’s personal assets to $4.5 billion. Having achieved complete financial freedom, Wang was able to step away from the company he founded and lead Meta’s AI business with peace of mind.

What did Zuckerberg see in Alexandr Wang? As co-founder and CEO of Scale AI, Wang has a deep understanding and practical experience in AI model training data, data infrastructure, and how to efficiently build and evaluate AI models. These are core capabilities that Meta urgently needs to develop and optimize its AI models—especially large language models (LLMs) and generative AI. Meta even stated that Wang’s arrival would “deepen our work in generating data for AI models.”

Furthermore, Wang not only has a technical background but also possesses exceptional business acumen and execution capabilities. This time, Zuckerberg is betting on a different type of leadership—one more focused on business thinking and practical implementation, rather than pure research. In the AI arms race, quickly translating research results into usable products and capabilities is crucial.

However, after joining Meta from a startup environment, Wang believed that Meta’s AI division suffered from bloat and inefficiency. A major overhaul was inevitable, and layoffs were only a matter of time. He wanted a more streamlined, agile, and execution-driven department.

Llama’s Fiasco Is the Root Cause

Alexandr Wang’s appointment to lead Meta AI and the subsequent mass layoffs in the division may both be directly linked to the underwhelming performance of Llama 4. As is well-known, Meta’s flagship open-source Llama series garnered widespread attention and praise globally after its release in February 2023. However, Llama 4—released in April this year—disappointed many. Meanwhile, the rapid rise of Chinese models represented by DeepSeek has put immense pressure on the Llama team.

Why did Llama 4 perform so poorly? A former Meta employee who directly participated in the Llama team’s development revealed that the issue largely stemmed from poor decision-making by mid-level and senior management. The Llama team had originally prioritized the multi-modal direction, given Meta’s diverse product ecosystem (e.g., the metaverse, smart glasses, social media).

However, when DeepSeek emerged earlier this year with significantly stronger reasoning capabilities than Llama, it caused “great panic” within the Meta team. The team tried to excel in both multi-modal development and reasoning capability enhancement, but time was too tight—leading to product chaos.

In the eyes of this former Meta employee (who left the company before Zuckerberg brought in Alexandr Wang), the root problem of Meta’s AI team was a mismatch between roles and capabilities. Some mid-level and senior managers at Meta, who originally worked in product roles, were put in charge of teams focused on actual AI development—a case of “outsiders leading insiders.” This misalignment of authority was one of the key reasons he decided to leave.

The former employee noted that Zuckerberg and other executives, such as Chief Product Officer Chris Cox, are far-sighted and strong corporate leaders who were also deeply involved in the Llama development process. However, they could only focus on overall strategic deployment and could not oversee every detail. The failure of Llama 4 was not unrelated to the poor judgment of some mid-level and senior managers with weak AI backgrounds.

He believes that Zuckerberg realized there were serious team leadership issues behind Llama 4’s failure but felt he could not drive change through internal means alone. He needed to bring in an “external catalyst” to restore the team’s competitiveness. This was the direct reason Zuckerberg tapped Alexandr Wang to lead Meta AI.

While Zuckerberg did not hold the senior management of the Llama team accountable for the product’s failure, he later brought in Alexandr Wang to lead all AI departments and appointed Shengjia Zhao—poached from OpenAI—as Chief Scientist. This effectively marginalized or demoted the original AI team.

After Wang’s arrival, he formed his own direct-reporting elite team and poached top AI development talents from competitors with high salaries, directly widening the salary gap between the new team and the original Llama team. This disparity affected morale within Meta AI, prompting some original employees to seek job opportunities elsewhere.

Nevertheless, the former Meta employee believes Zuckerberg’s strategy is understandable. In his view, the newly recruited star employees all have unique strengths: “Competition in the AI track is fierce, and products iterate rapidly. Even a 15%-20% gap can determine success or failure. That’s why Zuckerberg is investing so heavily, regardless of the cost.”

Fundamental Research Becomes a Casualty

However, after Alexandr Wang’s arrival, the position of Yann LeCun—Meta’s former Chief Scientist, renowned AI scholar, French national, and self-given Chinese name “Yang Likun”—gradually became marginalized. LeCun is one of the “three godfathers of AI,” a Turing Award winner, the founder of the FAIR department, and a flagship figure for Meta AI.

FAIR was co-founded by Zuckerberg and LeCun in December 2013, with a mission statement: “Advance the state of AI through open research for the benefit of all.” Over the past decade, FAIR has made substantial contributions to fundamental AI research.

Unlike the intense pace of product development, fundamental research requires a more relaxed work environment. With Zuckerberg’s support, the FAIR team enjoyed a far more flexible atmosphere than the Llama development department. This distinction between “research” and “R&D” also exists at Microsoft Research.

However, after Alexandr Wang joined Meta, Zuckerberg merged the FAIR fundamental research team—led by LeCun—into the Super Intelligence Lab. From that point on, this fundamental research department was likely destined to become a casualty of Wang’s restructuring.

LeCun is presumably unhappy about the layoffs, as many of the dismissed employees worked directly under him. Rumors have already circulated within Meta that “LeCun will soon leave to start his own open-source AI startup.”

A former FAIR employee stated that LeCun has made significant contributions to the entire AI field and hopes society will respect and be tolerant of scientists like him: “His achievements are already extremely high. Moving forward, he carries a great deal of responsibility and a strong sense of mission—after all, he already has enough fame and fortune.”

Why would Alexandr Wang lay off staff from a research team like FAIR? Could it be because FAIR frequently competed with product model development teams for computing resources? A former Meta employee who worked on Llama development denied this. According to his account, the FAIR team does not require much computing power support. Meta provided unreserved support to the Llama development department, deploying most of the NVIDIA GPUs it purchased to that team.

This integration may reflect the strategic considerations of Zuckerberg and Alexandr Wang. Under the new AI leadership, the role of the FAIR team will also change. Meta will more actively integrate many of FAIR’s research ideas into projects managed by TBD Lab, with the traditional research role of publishing papers being replaced by an engineering role focused on launching products.

Clearly, Zuckerberg’s current AI priority is to accelerate model and product development to achieve immediate returns, rather than investing endlessly in fundamental research that may only show value a decade later. While Meta will continue to advance AI fundamental research, only projects aligned with the company’s priorities in the next phase are likely to receive support.

Competitors Rejoice Over Talent Drain

However, Zuckerberg’s approach—offering sky-high salaries to recruit AI talent while allowing Alexandr Wang to carry out large-scale layoffs—amounts to “gifting” numerous senior AI researchers to competitors, even with severance packages. Notably, Yuandong Tian, Research Director at FAIR and a well-known Chinese AI researcher, was also among those laid off. This news shocked many in the industry.

Yuandong Tian is the Director of Research Science at FAIR and enjoys high prestige in the field. He graduated from Shanghai Jiao Tong University and Carnegie Mellon University and has published numerous influential papers in reinforcement learning and large language models. Tian confirmed the layoff news to Silicon Valley Observer but politely declined an interview.

During his tenure at Meta, Tian led the development of DarkForestGo and OpenGo—Go-playing AIs that predated AlphaGo. OpenGo defeated four South Korean professional Go players (with no time limits for the players) in 20 consecutive matches, including Shin Jin-seo, currently ranked first in the world. In the field of large models, the “Attention Sink” phenomenon he discovered in 2023 has attracted widespread attention and been adopted in open-source GPT models.

Last year, his team launched Searchformer and Dualformer, which were among the first to identify the positive impact of ultra-long chains of thought on Scaling Laws and the possibility of mixed long-short chain-of-thought reasoning. The team also pioneered the “Continuous Chain of Thought” (Coconut) paradigm and demonstrated its superiority over discrete chains of thought. Most recently, Tian’s team discovered and characterized the dynamic mechanisms of neural network emergence and insight. Notably, Tian is also a science fiction novelist.

It is worth mentioning that a few months ago, Tian’s team was transferred from FAIR’s core projects to urgently address issues before Llama 4’s release. However, this did not change Llama 4’s failed outcome. Now, Tian himself has been laid off. Tian posted on X (formerly Twitter): “Those being laid off are not the ones responsible for the problems”—a statement that clearly carries profound implications.

As a well-known researcher in the industry, Tian does not need to worry about his future. His post was primarily intended to help his laid-off team members find new jobs. Thanks to his reputation, his X post—titled “Feel free to contact me”—quickly turned into an online job fair.

Colleagues from popular AI startups such as OpenAI and xAI flocked to comment below, lamenting that Meta had made a major mistake and expressing hope that Tian and his laid-off team members would join their companies. One comment read: “Yuandong Tian is undoubtedly one of the world’s top AI scientists. It’s a shame Meta let him go.”

There is no denying that the fact that Meta’s once-leading Llama large model has been surpassed by Chinese products has plunged Zuckerberg into a deep sense of crisis, prompting him to overhaul the team and organizational structure. Perhaps only time will tell whether Zuckerberg’s decision to bring in Alexandr Wang and restructure Meta AI will ultimately succeed.