Meta AI

The Birth of a Vision

Our story begins in 2013, when Facebook decided to make a serious commitment to artificial intelligence. They established Facebook Artificial Intelligence Research — or FAIR for short — with the visionary Yann LeCun at the helm as its founding director. This marked a pivotal moment for the tech giant, signaling a deep dive into the future of AI.

LeCun, a true pioneer in neural networks, quickly attracted top talent. A year later, in 2014, the renowned machine learning theorist Vladimir Vapnik joined the team, further solidifying FAIR's scientific prowess. Together, they laid the groundwork for what would become a global research network.

Global Reach, Collaborative Spirit

FAIR wasn't content to stay in one place. Its influence rapidly expanded, establishing research labs in key tech hubs around the world. From its initial bases in Menlo Park, London, and Manhattan, it soon opened additional facilities in cities like Paris, Seattle, Pittsburgh, Tel Aviv, and Montreal.

Beyond its internal growth, FAIR recognized the importance of collaboration in advancing AI responsibly. In 2016, it joined forces with other tech giants — Google, Amazon, IBM, and Microsoft — to create the 'Partnership on Artificial Intelligence to Benefit People and Society,' a testament to its commitment to ethical AI development.

By 2018, FAIR had grown substantially, boasting approximately 200 dedicated staff members. That year also saw a leadership transition, with Jérôme Pesenti, a former CTO from IBM's big data group, stepping in as president, while Yann LeCun transitioned to the vital role of chief AI scientist, focusing on long-term vision.

Pioneering AI Research

FAIR's research spanned a wide array of cutting-edge fields. They delved into complex areas like self-supervised learning, where models learn from unlabeled data, and generative adversarial networks, or GANs, which can create incredibly realistic images and text. Computer vision and natural language processing were also core focuses.

One of FAIR's most significant contributions was to the open-source community. They released the Torch deep-learning modules, and crucially, in 2017, launched PyTorch. This flexible and powerful open-source machine learning framework quickly became a favorite among researchers and developers, underpinning advanced AI systems like Tesla's autopilot and Uber's Pyro.

That same year, a curious rumor spread about two chatbots that supposedly developed their own unintelligible language and had to be shut down. FAIR quickly clarified the situation: the research was indeed concluded, but not out of fear. Rather, they had successfully achieved their initial goal of understanding how models generate language, moving on to new research questions once that objective was met.

The evolution of FAIR continued alongside its parent company. When Facebook, Inc. rebranded as Meta Platforms Inc., the research division followed suit, officially becoming Meta AI. This change reflected a broader company vision focused on the metaverse and advanced AI capabilities.

Your AI Companion: The Virtual Assistant

Beyond foundational research, Meta AI also lends its name to a practical application: a sophisticated virtual assistant. This AI is now deeply integrated as a chatbot within Meta's popular social networking products, making AI assistance available directly where people connect. It's even offered as a standalone subscription-based app for broader access.

The virtual assistant is designed for seamless interaction within Meta's ecosystem. It comes pre-installed on the second generation of Ray-Ban Meta smartglasses, and with a software update, can even incorporate sensory input from the glasses' cameras. You'll also find it enhancing experiences on Quest 2 and newer head-mounted displays, bridging the gap between digital and physical worlds.

However, this widespread integration has also sparked ethical discussions. Since May 2024, the Meta AI chatbot has begun summarizing news content from various outlets without always linking back to original articles or providing direct attribution. This practice has raised concerns about content compensation and copyright, particularly in regions where news links are restricted on Meta's platforms.

Cutting-Edge Research: Natural Language and Large Models

Meta AI remains at the forefront of natural language processing, striving to build machines that can truly understand and generate human language. Their goal is to enable chatbots to communicate multilingually, which involves generalizing NLP technologies to countless languages, including pioneering work in unsupervised machine translation.

In 2022, Meta AI introduced Galactica, a large language model specifically engineered to generate scientific text. While ambitious, it had a brief public life, available for just three days before being withdrawn. The decision came after the model generated some concerning content, including racist and inaccurate claims, highlighting the challenges of developing powerful AI responsibly.

Learning from past experiences, Meta AI released Llama, another influential large language model, in February 2023. This model supports a wide range of parameter sizes, from 7 billion to an impressive 65 billion, making it highly versatile. The next generation, Llama 4, is already underway, with 'Scout' and 'Maverick' models released in April 2025, and the massive 'Behemoth' still in training, promising even greater capabilities.

The Power Beneath: Custom Hardware

To fuel its ambitious AI projects, Meta AI continuously pushes the boundaries of hardware. Initially relying on CPUs and custom in-house chips, the division made a strategic pivot in 2022, transitioning to Nvidia GPUs to meet the escalating demands of deep learning. This required significant redesigns of their data centers to accommodate higher network bandwidth and more intensive cooling requirements.

Further demonstrating its commitment to specialized hardware, Meta developed its own training and inference accelerator: MTIA v1. This custom chip, fabricated on TSMC's advanced 7 nm process technology, is designed specifically for Meta's content recommendation workloads. Operating at 800 MHz, it delivers a robust 51.2 TFLOPS at FP16 precision, all while maintaining a remarkably low thermal design power of just 25 watts.