Meta AI business earthquake! Three months after taking office, the new director started laying off employees. The basic research department suffered heavy losses, and even star researchers were unfortunately unemployed. Is Zuckerberg eager for quick success and self-destruction of the Great Wall? Or is he streamlining his organization and revitalizing his business?
A former Meta AI department employee told Silicon Valley Watch his personal experience in developing the Llama large model, revealing the story behind the massive layoffs in the Meta AI department.
In his view, it was the large open source model Llama 4 that clearly lagged behind DeepSeek and other Chinese competitors, which gave Zuckerberg a strong sense of crisis and prompted him to make up his mind to recruit AI leaders and elite talents from outside at all costs to completely restructure Meta AI's business and strategy.
Meta lays off AI department
Competition in the AI industry has become fierce, and talent has become the most in-demand asset that companies are competing for. In order to concentrate resources and invest in AI R&D competition, many technology giants such as Google and Microsoft have continued to expand their AI R&D departments and even resorted to large-scale layoffs in other non-core departments.
However, while Meta was eager for talent and offered sky-high salaries to poach competitors, it also targeted its own AI department with layoffs, firing up to 600 employees, including well-known senior researchers in the industry. This confuses and surprises many observers, and also allows many AI colleagues to see a great opportunity to recruit talents.
Last Wednesday, Meta announced a reorganization of its artificial intelligence department, laying off about 600 people. The news was announced by Alexander Wang, Meta’s chief AI officer and head of the Super Intelligence Laboratory, and is an internal adjustment within the department.
Alexander Wang wrote in a memo to employees: "By reducing the size of the team, fewer conversations will be needed to make decisions, and each person will assume more responsibility and have greater scope and influence." After he came to Meta, he has been committed to reducing the bloated organization and personnel size.
After the layoffs, the total number of employees in the Meta AI department dropped to less than 3,000. Affected employees were told that November 21 would be their departure date, and that until then they would be in a "non-work notice period" during which internal access would be removed.
In accordance with California layoff regulations, laid-off employees will retain their contracts and salaries for two months, and specific compensation will be determined separately. Some employees holding H1B work visas need to find a new employer within two months in order to keep their visas and continue to stay in the United States.
The core department remains unscathed
Let’s briefly introduce Meta’s current AI department, the Super Intelligence Laboratory, which includes four departments: TBD Lab (model training and expansion department), FAIR (basic AI research department), product application (product integration department), and MSL Infra (infrastructure department).
The layoffs involve three departments. Only TBD Lab has not been affected and will continue to expand recruitment. Obviously, this is the AI strategic department that Zuckerberg values most. The purpose of streamlining personnel in other departments is to better support TBD Labs. After all, this department is responsible for the core responsibility of developing Meta large models and AI products.
TBD Lab is a new AI team formed by Meta at the end of June this year with the arrival of Alexander Wang. It is personally responsible for Alexander Wang and focuses on developing the next generation of basic models, including iterations of the Llama series, with the goal of achieving stronger reasoning capabilities and "super intelligence."
TBD Lab includes dozens of key employees, including not only the elites of the original AI team, but also many well-known AI talents in the industry that Zuckerberg has spent a lot of money to poach from competitors such as Google, OpenAI, and Apple in the past few months. The average salary level is much higher than that of the other three departments.
How luxurious is the lineup of TBD Lab department? The top technical talents in the AI industry are gathered here. Meta wields a checkbook and poaches star employees of AI companies everywhere. They took away eight core development talents from OpenAI at one time, including several core architects of OpenAI's development of GPT-4. They also poached Pang Ruoming, the head of Apple's AI large model team, and even offered an incredible $200 million contract.
Foreign executives airborne into Meta
In fact, Meta’s restructuring and adjustment of the AI department this time may have been expected by the outside world. When Zuckerberg introduced Alexander Wang to take full charge of the AI department in June this year, the outside world expected that he would make major adjustments to Meta's major AI business focus and resource allocation, especially the FAIR basic research department.
In June of this year, Zuckerberg decided that Meta would invest US$14.8 billion in the startup Scale AI, acquiring half of the latter's non-voting shares. As part of the deal, Zuckerberg also brought Scale AI's founder and CEO Alexander Wang to Meta to take full responsibility for the AI business.
Zuckerberg has always been generous with deals as long as they feel they are strategically important to the company. With Meta's investment, Scale AI's valuation has reached US$30 billion, which has also brought Alexander Wang's personal assets to US$4.5 billion. With complete financial freedom, he can safely put down the company he founded and come to Meta to lead the AI business.
What does Zuckerberg see in Alexander Wang? As co-founder and CEO of Scale AI, he is passionate aboutAIModel training data, data infrastructure, and how to build and evaluate efficientlyAIModel with deep understanding and practical experience. This is a core capability Meta urgently needs to develop and optimize its AI models, especially large language models and generative AI. Meta even says that bringing in Alexander will “deepen the work we do in producing data for AI models.”
Moreover, Alexander Wang not only has a technical background, but also has excellent business acumen and execution capabilities. Zuckerberg is betting on a different type of leadership this time—one that is more business-minded and practical than a pure researcher. In the AI arms race, quickly translating research results into usable products and capabilities is critical.
However, after Alexander Wang, who leads the startup company, came to Meta, he believed that the Meta AI department was bloated and inefficient, so drastic rectification was bound to be carried out, and layoffs were only a matter of time. He wants to see a department that is leaner, more nimble and more capable of executing.
Llama fiasco is the root cause
Alexander Wang's airborne leadership of the Meta AI department, as well as the large layoffs in the AI department, may be directly related to the weak performance of Llama 4. As we all know, Meta’s flagship open source model Llama series has received a lot of attention and praise around the world after its release in February 2023. But Llama 4, released in April this year, was a huge disappointment. At the same time, the rapid rise of Chinese models represented by DeepSeek has put tremendous pressure on the Llama team.
Why does Llama 4 perform so poorly? A former Meta employee who was directly involved in the development of the Llama team revealed that this is more of a decision-making issue from the middle and senior levels. The Llama team originally prioritized investment in the multi-modal direction because Meta has a diversified product ecosystem, such as the metaverse, smart glasses, social media, etc.
However, after the emergence of DeepSeek at the beginning of this year, its reasoning capabilities significantly surpassed Llama, causing "very panic" within the Meta team. The team wanted to do it well on both sides, but time was running out, which led to product chaos.
According to this former Meta employee who has resigned, the root of the Meta AI team’s problems lies in the mismatch of people. Some of Meta's mid-level and senior managers who were originally working on products are now leading the people who are actually working on AI. This is the so-called "layman supervising the expert". This dislocation of voice was one of the important reasons for his decision to resign. (Note: He left Meta before Zuckerberg brought in Alexander Wang.)
According to this former Meta employee, management such as Zuckerberg and chief product officer Chris Cox are very far-sighted and strong business leaders. They are also deeply involved in the development process of Llama. However, they can only focus on the general direction of deployment and cannot pay attention to all details. The failure of Llama 4 is not unrelated to the judgment errors of some middle and senior managers with weak AI backgrounds.
He believes that Zuckerberg has realized that there are serious team leadership problems behind the failure of Llama 4, but also believes that it is difficult for him to bring about changes through internal means. He needs to find external catfish to make the team competitive again. This is the direct reason why Zuckerberg asked Alexander Wang to lead the Meta AI department.
Although Zuckerberg did not let the senior leaders of the Llama team take responsibility for the product failure, he subsequently brought in Alexander Wang to lead all AI departments, and appointed Zhao Shengjia, who was poached by OpenAI, as the chief scientist. This was tantamount to marginalizing or downgrading the original AI team in disguise.
After Alexander Wang arrived, he formed his own elite team and hired top AI development talents from competitors with high salaries, which directly widened the salary gap with the original Llama team. This gap has had a certain impact on morale within the Meta AI department, prompting some former employees to seek job-hopping opportunities.
But according to the former Meta employee, Zuckerberg’s strategy makes sense. In his opinion, the new star employees have their own merits. "The competition in the AI field is fierce, and products are rapidly iterated. Even a 15%-20% gap can determine life or death. That's why Zuckerberg invests so much regardless of the cost."
Basic research falls victim
However, after Alexander's arrival, the position of Meta's chief scientist and well-known AI scholar Yann LeCun (French, given his Chinese name) gradually became marginalized. Yang Likun is one of the "three godfathers" in the AI industry, a Turing Award winner, the founder of the FAIR department, and a signature business card of the Meta AI department.
FAIR was founded by Zuckerberg and Likun Yang in December 2013, with a mission statement to "advance the latest technologies in artificial intelligence through open research for the benefit of all." Over the past decade or so, FAIR has done a lot of fundamental work in AI research.
Unlike the intense pace of product development, basic research requires a more relaxed working atmosphere. Therefore, with Zuckerberg's support, the FAIR team enjoys a relaxed environment that is completely different from the Llama development department. But this is also the difference between research and development, and the same situation also exists at Microsoft Research.
But after the arrival of Alexander Wang, the basic research team FAIR led by Yang Likun was also merged into the Super Intelligence Laboratory by Zuckerberg. Perhaps, from then on, this basic research department was destined to become a victim of Alexander Wang's reforms.
Yang Likun may not be happy about this layoff, because many of the laid off people work directly under him. There are already rumors within Meta that "Yang Likun will leave soon to start his own open source AI startup company."
A former employee of the FAIR team said that Yang Likun has made many contributions to the entire AI field and hopes that society can respect and tolerate such scientists. "His achievements are already very high. He shoulders a lot of things moving forward and has a strong sense of mission. After all, fame and fortune are enough."
Why does Alexander Wang want to lay off research teams like FAIR? Is it because they often compete with product model development teams for computing resources? A former Meta employee who worked on Llama denied this. According to his introduction, the FAIR team does not require much computing power support. Meta provides almost unreserved support to the Llama development department, and the vast majority of purchased Nvidia GPUs are deployed in their development team.
This integration may reflect the strategic considerations of Zuckerberg and Alexander Wang. Under the leadership of the new AI leader, the FAIR team's responsibilities will also change. Meta will be more active in integrating many of FAIR's research ideas into projects run by TBD Lab. Traditional research roles that publish papers are being replaced by engineering roles that can launch products.
Obviously, Zuckerberg’s current AI priority is to accelerate the development of models and products and obtain immediate returns, rather than endlessly investing in basic research that will only see value ten years later. Although Meta will continue to promote basic AI research, it may only receive certain support if its priorities are consistent with the company's next stage.
Colleagues rejoice over brain drain
However, while Zuckerberg recruited AI talents at sky-high prices, he also allowed Alexander Wang to lay off employees on a large scale, which was equivalent to handing over many senior AI researchers to competitors. It is worth mentioning that Tian Yuandong, the research director of the FAIR team and a well-known Chinese AI researcher, is also in the scope of layoffs. The news shocked many in the industry.
Tian Yuandong is the director of research scientists at FAIR and enjoys a high reputation in the industry. He graduated from Shanghai Jiao Tong University and Carnegie Mellon University and published many influential papers in the fields of reinforcement learning and large language models. Tian Yuandong confirmed the news to Silicon Valley Observer, but declined the interview request.
While working at Meta, Tian Yuandong led the development of the Go AI DarkForestGo and OpenGo, which were earlier than AlphaGo. The latter's single-card reasoning defeated four Korean professional players in a row with a record of 20-0 (no time limit for players), including Shin Zhenzhen, who is currently ranked No. 1. In the field of large models, the Attention Sink phenomenon discovered in 23 years has received widespread attention and is used in the open source GPT model.
Last year, his team launched Searchformer and Dualformer and early discovered the positive impact of ultra-long thinking chains on Scaling Laws and the possibility of mixed long and short thinking chain reasoning. The team also pioneered the "continuous thinking chain" (Coconut) paradigm and proved its superiority over discrete thinking chains. He recently discovered and characterized the dynamic mechanism of neural network emergence and epiphany. It is worth mentioning that he is also a science fiction novelist.
It is worth mentioning that a few months ago, Tian Yuandong’s team was transferred from the core project of FAIR to urgently solve the problems before the release of Llama 4. However, this did not change the failure result of Llama 4. And now, he himself has been laid off. Tian Yuandong posted on X: "The person responsible for the problem is not the person who was laid off." This sentence obviously means a lot.
As a well-known researcher in the industry, he does not need to worry about his future. The purpose of posting is mainly to seek job opportunities for colleagues in his group who have been laid off. Thanks to Tian Yuandong's reputation in the industry, his X article "Welcome to Contact Me" immediately became an online job fair.
Colleagues from popular AI startups such as OpenAI and xAI left messages below, lamenting that this was a big mistake made by Meta and hoping that Tian Yuandong and his laid-off colleagues could join their company. "Tian Yuandong is undoubtedly one of the best AI scientists in the world. It is a pity that Meta let him leave." One comment wrote.
It is undeniable that the originally leading Llama large model was surpassed by Chinese products, which plunged Zuckerberg into a deep sense of crisis, prompting him to change the team's status quo and organizational structure. Perhaps, only time will tell whether Zuckerberg’s reorganization of the Meta AI department by introducing Alexander Wang is successful.
Silicon Valley Observation/Zheng Jun