0% read
Last updated:
Meta Superintelligence Labs
Meta Superintelligence Labs (MSL) is a dedicated research and development division within Meta Platforms, established with the explicit goal of advancing artificial intelligence to achieve superintelligence, a form of AI with capabilities surpassing human intellect. The formation of this team reflects Meta's increased focus on advanced AI development and its role as a competitor in the AI sector.[1] In April 2026, MSL introduced Muse Spark, the first model family designed as a step toward Meta’s vision of a “personal superintelligence” that can assist users across many aspects of their daily lives.[8]
History
Meta's push into superintelligence development is an initiative led by CEO Mark Zuckerberg, who has publicly stated his view that superintelligence could mark "the beginning of a new era for humanity." The company is treating the development of superintelligence as a concrete engineering goal rather than a speculative concept, with the aim of creating an AI that could rival or exceed human abilities across disciplines, including scientific creativity and emotional understanding. Zuckerberg's stated vision is to build a "personal superintelligence" that could be integrated into personal devices, such as smart glasses, to act as an assistant managing information and anticipating user needs.[7]
The establishment of Meta Superintelligence Labs reflects a shift in Meta's AI strategy toward ambitious, long-term research goals. The initiative consolidates AI talent and resources under a unified vision, with a focus on advancing capabilities in areas such as reasoning, problem-solving, and general intelligence. The company's financial backing and strategic partnerships indicate the priority of this endeavor within Meta's overall business objectives.[1][2][5]
To build out Meta Superintelligence Labs, Meta undertook an aggressive talent acquisition campaign involving substantial investment to attract AI researchers and engineers from rival organizations. This strategy relied on highly competitive compensation packages, with some reportedly exceeding $100 million for top-level hires and total packages for certain individuals described as reaching into the hundreds of millions of dollars.[9][1] The lab recruited individuals from OpenAI, Google DeepMind, Anthropic, Apple, and Ilya Sutskever's Safe Superintelligence, positioning MSL as a major destination for frontier AI talent.[1][2][3] In 2026, for example, prominent AI researcher Luke Metz, previously at OpenAI, joined Meta Superintelligence Labs to work under Alexandr Wang, illustrating how high-profile hiring was treated as a core part of the lab’s initial build-out.[12]
Products & Programs
In April 2026, Meta Superintelligence Labs introduced Muse Spark, the first model in the Muse family developed as part of Meta’s strategy to build a personal superintelligence.[8] Muse Spark is described as a natively multimodal reasoning model that integrates text and visual understanding, supports tool use, and employs visual chain of thought and multi-agent orchestration to solve complex tasks.[8] The model targets applications such as interactive health guidance, visual STEM problem-solving, and personalized assistance that can analyze a user’s environment and provide step-by-step support.
Muse Spark introduces a “Contemplating” mode that orchestrates multiple agents reasoning in parallel to improve performance on challenging benchmarks, including tasks such as Humanity’s Last Exam and FrontierScience Research.[8] The model is available to users through meta.ai and the Meta AI app, and Meta has opened a private API preview for selected developers, positioning Muse Spark as the first productized output of MSL’s rebuilt training and reinforcement learning stack.[8]
As part of its deployment, Muse Spark underwent safety evaluations under Meta’s Advanced AI Scaling Framework, which defines risk categories, evaluation protocols, and deployment thresholds for advanced models.[8] These evaluations included testing refusal behavior in high-risk domains, behavioral alignment, and adversarial robustness before broader rollout.
Governance
The Meta Superintelligence Labs (MSL) organization was announced by Mark Zuckerberg on Threads, with details further elaborated in an internal memo. Zuckerberg stated that MSL encompasses Meta's foundation model, product, and FAIR (Fundamental AI Research) teams, along with a new lab focused on developing the next generation of models.[1] Alexandr Wang, the former CEO of Scale AI, was appointed as Meta's Chief AI Officer and leads MSL.[9] Within the lab, Wang co-leads the new model group with Nat Friedman, the former head of GitHub, who focuses on AI products and applied research.[1][2]
Shengjia Zhao, an engineer formerly with OpenAI, was named the chief scientist for Meta Superintelligence Labs. In this role, Zhao sets the research agenda and scientific direction for the lab, working directly with Mark Zuckerberg and Alexandr Wang.[1][2][3] As part of its governance and safety structure, MSL evaluates frontier models such as Muse Spark using Meta’s Advanced AI Scaling Framework, which defines threat models, evaluation protocols, and deployment thresholds before wide release.[8]
According to internal communications, MSL also incorporates specialized safety leadership. Co-founders of AI security startup Virtue AI—Bo Li and Dawn Song—report to Nat Friedman at MSL, while fellow co-founder Sanmi Koyejo reports to Rob Fergus, head of Meta’s FAIR lab, aligning safety research and policy work across the organization.[11]
Senior staffing at MSL includes researchers and engineers such as Andrew Tulloch, Joshua Gross, Mark Jen, Yinghai Lu, Jason Wei, Hyung Won Chung, and Zhiqing Sun, along with executives like Dina Powell McCormick, who support the expansion of the lab’s research programs and compute infrastructure.[10]
Key Team Members
The Meta Superintelligence Labs includes researchers, engineers, and technical professionals with prior experience at major organizations involved in artificial intelligence research and development.
Shengjia Zhao, formerly involved in the development of ChatGPT at OpenAI, was appointed Chief Scientist of Meta's Superintelligence Lab. Other individuals from OpenAI on the team include Trapit Bansal, Ji Lin, Hongyu Ren, Jiahui Yu, Shuchao Bi, Huiwen Chang, Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai. Their previous work encompassed areas such as reinforcement learning, voice interfaces, image processing, and large language model development. Trapit Bansal pioneered reinforcement learning on chain of thought and was a cocreator of o-series models at OpenAI. Shuchao Bi cocreated GPT-4o voice mode and o4-mini, and previously led multimodal post-training at OpenAI. Huiwen Chang cocreated GPT-4o's image generation and previously invented MaskIT and Muse text-to-image architectures at Google Research. Ji Lin contributed to o3/o4-mini, GPT-4o, GPT-4.1, GPT-4.5, 40-imagegen, and the Operator reasoning stack. Hongyu Ren cocreated GPT-4o, 4o-mini, o1-mini, o3-mini, o3 and o4-mini, and previously led a group for post-training at OpenAI. Jiahui Yu cocreated o3, o4-mini, GPT-4.1 and GPT-4o, and previously led the perception team at OpenAI and co-led multimodal at Gemini. Shengjia Zhao cocreated ChatGPT, GPT-4, all mini models, 4.1 and o3, and previously led synthetic data at OpenAI. Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, who previously helped launch OpenAI’s Zurich office, were noted to have joined Meta, though their inclusion on specific internal lists was subject to "technicalities."
Former Google DeepMind contributors Jack Rae and Pei Sun are part of the team, along with engineers Tianhe Yu, Cosmo Du, and Weiyue Wang, who were involved with the Gemini model. Jack Rae was a pre-training tech lead for Gemini and reasoning for Gemini 2.5, and led Gopher and Chinchilla early LLM efforts at DeepMind. Pei Sun contributed to post-training, coding, and reasoning for Gemini at Google Deepmind, and previously created the last two generations of Waymo's perception models. From Apple, the group includes Ruoming Pang, who directed work related to large language models, as well as Mark Lee and Tom Gunter.
The team also includes individuals from other AI-focused companies. Joel Pobar was previously affiliated with Anthropic and had prior experience at Meta for 11 years on HHVM, Hack, Flow, Redex, performance tooling, and machine learning. Johan Schalkwyk worked at Sesame AI, and was a former Google Fellow and early contributor to Sesame, and technical lead for Maya. Alexandr Wang is the founder of Scale AI. Daniel Gross, co-founder of Safe Superintelligence, and Nat Friedman, former CEO of GitHub, are also identified as contributors to Meta's AI initiatives.
These individuals are part of a broader group of 44 professionals reported to be involved in Meta’s Superintelligence Labs efforts.[1][2][3][4][6]
Team Members Full List
- Nat Friedman (GitHub)
- Daniel Gross (Safe Superintelligence / NFDG)
- Yann Le Cun (Meta AI / NYU)
- Joel Pobar (Anthropic)
- Mat Velloso (Microsoft / DeepMind)
- Annie Hu (Scale AI)
- Ruben Mayer-Hirschfeld (Scale AI)
- Linda Gong (Two Sigma / Anrok)
- Jon Wilfong (E&A / Mulesoft)
- Summer Yue (DeepMind)
- Alexander Kolesnikov (DeepMind)
- Allan Jabri (OpenAI)
- Anton Bakhtin (Meta / Yandex)
- Bowen Cheng (Tesla / OpenAI)
- Chengxu Zhuang (OpenAI / MIT)
- Chenxi Liu (Waymo / DeepMind)
- Chunyuan Li (Microsoft Research / X.AI)
- Gabriel Ilharco (X.AI)
- Haotian (Ken) Tang (DeepMind)
- Hongyu Ren (OpenAI)
- Huiwen Chang (OpenAI / Google)
- Jack Rae (DeepMind)
- James Lee-Thorp (DeepMind / Goldman Sachs)
- Ji Lin (OpenAI)
- Jiahui Yu (OpenAI)
- Julian Michael (NYU / OpenAI)
- Lu Liu (Google / OpenAI)
- Lucas Beyer (DeepMind / OpenAI Zurich)
- Matt Deitke (Vercept / Allen Institute for AI)
- Michael Zhang (DeepMind)
- Qingqing Huang (Google / TikTok)
- Rui Hou (X.AI / Meta)
- Pingchuan Ma (OpenAI)
- Shengjia Zhao (OpenAI)
- Shuchao Bi (OpenAI)
- Trapit Bansal (OpenAI)
- Xiaohua Zhai (DeepMind / OpenAI Zurich)
- Yinghao Li (OpenAI)
- Yuanzhi Li (DeepMind)
- Zhishuai Zhang (Waymo / DeepMind)
- Hammad Syed (Dubizzle / PlayAI)
- Yu Zhang (Twitter / Dropbox / DeepMind)
- Pei Sun (Waymo / DeepMind)
- Tao Zhu (DeepMind)
- Jason Wei (OpenAI)
- Hyung Won Chung (OpenAI)
- Zhiqing Sun (OpenAI)
Funding
Meta is committing substantial financial and computational resources to its superintelligence initiative. In June 2025, the company invested about $14.3 billion to acquire a 49% stake in Scale AI, a company specializing in labeling data for AI systems, which is considered crucial for training advanced models and brought Scale cofounder Alexandr Wang to Meta as head of MSL.[9] To support this initiative, Meta is developing several multi-gigawatt "supercluster" data centers. Two of these facilities are reportedly codenamed Prometheus and Hyperion, with one planned data center described as having a physical footprint comparable in size to most of Manhattan.[7]
Analysis of Meta’s infrastructure program describes a build-out of five 1 GW–plus “titan” clusters, including the Hyperion campus in Louisiana and the Prometheus campus in Ohio, along with three additional sites in El Paso, Iowa, and Indiana. For Hyperion, Meta is reported to be constructing three 400 MW buildings and three 100 MW buildings, totaling about 1.5 GW under construction, while Prometheus is projected to expand from an initial capacity of around 1 GW to more than 3 GW within two years.[10] Beyond talent acquisition and strategic investments, Mark Zuckerberg has stated that Meta plans to invest "hundreds of billions of dollars into compute" for superintelligence development.[1][2] Company guidance indicates total capital expenditures of $130–145 billion in 2026, more than double Meta’s spending in 2024 and 2025 combined, with the bulk directed toward AI-focused data centers and infrastructure.[9]
Industry analysis projects that, based on these data center investments, Meta could have more AI training compute than both OpenAI and Anthropic by the end of 2026. Even under conservative assumptions that assign only high-profile data center sites to Meta Superintelligence Labs, the lab’s training compute is expected to be comparable to other leading frontier labs through 2026 and 2027.[10]
Partnerships
Meta’s superintelligence initiative is closely linked to its strategic investment in Scale AI. In 2025, Meta acquired a 49% stake in Scale AI for about $14.3 billion, securing access to large-scale data labeling services considered critical for training advanced models and integrating Scale’s expertise into Meta Superintelligence Labs through Alexandr Wang’s leadership position.[9][1]
MSL has also expanded its capabilities in AI safety and security through the acqui-hire of Virtue AI’s founding team. Virtue AI develops automated red teaming tools, real-time safety systems to block harmful or illegal outputs, and compliance frameworks for enterprise AI deployments, and its co-founders Bo Li, Dawn Song, and Sanmi Koyejo joined Meta to work on safety and security across MSL and FAIR.[11]
See something wrong?
