Meta AI

• 1. Introduction

  ›

  Meta AI traces its public profile to a misunderstanding. In 2017, a pair of chatbots built inside the division became the subject of a viral claim. Reports said the two had developed a private language, one incomprehensible to humans. Researchers explained that the experiment had been stopped because it succeeded. The goal had been to understand how language is generated inside their models, and when that goal was met, the project ended. The story that spread was not the story that happened. That gap between public perception and actual practice has run through Meta AI's entire existence. Born as a basic-research arm of Facebook, the division has since grown into one of the most widely deployed AI operations in the world. It raises questions about what it means to pursue open scientific inquiry inside one of the world's largest advertising companies. The workplaces alone suggest the scale. As of 2025, the division operates out of eight cities across three continents, including Menlo Park, London, Tel Aviv, and Montreal.

  71141

• 2. The FAIR Years

  ›

  Facebook Artificial Intelligence Research, abbreviated FAIR, came into existence in 2013. Yann LeCun directed the division through its opening years, until 2018. LeCun's departure brought Jérôme Pesenti to the role; Pesenti had previously served as chief technology officer of IBM's big data group.

  The research agenda FAIR pursued included self-supervised learning, generative adversarial networks, document classification, machine translation, and computer vision. In 2016, the division joined a coalition that included Google, Amazon, IBM, and Microsoft. Together they established the Partnership on Artificial Intelligence to Benefit People and Society.

  When Facebook, Inc. changed its name to Meta Platforms Inc., FAIR was renamed Meta AI. The researchers had, in the years before that rename, built tools that were already running inside systems the division had no direct part in building.

  34579

• 3. Open-Source Infrastructure

  ›

  In 2017, the division released PyTorch as an open-source machine learning framework, alongside Torch deep-learning modules it had developed earlier. Tesla's autopilot system was built using PyTorch. So was Uber's Pyro, a separate deep learning project that drew on the same foundation. The decision to publish the framework as open-source software changed who could use it. Work produced inside a private company became available to any engineer or researcher who wanted it. Galactica, a model designed specifically to generate scientific text, would test the outer limits of what these tools could produce.

  108117

• 4. From Science to Scale

  ›

  On the 15th of November 2022, Galactica opened to public access. Three days later, Meta withdrew it. Reviewers had found that its output contained racist and inaccurate material. The content was not defensible, and the company removed it.

  The experience with Galactica did not stop the division from releasing Llama, a large language model that arrived in February 2023. Llama became the basis for a continuing series of releases. As of January 2026, the most recent version in the series is Llama 4. The models these researchers trained ran on hardware the company had recently overhauled.

  2524262730

• 5. Inside the Machine

  ›

  Before 2022, Meta ran its AI workloads on CPUs and custom chips designed in-house. That year, the company switched to Nvidia GPUs.

  Among the chips the team had previously built was MTIA v1, designed to handle the company's content recommendation algorithms. MTIA v1 was fabricated on TSMC's 7-nanometer process, consumed 25 watts, and was capable of 51.2 teraflops at FP16 precision. The products the division eventually shipped placed this processing power in the hands of ordinary users, including through a device designed to be worn.

  303111

• 6. Wearables and Feeds

  ›

  The virtual assistant bearing the Meta AI name was pre-installed on the second generation of Ray-Ban Meta smartglasses. A later software update added the ability to process input from the glasses' cameras. The assistant also runs on Quest 2 headsets and the newer mixed-reality hardware that followed. Beyond wearables, Meta AI functions as a chatbot inside Meta's social networking platforms and is offered separately as a subscription-based application. The expansion of Meta AI across these platforms coincided with a set of legal and ethical disputes that had been building since at least 2022.

  1718121516

• 7. Data, Access, and Accountability

  ›

  The French media outlet Mediapart reported that Facebook's parent company had illegally used works from the pirate site LibGen. The alleged breach occurred in 2022 and concerned material used to train artificial intelligence systems.

  Since May 2024, Meta AI has summarized content from news outlets without providing links to the original articles. This extended to Canada, where Meta had already prohibited news links on its platforms. The use of news content without compensation or attribution raised ethical and legal questions.

  The company had not publicly resolved those questions when, on the 1st of October 2025, it issued an announcement. The text read: 'We will soon use your interactions with AI at Meta to personalize the content and ads you see.'

  3219131