CtrlK
| Meta AI | |
|---|---|
| * | |
| File:Meta headquarters Menlo Park.jpg | |
| Meta headquarters in Menlo Park, California | |
| Type | Research division |
| Industry | Artificial intelligence Machine learning Computer vision Natural language processing |
| Founded | December 11, 2013 (as Facebook AI Research) 2021 (renamed to Meta AI) |
| Founders | Mark Zuckerberg Yann LeCun Rob Fergus |
| Headquarters | Menlo Park, California, United States (Additional labs: NYC, Paris, Montreal, Seattle, Pittsburgh, Tel Aviv, London |
| Key people | Yann LeCun (Chief AI Scientist) Alexandr Wang (Chief AI Officer) Daniel Gross (AI Executive) Nat Friedman (Head of Products & Applied Research) Chris Cox (Chief Product Officer) Andrew Bosworth (CTO) |
| Parent | Meta Platforms |
| Owner | Meta Platforms, Inc. |
| Products | PyTorch (ML framework) LLaMA (language models) Meta AI Assistant Segment Anything Model Cicero MTIA (AI chips) DINOv3 (vision model) Ray-Ban Meta integration |
| Website | [https://[Expression error: Unexpected < operator. Script error: No such module "String".] [Expression error: Unexpected < operator. Script error: No such module "String".]] |
See also: Meta AI* Meta AI** is a research division of Meta (formerly Facebook) that develops artificial intelligence, augmented reality, and superintelligence technologies. The division encompasses both Meta's AI research efforts through Meta Superintelligence Labs (MSL) and an AI assistant product available across Meta's platforms.[1]
Meta AI was originally founded on December 11, 2013 under the name Facebook Artificial Intelligence Research (FAIR).[2] The foundation of the laboratory was announced by CEO Mark Zuckerberg with the goal of advancing the field of machine learning and artificial intelligence through open research. FAIR was first directed by New York University's Yann LeCun, a deep learning professor and Turing Award winner, alongside co-founder Rob Fergus.[1]
The initial goal of FAIR, working with NYU's Center for Data Science, was to research data science, machine learning, and artificial intelligence.[3] Initial hubs opened in Menlo Park and New York City, specifically at Astor Place.[2] Vladimir Vapnik, a pioneer in statistical learning, joined FAIR in 2014.[3]
FAIR expanded its presence globally by opening research centers in multiple locations:
| Lab | City | Country | Opened |
|---|---|---|---|
| Menlo Park | California | United States | 2013 |
| Astor Place, New York | New York State | United States | 2013 |
| Paris | Île-de-France | France | June 2015 |
| Seattle | Washington | United States | 2016 |
| Pittsburgh | Pennsylvania | United States | 2016 |
| Montreal | Quebec | Canada | September 2017 |
| Tel Aviv | - | Israel | 2019 |
| London | - | United Kingdom | 2020 |
The Paris lab, opened in June 2015, marked the first major U.S. tech AI center in continental Europe.[5] FAIR Montreal began in September 2017 under Joelle Pineau, expanding Canadian federal AI initiatives.[6]
In 2016, FAIR partnered with Google, Amazon, IBM, and Microsoft in creating the Partnership on Artificial Intelligence to Benefit People and Society.[3] By 2018, FAIR had approximately 200 staff members, with Jérôme Pesenti as president and Yann LeCun as chief AI scientist.[4]
Following the rebranding of Facebook, Inc. to Meta Platforms Inc. in October 2021, FAIR was renamed to Meta AI.[7] This change reflected the company's broader focus on the metaverse and advanced AI technologies.
In January 2022, Meta unveiled the Research SuperCluster (RSC), a 5-exaflop AI supercomputer built with 16,000 NVIDIA A100 GPUs, accelerating model training up to 20× versus previous clusters.[8]
In June 2025, following concerns about the company's AI competitiveness, Mark Zuckerberg announced the formation of Meta Superintelligence Labs (MSL), a new division focused on achieving artificial general intelligence and superintelligence.[9]
The formation came after Meta's aggressive acquisition and hiring strategy, including:
A $14.3 billion investment in Scale AI for a 49% non-voting stake, with founder Alexandr Wang joining as Chief AI Officer[10]
Failed attempts to acquire Safe Superintelligence (valued at $32 billion) and Perplexity AI[11]
Recruitment of Daniel Gross (CEO of Safe Superintelligence) and Nat Friedman (former GitHub CEO)[11]
MSL comprises four groups:
TBD Lab - managing Meta's large language models, led by Wang
FAIR - artificial intelligence research team
Products and Applied Research - consumer integration team led by Friedman
MSL Infra - infrastructure team led by Aparna Ramani[12]
One of Meta AI's most significant contributions to the AI community is PyTorch, a machine learning library released in 2017.[13] PyTorch was developed as a successor to the Torch library and quickly gained traction among researchers for its developer-friendly approach and flexibility.[14]
Key milestones in PyTorch's development:
January 2017: Initial public release
March 2018: Merger with Caffe2
December 2018: PyTorch 1.0 release
September 2022: PyTorch Foundation established under the Linux Foundation[15]
March 2023: PyTorch 2.0 release with TorchDynamo
2025: Used by approximately 80% of researchers at major ML conferences[15]
Meta AI has developed the LLaMA (Large Language Model Meta AI) series of large language models:
| Model | Release Date | Parameters | Key Features | License |
|---|---|---|---|---|
| LLaMA 1 | February 2023 | 7B-65B | Foundation models only | Non-commercial |
| LLaMA 2 | July 2023 | 7B-70B | Chat models included, commercial use allowed | Llama 2 Community License |
| LLaMA 3 | April 2024 | 8B-70B | Improved performance | Llama 3 Open License |
| LLaMA 3.1 | July 2024 | 8B-405B | Largest open-source model (405B) | Llama 3 Open License |
| LLaMA 3.2 | September 2024 | 1B-90B | Multimodal capabilities | Llama 3 Open License |
| LLaMA 3.3 | December 2024 | 70B | Optimized performance | Llama 3 Open License |
| LLaMA 4 Scout | April 2025 | 109B (17B active) | 10M token context, MoE architecture | Llama 4 Community License |
| LLaMA 4 Maverick | April 2025 | 400B (17B active) | 1M token context, 128 experts | Llama 4 Community License |
| LLaMA 4 Behemoth | In training | 2T (288B active) | 16 experts, STEM-focused | TBD |
LLaMA 4, released in April 2025, introduced Meta's first mixture-of-experts (MoE) architecture models and native multimodality, supporting text, image, and video inputs.[18] The unreleased Behemoth model, with 2 trillion total parameters, reportedly outperforms GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro on STEM benchmarks.[17]
Meta AI has made significant contributions to computer vision:
Segment Anything Model (SAM): Released in April 2023, SAM segments any object in an image zero-shot and ships with the SA-1B dataset of one billion masks[19]
DINOv3: Released in August 2025, scales self-supervised learning for images to create universal vision backbones[20]
Detectron2: Open-source object detection library
Self-supervised learning techniques: Development of methods that learn from unlabeled data
Meta AI's NLP research focuses on:
Machine translation: Including unsupervised machine translation techniques
Language understanding: Developing systems for natural human-machine interactions
Conversational AI: Creating more engaging and context-aware conversational agents[1]
Protein Structure Prediction: In 2022, Meta AI predicted the 3D shape of 600 million potential proteins in two weeks[1]
HyperTree Proof Search (HTPS): AI system that proved 10 International Mathematical Olympiad problems in Lean in 2022[22]
Galactica: A large language model designed for scientific text generation released in November 2022, but withdrawn within days due to issues with offensiveness and inaccuracy[23]
In addition to research, Meta AI also refers to an AI assistant developed by the division. The Meta AI assistant has been integrated across Meta's platforms:
The Meta AI assistant was rolled out across multiple platforms with the following timeline:
September 27, 2023: Debuted as voice assistant on Ray-Ban Meta (2nd generation) smart glasses[1]
April 23, 2024: Evolved to support multimodal input via computer vision
July 23, 2024: Meta AI with Vision integrated into Horizon OS v68 on Meta Quest 3 and Quest Pro, made available on Quest 2 without Vision[1]
April 29, 2025: Standalone Meta AI app launched with Llama 4 integration[24]
December 2024: Reached nearly 600 million monthly active users[25]
Integration platforms include:
Messenger
Ray-Ban Meta smart glasses
Meta Quest VR headsets
meta.ai website
The assistant offers various capabilities:
Conversational AI powered by LLaMA models
Image generation and editing
Real-time information access (via Google and Microsoft Bing)
Multimodal input via computer vision (on supported devices)
Voice interaction capabilities with full-duplex speech technology
"Discover feed" for sharing and exploring AI prompts[24]
Users can mute the Meta AI assistant on platforms like Facebook, Instagram, and WhatsApp, and hide AI elements via settings. However, the assistant cannot be completely disabled on all platforms.[26]
Meta AI has evolved its hardware infrastructure over time:
| Infrastructure | Details | Year |
|---|---|---|
| Pre-2022 | Primarily used CPUs and custom chips | Before 2022 |
| Research SuperCluster (RSC) | 5-exaflop AI supercomputer with 16,000 NVIDIA A100 GPUs | 2022 |
| MTIA v1 | 7nm chip, 800 MHz, 102.4 TOPS for INT8, 25W TDP, 3× efficiency over GPUs | 2023 |
| MTIA v2 | Training chip for ranking and recommendations | 2025 |
| Prometheus | 1 gigawatt AI supercluster in Ohio | Expected 2026 |
| Hyperion | Up to 5 gigawatt cluster in Louisiana, Manhattan-sized footprint | Multi-year rollout |
Meta's AI infrastructure investment has grown dramatically:
2024: ~$40 billion (approximate)
2025: $66-72 billion (projected)[28]
2026: Expected to exceed 2025 growth rate[29]
The company is exploring partnerships with financial institutions to co-develop data centers while maintaining flexibility for changing infrastructure requirements.[28]
Meta has invested heavily in developing custom AI accelerators:
MTIA v1: First-generation inference accelerator deployed at scale in 16 data center regions
Training chips: Custom silicon for training recommendation systems in production**
Rivos acquisition: Acquired chip startup Rivos in October 2025 to bolster silicon development[30]
Multiple chips in various stages of development expected to deploy in coming years[31]
CEO Mark Zuckerberg argues that open models "make AI safer and more accessible," and Meta regularly releases code, weights and datasets under permissive terms.[32] This philosophy has led to:
Open-sourcing PyTorch, now used by approximately 80% of researchers at major ML conferences[15]
Releasing LLaMA models under increasingly permissive licenses
Publishing research papers and datasets like SA-1B (1 billion segmentation masks)
Contributing libraries such as FAISS and Detectron2 to the community
Enabling developers to use LLaMA models for synthetic data generation
Notable figures associated with Meta AI:
| Name | Role | Tenure | Notes |
|---|---|---|---|
| Yann LeCun | Chief AI Scientist, Co-founder | 2013–present | Turing Award winner, founding director of FAIR |
| Joelle Pineau | VP of AI Research | 2017–2025 | Led Montreal lab and FAIR, departed May 30, 2025[33] |
| Alexandr Wang | Chief AI Officer | 2025–present | Scale AI founder, joined via $14.3B investment deal[10] |
| Daniel Gross | AI Executive | 2025–present | Former CEO of Safe Superintelligence, Apple AI veteran[11] |
| Nat Friedman | Head of Products and Applied Research | 2025–present | Former GitHub CEO, venture investor[11] |
| Jérôme Pesenti | President of FAIR | 2018–? | Former CTO of IBM's big data group |
| Mark Zuckerberg | CEO of Meta | 2013–present | Driving force behind AI strategy and MSL formation |
| Rob Fergus | Co-founder | 2013–? | Computer vision researcher |
| Soumith Chintala | PyTorch Lead | 2017–? | Led PyTorch development |
| Vladimir Vapnik | Researcher | 2014–? | Statistical learning pioneer |
| Chris Cox | Chief Product Officer | Various | Oversees AI integration into products |
| Andrew Bosworth | CTO | Various | Oversees technical strategy |
Meta has engaged in aggressive talent acquisition strategies to compete with rivals:
Offered signing bonuses reportedly as high as $100 million to recruit from competitors[34]
OpenAI CEO Sam Altman confirmed Meta's recruitment attempts with large compensation packages[34]
Successfully recruited researchers from OpenAI, Google DeepMind, and other leading labs
CTO Andrew Bosworth acknowledged the "unprecedented" market rates for AI talent[9]
Meta AI has had substantial impact on the AI research community:
PyTorch Adoption: Used by approximately 80% of researchers at major ML conferences[15]
Open Source Philosophy: Meta's commitment to open-source AI has democratized access to advanced models
LLaMA Models: Provided alternatives to closed-source models from OpenAI and Google
Research Publications: Hundreds of papers advancing the state-of-the-art in AI
Industry Influence: Set standards for open AI research in corporate settings
User Reach: Meta AI assistant used by nearly 600 million monthly active users as of December 2024[25]
In November 2022, Meta AI released Galactica, a large language model designed for generating scientific text. The model was withdrawn within three days due to issues with offensiveness, inaccuracy, and fabricated scientific content.[23]
Meta's use of public data for AI training has raised privacy concerns, particularly in the European Union where users can opt out of data collection for AI purposes[26]
In May 2024, the Meta AI chatbot was summarizing news articles without providing direct links to original sources, raising ethical and legal concerns, particularly in Canada where news links are banned on Meta's platforms[1]
Reports in 2025 suggested Meta may have used pirated books for training AI models, raising intellectual property concerns[35]
High-profile departures including Joelle Pineau raised questions about research direction[33]
Tension between fundamental research (FAIR) and product-focused development (GenAI team)[34]
Some insiders claimed in 2025 that the research lab is "dying a slow death" due to focus shift[36]
The April 2025 release of LLaMA 4 received mixed reactions:
Allegations of inflated performance metrics
Reports of a rushed release
Concerns about transparency
Meta has denied claims of gaming performance metrics and attributed mixed performance to early bugs.[37]
In May 2025, Meta announced plans to replace human reviewers with AI systems for assessing privacy and societal risks, raising safety concerns among employees[38]
The company has faced scrutiny over content moderation approaches and AI safety protocols[35]
Meta AI's stated goals include:
Achieving "personal superintelligence for everyone" through MSL[9]
Developing AI agents capable of complex reasoning and action[18]
Reaching one billion Meta AI chatbot users[25]
Continuing investment in AI infrastructure with plans for multi-gigawatt data centers[28]
Advancing AI glasses and wearables as ideal form factors for AI[29]
Hosting LlamaCon, Meta's first AI developer conference (April 29, 2025)[39]
Artificial General Intelligence
Large Language Models
PyTorch
Meta Platforms
Meta Superintelligence Labs
Partnership on AI
Scale AI
Safe Superintelligence
