> **来源：[研报客](https://pc.yanbaoke.cn)** # The Future of Meta Superintelligence: A 1 Year Progress Update Summary ## Core Content This document provides an overview of Meta's progress in the AI domain over the past year, focusing on their strategic moves in data, talent, and compute. It compares Meta's efforts with those of OpenAI, Anthropic, and Google, highlighting their unique position and potential to catch up in the AI race. ## Main Points - **Meta's Rebuilding Efforts**: After the Llama 4 release, Meta has been rebuilding its AI organization, investing heavily in talent and infrastructure. - **Talent Acquisition**: Meta has made significant moves to acquire top AI talent, including a \$14.3B investment to poach Alexandr Wang and his team from SEAL, as well as offering multi-hundred million dollar pay packages to top researchers and engineers. - **Data Strategy**: Meta has begun tracking employees' screen, keyboard, and mouse activities, creating high-quality, real-world data for training AI models. This is seen as a major advantage, especially in comparison to other companies that rely on third-party data providers. - **Compute Expansion**: Meta is aggressively expanding its AI compute infrastructure, building multiple "titan" datacenter clusters with capacities exceeding 1GW, including Prometheus, Hyperion, and three unnamed campuses in El Paso, Iowa, and Indiana. - **Network Architecture**: To manage the scale-across of their datacenters, Meta has introduced AI-Backbone (AIBB), a specialized network architecture that allows for efficient communication across multiple locations. - **Scale-Across Challenges**: While AIBB helps with communication, the physical distance between datacenters introduces latency, forcing Meta to use asynchronous training strategies. - **Competitive Landscape**: The AI field is currently a two-horse race between OpenAI and Anthropic, with Google lagging behind. Chinese labs are also struggling with compute limitations. - **MSL's Potential**: Despite the initial underperformance of Muse Spark, MSL is believed to be on track to catch up with OpenAI and Anthropic due to their strong data, talent, and compute foundations. ## Key Information ### Data Acquisition - Meta has access to a vast amount of real-world data through internal employee tracking. - They are now building their own RL environments, which is a key advantage over other companies. - The new data strategy includes hiring experts and using screen recordings to create realistic tasks. ### Talent Strategy - Meta has recruited top talent from OpenAI, Anthropic, and Google. - Notable hires include Shengjia Zhao, Trapit Bansal, Joel Pobar, Jack Rae, and others. - They have also hired Dina Powell McCormick, a prominent figure in finance, to lead their compute expansion. ### Compute Strategy - Meta is constructing multiple large-scale datacenters, including Prometheus (3GW+), Hyperion (400MW), and others in El Paso, Iowa, and Indiana. - They are investing in a new "applied AI engineering org" with ~3000 engineers dedicated to RL task creation. - Their compute strategy is expected to surpass that of OpenAI and Anthropic by the end of 2026. ### Network Infrastructure - Meta has developed AI-Backbone (AIBB), a network architecture designed for massive AI compute clusters. - AIBB includes L3 Superspines and L4 Inter-BAG hubs, providing high-speed, bi-directional bandwidth across their datacenters. - Despite the benefits, the physical distance between datacenters introduces latency, requiring asynchronous training approaches. ### Strategic Implications - The document suggests that Meta's focus and resources position them to potentially overtake OpenAI and Anthropic in the next 6 months. - It also notes that the AI field is becoming increasingly dominated by a few key players, with Meta's unique advantages in data and compute giving them a strong edge. ## Conclusion Meta has made significant strides in the AI field over the past year, leveraging their internal resources, talent acquisition, and infrastructure investments to build a robust foundation for superintelligence. While challenges remain, particularly with network latency and the need for continuous task refinement, the document argues that Meta is well-positioned to compete with and potentially surpass OpenAI and Anthropic in the near future.