Meta AI

• 1. Founding And Evolution

  ›

  Facebook Artificial Intelligence Research opened its doors in 2013. This division emerged from a company that would later change its name to Meta Platforms Inc. The team established workspaces across the globe by 2025, including locations in Menlo Park, London, and Paris. Yann LeCun directed the group until 2018 when Jérôme Pesenti took over leadership. Pesenti previously served as CTO of IBM's big data group. In 2016, the organization joined forces with Google, Amazon, IBM, and Microsoft. They formed the Partnership on Artificial Intelligence to Benefit People and Society. This collaboration aimed to guide ethical development in the field. The unit eventually adopted the name Meta AI following the corporate rebranding.

  12

• 2. Global Research Infrastructure

  ›

  The division maintains physical offices in Seattle, Pittsburgh, Tel Aviv, and Montreal alongside its US headquarters. These international hubs support research into self-supervised learning and computer vision. Teams also focus on generative adversarial networks and document translation. The infrastructure supports large-scale experiments requiring significant computational power. Partnerships extend beyond domestic borders to include major technology corporations globally. Strategic alliances allow for shared resources and accelerated innovation cycles. The network spans multiple continents to facilitate diverse perspectives on artificial intelligence challenges.

  34

• 3. Open Source Frameworks

  ›

  Torch deep-learning modules appeared before PyTorch launched in 2017. This open-source machine learning framework became widely adopted by researchers worldwide. Tesla utilized similar technologies for their autopilot systems while Uber employed Pyro for logistics. A pair of chatbots generated rumors about being discontinued after developing unintelligible language. Researchers clarified that the project ended because they achieved their goal of understanding language generation. Fear was not the reason for shutting down the experiment. The framework remains a cornerstone for modern deep learning applications today.

  7910

• 4. Virtual Assistant Integration

  ›

  Meta AI functions as a virtual assistant integrated into social networking products since May 2024. It summarizes news from various outlets without linking directly to original articles. Canadian users face restrictions where news links are banned on these platforms. The system is pre-installed on second-generation Ray-Ban Meta smartglasses. Camera inputs from the glasses can be incorporated following software updates. Quest 2 headsets and newer HMDs also support the subscription-based stand-alone app. On the 1st of October 2025, Facebook announced plans to use interactions with AI to personalize content and ads.

  161712

• 5. Large Language Model Development

  ›

  Galactica targeted scientific text generation but existed for only three days starting the 15th of November 2022. The model withdrew quickly after generating racist and inaccurate content. Llama arrived in February 2023 as another large language model release. By January 2026, the most recent version reached Llama 4 status. These models represent significant steps in natural language processing capabilities. Teams continue researching unsupervised machine translation and multilingual chatbots. Ethical concerns persist regarding how these systems handle sensitive information during training phases.

  222325

• 6. Custom Hardware Strategy

  ›

  Meta utilized CPUs and in-house custom chips before switching to Nvidia GPUs in 2022. MTIA v1 represents an early proprietary chip designed for content recommendation algorithms. This hardware runs on TSMC's 7 nm process technology consuming 25W of power. It delivers 51.2 TFlops FP16 performance for specific tasks. The transition reflects a strategic shift away from external suppliers toward internal solutions. Custom silicon allows greater control over data processing speeds and efficiency metrics. Content recommendation algorithms benefit directly from this specialized hardware architecture.

  27

• 7. Legal And Ethical Controversies

  ›

  Mediapart reported that Facebook used works from pirate site LibGen to train artificial intelligence in 2022. This action raised allegations of illegal data scraping and copyright violations. The French media outlet highlighted the ethical implications of using unauthorized sources. Meta continues to reduce news visibility on its platforms while facing legal scrutiny. Questions remain about compensation and attribution for creators whose work appears in datasets. These controversies challenge the company's commitment to responsible AI development practices globally.

  28