Young Wartime CEO: How to Win Meta's AI Comeback Battle

The competition in the field of artificial intelligence has only just begun. Every company has a shot, and major tech giants are all placing their bets. Yao Shunyu is the head of AI at Tencent, while Xiaomi has placed its bet on female scientist Luo Fuli. Although Alibaba's Qwen team went through the event of Lin Junyang stepping down, the comprehensive strength of its other leading figures still cannot be underestimated. This article starts with the story of Meta's AI comeback, looking at how these large capitals bet on these top-educated, highly young AI leaders.

In the spring of 2025, an atmosphere of being left behind permeated inside Meta. The release of Llama 4 received a dismal response, and its benchmark test was exposed to the practice of "using different models for different tests", which completely exhausted Mark Zuckerberg's patience with the generative AI division. Turing Award winner Yann LeCun later stated bluntly after leaving his post that Meta only chose "safe, proven things that work", and this choice itself meant falling behind.

Fourteen months later, the situation reversed. On October 12, 2026, Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full. The independent evaluation agency Artificial Analysis gave them an intelligence index score of 52 and 56 respectively. The latter is on a par with Anthropic's Claude Sonnet 5.5, surpassing GPT-6 Astra and Gemini 4 Argon. The headline of Wired was full of praise: "Muse Wonder Puts Meta Back at the AI Frontier".

The one who pulled Meta back from the trough is a 29-year-old Chinese-American entrepreneur. His name is Alexandr Wang, Chinese name Wang Tao, the founder of Scale AI. In June 2025, Mark Zuckerberg poached him along with his company into Meta with an investment of 14.3 billion US dollars, appointing him as Chief AI Officer to take charge of the newly established Superintelligence Lab.

Zuckerberg did not bet on Wang's scientific research ability. Yann LeCun put it plainly: "He has no scientific research experience, does not know how to carry out scientific research activities, let alone understand the preferences of researchers". What Zuckerberg bet on is something else: an entrepreneur's sense of urgency at the critical moment of life and death, a "wartime state" that has almost disappeared among professional managers of large companies.

Wang brought this state into Meta. The first thing he did was to transform himself from a "founder" to a "coach". During the Scale AI period, all decisions had to go through him personally, which made him a bottleneck for business advancement. After arriving at Meta, he clearly knew that he was "neither an AI researcher nor a product designer", so he shifted his focus entirely to talent recruitment, scouting top talents in various professional fields, and then creating an environment for them to give full play to their expertise.

This is not modesty, but a strategy. Meta's AI division laid off about 600 people in October 2025, affecting established research teams such as FAIR, and Chinese scientist Tian Yuandong was also among the laid-off. But TBD Lab led by Wang himself — the core team focused on building "superintelligence" — was completely unharmed in the layoffs. He replaced Meta's old blood with new blood, and used the sense of urgency of "outsiders" to clean up the slackness of the "old-timers".

The cost of this cleanup was drastic. Yann LeCun left his job at the end of 2025 and founded his own AI startup. Before leaving, he left a meaningful sentence: Meta's new team "is full of large language models in their minds". In other words, the world model route he insisted on was marginalized. Some people inside Meta also refused to accept Wang. A former employee commented on Muse Spark: "Those people in TBD, Alex and Zuckerberg, set very low standards for both internal and external parties".

But Wang bet on another path. He did not catch up head-on with OpenAI and Anthropic on the programming and abstract reasoning tracks where Meta was the weakest, but placed his bets on "personal agents".The core selling point of Muse Wonder is not chatting, but Active Memory and Active Focus — allowing AI agents to remember user context across conversational scenarios, and actively adjust plans instead of giving up goals when encountering obstacles. Wang said frankly in an interview with Bloomberg on the release day: "Other ecosystems have been focusing on programming agents. We have adopted a differentiated strategy to make the application scenarios of personal agents truly perform excellently".

This judgment makes sense. Meta owns Facebook, Instagram and WhatsApp, and the 3 billion users' daily behavior data is a rich ore that no competitor has. Feeding these data to a personal agent that truly "understands you" has a much higher chance of winning than competing head-on with OpenAI on the programming track.

Wang's other card is speed.Muse Spark was released in April 2026, and Muse Wonder was launched in October, with only six months in between. This iteration rhythm is rare in Meta's AI history. Russ Salakhutdinov, a professor at Carnegie Mellon University and former Meta Vice President of AI Research, commented: "The amount of work TBD Lab has completed in such a short time is impressive. Alex knows what he doesn't know, and he is willing to listen".

"Knowing what you don't know" — this is probably Wang's most underrated ability. A 29-year-old non-researcher leading a group of top AI scientists, it is easy for others to fall into two extremes: either forcibly play the role of technical authority, or completely delegate power and lose direction. Wang chose the third path: he recruited his Tsinghua alumni Zhao Shengjia as Chief AI Scientist to supervise the entire scientific agenda of MSL; he recruited Dawn Song, a professor at the University of California, Berkeley, to lead the new AI research division. He does not do research himself, but ensures that the people doing the research are the right ones.

Of course, Meta has not won yet. Muse Wonder Full is currently only open to Muse Code subscribers, the standard version of Wonder is priced via API, and it is still unknown whether the most powerful version will be open to enterprise customers. Anthropic's Claude Opus 5.5 still leads the intelligence index with a score of 58, and GPT-6.1 Sol still has an advantage in cost efficiency. Meta's AI talent war has also paid an astonishing price, offering signing bonuses as high as 100 million US dollars to researchers from OpenAI and Google.

But at least, Meta is back at the poker table. From the fiasco of Llama 4 to the cutting-edge performance of Muse Wonder, from Yann LeCun's departure to the efficient operation of TBD Lab, Meta has completed a costly but clearly directed transformation in 14 months. Zuckerberg did not bet on a technical genius, but on an entrepreneur who has been in the data labeling industry for ten years and knows that "trust" and "delivery" are more valuable than "vision".

The most important lesson Wang learned at Scale AI may not be how to label data, but how to make top talents believe that they can win by following him. The people left after the turmoil of Meta's AI team, Zhao Shengjia, Dawn Song, and those researchers in TBD Lab who chose not to leave, form a more streamlined, more focused, and more dangerous team than the Llama era.

Muse Wonder is not the end. Wang said a sentence in the interview on the release day, which is like an invitation to skeptics and a reminder to competitors: "I would suggest everyone try it out for themselves and then judge for yourself".

Whether Meta can overtake others at the last minute, the answer to this question is not on Zuckerberg's balance sheet, but on what Wang can deliver in the next six months.

This article is from WeChat Official Account "Co-opetition Artificial Intelligence", author: Co-opetition, published with authorization from 36Kr.