Meta has made another high-profile move in the fierce battle for AI talent by recruiting Yang Song, a prominent AI researcher formerly leading OpenAI’s strategic explorations team, to lead its Superintelligence Lab.
Song, widely known for pioneering diffusion model research that was used in building state-of-the-art generative AI like DALL-E 2, joined Meta as research principal earlier this month. Reporting to Meta Superintelligence Labs’ chief scientist Shengjia Zhao, himself an OpenAI alumnus, Song’s arrival marks a strategic boost for Meta’s ambition to further develop artificial general intelligence (AGI).
Song’s expertise in multimodal AI systems, which is the foundation of advanced generative AI technologies, is also a perfect fit for Meta’s vision of developing AI that can operate seamlessly across text, images, and complex data landscapes.
A game-changing talent acquisition
Yang Song has been at the forefront of AI innovation, contributing foundational research that significantly influenced OpenAI’s breakthrough image generation system, DALL-E 2. Song also directly contributed to the evolution of AI capabilities, where he led strategic explorations to improve generative model architecture, training objectives, and data efficiency.
Now at Meta, Song will assume the role of Research Principal at Superintelligence Labs, working closely under Shengja Zhao, the Lab’s newly appointed chief scientist and another former OpenAI prodigy.
Both Song and Zhao have a lot in common, as they were both undergraduates in Tsinghua University, as well as classmates as doctorate students at Stanford University.
At OpenAI, Song led the strategic explorations team, where his work contributed to the diffusion models that helped lay the groundwork for multimodal generative AI systems, most notably OpenAI’s DALL-E 2, a landmark project in AI-based image generation.
Similarly, Zhao played a key role in foundational model research and development at OpenAI, focusing on enhancing ChatGPT’s ability to generalize knowledge across diverse data types. In a memo that was seen by WIRED weeks ago, Zuckerberg touted Zhao as the co-creator of GPT-4, all mini models, 4.1, and o3 at OpenAI.
Bringing both talents together, although at different periods, at Meta’s Superintelligence Lab (MSL) sets the stage for a powerful synergy. The shared background, shaped by their complementary strengths and research standards, will be effective in advancing diffusion models, generative AI, and multimodal learning in Meta AI, which are also fields critical to the future of AGI; a path to “superintelligence.”
Meta’s bold AI ambitions
Since its launch in Mid-2025, Meta’s Superintelligence Labs has been the heart and core of the company’s AI innovation strategy. The Lab’s mission surpasses traditional AI applications, as it seeks to build AI systems that are superior to human cognition across all areas.
Under the leadership of former Scale AI CEO Alexandr Wang, MSL will consolidate foundational research, infrastructure, and product teams to streamline efforts toward achieving the superintelligence goal.
Yang Song’s expertise will crucially bolster Meta’s capacity to innovate in diffusion-based generative AI and multimodal learning, areas instrumental in building AI with broad understanding and creative capability.
To achieve these huge AI ambitions, especially while developing MSL, Meta has also notably shifted from offering open-source LLMs. This means that open-source AI development will take a back seat in the face of the development of AGI.
The fierce AI talent war
Yang Song’s move to Meta highlights the intensifying competition among tech giants to secure top-tier AI talent.
So far in 2025, Meta has embarked on one of the most aggressive recruitment campaigns in AI history, where it has successfully attracted over a dozen senior researchers from OpenAI, Google DeepMind, Anthropic, and other leading AI institutions.
In what many industry insiders now call the “Zuck Bucks,” compensation packages under this war for talent reportedly reach $300 million, reflecting the extraordinary and immense value these AI companies place on AI talents and researchers that are capable of driving the next wave of AI breakthroughs.
In addition to this, Meta also pledged between $64 billion and $72 billion to build AI infrastructure in 2025 alone, rivaling the investments of tech giants like Google and Microsoft, who are also actively involved in the AI arms race.
Yet, the battle to achieve superintelligence is not without its own challenges. Several high-profile hires have reportedly departed Meta shortly after joining, with one researcher reportedly not showing up for work on their first day of work. Another AI researcher, Aurko Roy, also left the company in July, after working there for just five months.
This suggests that cultural project alignment problems or factors might be as critical as financial incentives in retaining elite talent.
However, the addition of Yang Song sends a clear message about Meta’s AI strategy, where there is a major investment and focus in generative AI models that are capable of multimodal synthesis and comprehension; perhaps exactly what Meta AI needs to catch up with other leading AI companies in the escalating AI arms race.
