Recently, Meta has launched a "reorganization action" in the field of AI - huge investment, top talent recruitment, new technology research and development, and Zuckerberg's every step has attracted global attention. Meta's series of changes not only concern its own future, but also reflect the fierce competition among global technology giants in the AI ​​track.

Mark Zuckerberg

Currently, Meta plans to acquire approximately 49% of the shares of ScaleAI, an American AI data annotation company, for approximately US$14.8 billion. If the acquisition is completed, this will become Meta's largest external investment to date. According to people familiar with the matter, in terms of the transaction structure, Meta will pay cash directly to existing shareholders of Scale AI and plans to hire Alexandr Wang, CEO of Scale AI, as the head of Meta's Super Intelligence Lab.

By purchasing shares directly from existing shareholders rather than through traditional mergers, Meta can avoid triggering the US antitrust review process. This transaction structure enables Scale AI to continue to cooperate with competitors such as OpenAI and Microsoft as an independent company, while also expanding Meta's social influence.

Public information shows that ScaleAI was founded in 2016 and is the world's leading data annotation and model evaluation company. Its main business is to provide data annotation solutions. Its customers include OpenAI, Microsoft, General Motors, etc. The company's revenue in 2024 is nearly US$870 million, and it is expected to exceed US$2 billion in 2025. Meta's US$14.8 billion investment will bring Scale AI's valuation to approximately US$30 billion.

The founder of ScaleAI was born in 1997. He began to self-study programming through the Internet in high school. At the age of 17, he became a full-time coder for Quora, a well-known American question-and-answer website. In 2015, he was admitted to the Massachusetts Institute of Technology in the United States. In 2016, 19-year-old Alexander Wang dropped out of school to establish Scale AI. Relying on the data annotation business, Alexander Wang, 24, became the youngest self-made billionaire in the world at the time in 2021.

Alexander Wang to join Super AI Lab

In this age where data is gold, Meta’s move is aimed at building its own data system to match its artificial intelligence strategy and build a data foundation to compete with OpenAI, Anthropic, Google and other companies. At the same time, Alexander Wang, the founder of Scale AI, will step down as CEO after the investment transaction is completed and join Meta’s “Super AI Lab”.

Previously, CEO Mark Zuckerberg publicly expressed disappointment with Meta AI’s progress, and the poor performance of the Llama 4 model also made Zuckerberg uneasy. In January this year, CEO Mark Zuckerberg listed artificial intelligence as Meta’s top priority and revealed plans to invest up to $65 billion in artificial intelligence by 2025.

According to multiple reports, Mark Zuckerberg is personally setting up a “Super AI Lab” consisting of 50 top AI engineers and researchers. According to people familiar with the matter, Zuckerberg is currently personally responsible for the recruitment of the lab and has even rearranged Meta’s office so that new employees can sit near him.

Meta's Super Intelligence Lab represents the company's full-scale promotion of AI strategy and its strong vision to remain competitive in the current increasingly fierce AI competition. It is determined to develop AI systems that surpass human cognitive abilities in all fields and achieve the goal of general artificial intelligence.

This layout may be a strong rebound after the sluggish performance growth of Llama4 and the departure of senior talents such as Joelle Pineau, vice president of Meta Artificial Intelligence Research Center and professor of computer science at the University of Montreal, and Dan Reed, COO of Reality Labs.

According to Bloomberg, Zuckerberg has been meeting with AI researchers and engineers at his homes in Lake Tahoe and Palo Alto, and Meta has offered seven- to nine-figure compensation packages to dozens of researchers from companies such as OpenAI and Google.

Some researchers have agreed to join Meta, including Jack Rae, a top researcher at Google DeepMind, and Johan Schalkwyk, head of machine learning at the popular voice assistant application Sesame AI.

Meta wants to join the "national team" with Scale AI

So far, unlike companies such as Microsoft, Amazon and Google, Meta has always adopted a go-it-alone strategy, focusing mainly on internal research and open source projects. Meta's competitors have reached cooperation with external companies and have made huge investments in AI unicorns such as OpenAI and Anthropic.

Meta's huge investment in Scale AI may not be a temporary idea. Previously, Meta had expressed its optimism about the prospects of Scale AI and participated in Scale AI's $1 billion F round of financing with companies such as Amazon.

In 2024, Scale AI won a defense contract from the US Armed Forces and was commissioned by the Pentagon's Chief Digital and Artificial Intelligence Office to test and evaluate the security and reliability of large language models used for military planning and decision-making. Behind this, it relies on Defense Llama, which is based on Meta's large model Llama 3.

Defense Llama is a large language model built on Meta Llama 3, which is specially customized and fine-tuned to support US national security missions. Defense Llama can assist in planning military or intelligence operations while understanding the military weaknesses of opponents during operations.

Through the investment, Meta can directly access Scale AI's existing high-quality database and conduct further business exploration at the defense and military level. Currently, Meta has cooperated with defense company Anduril Industries to jointly develop artificial intelligence helmets with augmented reality and virtual reality functions for the US military, and authorized US government agencies and contractors to use its company's artificial intelligence models.

World Model V-JEPA 2

When Meta's reorganization and heavy money poaching were making a lot of noise, some netizens posted a post speculating that LeCun would sit on the bench at Meta.

But Meta's chief AI scientist Yann LeCun personally appeared on the screen to introduce Meta's new video-based training world model V-JEPA 2 (Video Joint Embedding Predictive Architecture 2).

AI scientist Yann LeCun

There are currently two schools of thought in general intelligence. The Transformer school (such as the GPT series) focuses on big data-driven autoregressive learning, and the other is the world model school represented by Yann LeCun. The world model school believes that it is necessary to emphasize common sense knowledge and environmental interaction by building an intrinsic understanding of the world.

Meta's latest world model V-JEPA 2 has 1.2 billion parameters and is built on the Joint Embedding Prediction Architecture (JEPA) proposed by Yann LeCun. In the video demonstration, Yann LeCun described V-JEPA 2 as a "realistic abstract digital twin". Unlike understanding language, world models enable machines to understand the physical world, while being able to predict the consequences of their actions and autonomously plan action plans to complete tasks given by humans.

The model breaks away from traditional pixel-level predictions and gives robots and agents human-like "common sense physical reasoning" capabilities, such as automatically understanding object interactions and predicting motion trajectories, thereby greatly improving generalization performance in real environments, marking a shift from language-based artificial intelligence to more spatially aware systems.

"As humans, we think language is essential for intelligence, but it's not," said Yann LeCun. "What if AI could develop common sense similar to that of humans and animals and predict what will happen in some abstract spatial representation?"

We no longer have to conduct millions of trials on AI in the real world, because the world model can simulate the real world for AI models in minutes and train them based on an understanding of how the world works.

According to officials, the model is designed to provide real-time spatial understanding for AI-driven technologies such as self-driving cars, warehouse robots, and drone delivery systems. LeCun said in the video: "The world model will usher in a new era of robotics, allowing real-world robots to help with housework and physical tasks without a lot of training data."

The model mainly uses a self-supervised learning model to learn from more than one million hours of video and one million images. The huge data set enables the model to capture complex physical interaction patterns without tedious manual annotation.

The training of V-JEPA 2 includes two stages: no-action and action-guided training. In the first stage, the model learns general world dynamics, such as object movement, gravity effects, and collisions, without considering any specific actions. In the second phase, about 62 hours of robot control data were introduced to enable the model to explicitly consider the robot's actions when predicting results.

In this way, using only small-scale robot data, V-JEPA 2 learned to predict and plan based on specific actions.

The model was trained based on the open source DROID dataset and can be directly deployed to the robot in the laboratory, making it truly "ready to use". Through the demonstration video, the robot in the Meta laboratory can successfully complete tasks such as grasping, picking up objects, and placing them in new locations.

V-JEPA 2 has been open sourced in the community, and the Meta team has also released three new benchmarks, IntPhys 2, MVPBench, and CausalVQA, to evaluate the ability of existing models to understand and reason about the physical world from videos.

Whether it is investing in data annotation companies, investing heavily in new laboratories, or releasing new world models, Meta's ambitions in the field of AI are all demonstrated. Let us wait and see when the next reshuffle of the global AI ecosystem will come.