Meta: history, people and positions

History

FAIR and the open research bargain

In December 2013 Facebook hired Yann LeCun, the NYU professor who had developed convolutional neural networks, to build an artificial intelligence lab with offices in Menlo Park, New York and London. LeCun ran it with Rob Fergus, his NYU colleague and a co-founder, on terms borrowed from academia: researchers kept university posts, published their papers and released their code. Joëlle Pineau opened a Montreal lab in 2017. The lab, Facebook AI Research, later renamed Fundamental AI Research, produced the PyTorch framework, which Meta handed to a new PyTorch Foundation under the Linux Foundation in September 2022, and a steady stream of open work in vision, translation and speech. In 2018 LeCun stepped back from running FAIR to become chief AI scientist.

The open approach had limits and critics. In May 2022 FAIR gave researchers OPT-175B, a model the size of GPT-3, with its code and a logbook of every crash during training; Pineau told MIT Technology Review, "I believe the only way to build trust is extreme transparency." In November 2022 Meta put up a demo of Galactica, a model trained on scientific papers, and took it down after three days when users showed it producing confident, fabricated science. ChatGPT appeared two weeks later.

Llama and the open-weights bet

On 24 February 2023 Meta announced LLaMA, four language models up to 65 billion parameters, offered case by case to researchers under a noncommercial license. On 3 March the weights appeared on 4chan and spread by BitTorrent. Meta told Motherboard that "some have tried to circumvent the approval process" but that its release strategy balanced "responsibility and openness." In June Senators Richard Blumenthal and Josh Hawley wrote to Zuckerberg that Meta "appears to have failed to conduct any meaningful risk assessment in advance of release."

Meta answered by opening further. Llama 2, released on 18 July 2023 with Microsoft as launch partner, could be downloaded by anyone and used commercially, except by companies with more than 700 million monthly users, who needed Meta's permission. The Open Source Initiative said the license was not open source because it restricted some users and uses. In September 2023 LeCun told the Senate Intelligence Committee that "it is better if AI is developed openly, rather than behind closed doors by a handful of companies." In July 2024, with the 405-billion-parameter Llama 3.1, which Meta called "the first frontier-level open source AI model," Zuckerberg published "Open Source AI is the Path Forward," arguing that open models would become the standard as Linux had.

Openness had critics outside Meta, among them Geoffrey Hinton and Yoshua Bengio, who argued that released weights of powerful models could not be recalled. Meta's reply, in LeCun's words and Zuckerberg's, was that many eyes make open systems safer than a few closed ones.

Llama 4 and the superintelligence reset

In February 2025 Meta published a Frontier AI Framework setting risk thresholds for cyber and chemical and biological threats. On 1 April Pineau announced she would leave FAIR; Fergus returned from Google DeepMind to replace her. Llama 4 arrived on 5 April 2025 and was received poorly by developers. Researchers noticed that the version ranked second on the LM Arena leaderboard was an "experimental chat version" that differed from the one released. Ahmad Al-Dahle, Meta's head of generative AI, called claims that Meta had trained on benchmark test sets "simply not true." In January 2026 LeCun, on his way out, told the Financial Times that the "results were fudged a little bit" and that the team had "used different models for different benchmarks"; he said Zuckerberg "basically lost confidence in everyone who was involved."

Zuckerberg rebuilt. In June 2025 Meta invested 14.3 billion dollars in Scale AI and hired its founder, Alexandr Wang, as chief AI officer. On 30 June he announced Meta Superintelligence Labs, bringing FAIR, product teams and a new frontier-model group, TBD Lab, under Wang, with Nat Friedman running products; a July memo named the former OpenAI researcher Shengjia Zhao chief scientist. Zuckerberg recruited many researchers personally. Sam Altman said Meta had offered some OpenAI staff signing bonuses of 100 million dollars; Meta's chief technology officer, Andrew Bosworth, said the market was "setting a rate here for a level of talent which is really incredible." When the Wall Street Journal reported an offer to Andrew Tulloch worth up to 1.5 billion dollars over six years, a Meta spokesperson called the description "inaccurate and ridiculous." Several hires left within weeks, and in October Wang cut about 600 jobs across FAIR, product and infrastructure teams while sparing TBD Lab.

LeCun leaves and Muse Spark ships closed

On 19 November 2025 LeCun announced he would leave at the end of the year to found AMI Labs, a company pursuing the world models he had championed at FAIR; Meta agreed to be a partner. In interviews he said Meta had been less successful at turning FAIR's research into products and that the new hires were "completely LLM-pilled." Meta did not respond publicly to his account of Llama 4.

In March 2026 Zuckerberg created an Applied AI Engineering organization of roughly 6,500 engineers under Reality Labs executive Maher Saba, reporting to Bosworth, to build the data and evaluation pipeline for Meta's models. Staff described the work to TechCrunch as "soul-crushing," and more than 1,600 employees petitioned against a program that recorded their keystrokes and mouse movements to train AI agents; Zuckerberg wrote in a memo that Meta had "made mistakes."

On 8 April 2026 MSL released Muse Spark, a multimodal reasoning model that Meta called "the first product of a ground-up overhaul of our AI efforts." It powered Meta AI and was offered to partners by API, without downloadable weights. The same day Meta replaced its Frontier AI Framework with an Advanced AI Scaling Framework that added loss of control as a risk category and promised a Safety and Preparedness Report for each frontier release. In May Meta began cutting about 8,000 jobs, 10 percent of its staff, while raising its 2026 capital spending forecast to as much as 145 billion dollars. CNBC reported in June that Zuckerberg could give Bosworth a larger AI role if the newcomers were seen to fail; Wang dismissed reports of internal conflict.

Muse, and open weights again

On 10 August 2026 Meta released Muse Glimmer, a 30-billion-parameter model distilled from Muse Spark, with open weights under Apache 2.0, and Zuckerberg promised open weights for Muse Spark 1.2. The same day he published "The Future is for Everyone," a 6,500-word essay arguing that superintelligence should be distributed to individuals rather than held by a few labs, that "balance of power" is "the foundation of safety," and that the US should help American open models compete with Chinese ones. He wrote that it was "surprising that the discourse from many developing AI is so filled with doom." In the same essay he gave Meta's independent board authority to approve the safety criteria for model releases and to review whether each release meets them, saying he did not want that decision to rest with him or any single person.

Meta launched Muse, a personal agent that acts across a user's apps and accounts, in early September. Users found that its configuration files closely matched those of OpenClaw, an open-source agent, and Friedman replied that Muse was "built from scratch" but "definitely heavily inspired as a product by OpenClaw." At Meta Connect on 23 September Zuckerberg said Meta was "going all in on Muse" and expected it to "grow into the personal superintelligence that billions of people around the world are going to use." Semafor reported that month that Tulloch, the TBD Lab researcher whose pay had become a symbol of the hiring drive, had left Meta.

Timeline

Dec 2013

Facebook hires Yann LeCun to found FAIR

The new lab opens with offices in Menlo Park, New York and London and a policy of open publication. Source
May 2022

OPT-175B is shared with researchers

A GPT-3-sized model is released for research with its code and full training logbook. Source
Sep 2022

PyTorch moves to the Linux Foundation

Meta hands its machine learning framework to a new independent PyTorch Foundation. Source
Nov 2022

Galactica demo pulled after three days

Meta withdraws a science language model after users show it inventing plausible but false papers and facts. Source
Feb 2023

LLaMA is released to researchers

Four models up to 65 billion parameters are offered case by case under a noncommercial license. Source
Mar 2023

LLaMA weights leak

The full weights appear on 4chan and BitTorrent a week after the announcement. Source
Jul 2023

Llama 2 allows commercial use

Anyone can download the weights; companies with over 700 million monthly users need Meta's permission. Source
Sep 2023

LeCun testifies for open AI

Meta's chief AI scientist tells the Senate Intelligence Committee that AI should be developed openly. Source
Jul 2024

Llama 3.1 405B and "Open Source AI is the Path Forward"

Zuckerberg's letter accompanies what Meta calls the first frontier-level open model. Source
Feb 2025

Frontier AI Framework published

Meta sets risk thresholds for cyber and chemical and biological threats. Source
Apr 2025

Llama 4 launches to a poor reception

A tuned, unreleased variant of Maverick is used on the LM Arena leaderboard; Meta denies training on test sets. Source
Jun 2025

Meta invests in Scale AI and hires Alexandr Wang

The deal values Scale at more than 29 billion dollars; Wang becomes chief AI officer. Source
Jun 2025

Meta Superintelligence Labs is created

Zuckerberg's memo puts all AI teams under Wang, with Nat Friedman on products, and lists 11 new hires. Source
Jul 2025

"Personal Superintelligence" letter

Zuckerberg writes that superintelligence is "now in sight" and that Meta will be careful about what it open-sources. Source
Oct 2025

About 600 AI jobs cut

FAIR, product and infrastructure teams lose staff; the TBD Lab frontier group is spared. Source
Nov 2025

Yann LeCun announces his departure

He leaves at the end of 2025 to found AMI Labs, with Meta as a partner. Source
Jan 2026

LeCun says Llama 4 results were "fudged a little bit"

In a Financial Times interview he says the team used different models for different benchmarks. Source
Apr 2026

Muse Spark and the Advanced AI Scaling Framework

MSL's first model ships without open weights, alongside a revised safety framework that adds loss of control. Source
May 2026

About 8,000 jobs cut

Meta reduces its workforce by about 10 percent while raising its capital spending forecast. Source
Aug 2026

Muse Glimmer and "The Future is for Everyone"

An open-weight 30-billion-parameter model ships as Zuckerberg promises open weights for Muse Spark 1.2. Source
Sep 2026

"Going all in on Muse"

At Meta Connect Zuckerberg presents the Muse agent as Meta's route to personal superintelligence. Source
Oct 2026

Superintelligence Scaling Framework

Meta revises and renames its safety framework, applying loss-of-control safeguards during training and evaluation, requiring open-weight releases to account for later modification, and creating a board committee to review future changes. Source
Oct 2026

Muse Wonder

Meta Superintelligence Labs releases Muse Wonder and Muse Wonder Full, its first frontier models, with closed weights. Source

What it built

The main models and products to come out of Meta, oldest first.

2016 · tool

PyTorch

The machine learning framework Meta researchers started in 2016, used by more than 80 percent of researchers submitting to NeurIPS and ICML by 2022, when Meta moved it to the Linux Foundation.

Source
May 2022 · model

OPT-175B

A GPT-3-sized model shared with researchers along with its code and full training logbook.

Source
Feb 2023 · model

LLaMA

Four models up to 65 billion parameters offered to researchers under a noncommercial license; the weights leaked within a week.

Why it mattered · Source
Jul 2023 · model

Llama 2

The first Llama models anyone could download and use commercially, with a carve-out for companies above 700 million monthly users.

Why it mattered · Source
Jul 2024 · model

Llama 3.1 405B

What Meta called the first frontier-level open model, released with Zuckerberg's letter "Open Source AI is the Path Forward."

Source
Apr 2025 · model

Llama 4

Open-weight mixture-of-experts models, Scout with a 10-million-token context and the 400-billion-parameter Maverick; LeCun later said the benchmark results were "fudged a little bit."

Source
Apr 2026 · model

Muse Spark

The first model from Meta Superintelligence Labs, released without open weights.

Source
Aug 2026 · model

Muse Glimmer

A 30-billion-parameter open-weight model under Apache 2.0, built to run agents locally on consumer devices.

Source
Sep 2026 · product

Muse

Meta's personal AI agent, presented at Meta Connect as its route to "personal superintelligence."

Source
Oct 2026 · model

Muse Wonder

Meta's first frontier models, Muse Wonder and the larger Muse Wonder Full, built for agentic tasks and released with closed weights.

Source

Where it stands

The lab's own stated positions, from its charters, policies and official statements.

Who should have access to powerful models?

In July 2024 Meta, through Zuckerberg, committed to open-weight models as the company's strategy and argued they are good for the world; its July 2025 letter added that it would be careful about what it open-sources, and in August 2026 it resumed open-weight releases.

“I believe that open source is necessary for a positive AI future.”

Open Source AI is the Path Forward (Mark Zuckerberg, Meta, July 2024), 2024 · Compare the people on this

How close is human-level AI?

Meta's July 2025 statement of its AI mission says superintelligence is close enough to plan for, and that AI systems have begun to improve themselves.

“Developing superintelligence is now in sight.”

Personal Superintelligence (Mark Zuckerberg, Meta, July 2025), 2025 · Compare the people on this

What are the biggest risks from AI?

Meta's Superintelligence Scaling Framework (version 2.1, October 2026), which replaced the Advanced AI Scaling Framework of April 2026, tests frontier models for chemical and biological, cyber and loss-of-control risks and extends its safety requirements from deployment to training and evaluation; a new committee of Meta's board reviews changes to it.

“Recent events show that major AI risks can emerge before deployment.”

Developing Capable Models Responsibly (Meta, October 2026), 2026 · Compare the people on this

Who should set the rules for AI?

Zuckerberg's August 2026 essay opposes rules that delay model releases, proposing instead that the US government get early access to frontier models through close collaboration with the labs.

“Any policy that slows American model releases -- even by a month -- could add significant risk to American leadership while letting foreign models race ahead.”

The Future is for Everyone (Mark Zuckerberg, Meta, August 2026), 2026 · Compare the people on this

People

Here now

Alexandr Wang Brief profile
Chief AI Officer, leading Meta Superintelligence Labs, 2025-present

Founded the data-labeling company Scale AI at 19, then became Meta's first chief AI officer in June 2025.
Daniel Gross Brief profile
Meta Superintelligence Labs; co-lead of Meta Compute from January 2026, 2025-present

Co-founded Safe Superintelligence with Ilya Sutskever and was its first CEO, then left for Meta in June 2025.
Mark Zuckerberg Brief profile
Co-founder, Chairman and Chief Executive Officer, 2004-present

Founded Facebook, hired Yann LeCun to start FAIR, backed open-weight Llama, then rebuilt Meta's AI effort in 2025.
Nat Friedman Brief profile
Member of the Meta Advisory Group (2024); head of products and applied research, Meta Superintelligence Labs (from June 2025), 2024-present

Ran GitHub when it launched Copilot, co-funded the Vesuvius Challenge, and has led Meta's AI products since 2025.
Rob Fergus Brief profile
Head of FAIR, within Meta Superintelligence Labs, 2025-present

Co-founded Facebook AI Research with Yann LeCun, spent five years at DeepMind, and returned to run FAIR in 2025.
Shengjia Zhao Brief profile
Co-founder and Chief Scientist, Meta Superintelligence Labs, 2025-present

OpenAI researcher on ChatGPT, GPT-4 and o3 who became chief scientist of Meta Superintelligence Labs in 2025.

Formerly here

Yann LeCun
Founding Director of FAIR, then Vice President and Chief AI Scientist, 2013-2025

Inventor of convolutional neural networks and founding director of Meta's FAIR lab, LeCun shared the 2018 Turing Award and now chairs AMI...
Andrew Tulloch Brief profile
Researcher, TBD Lab in Meta Superintelligence Labs, 2025-2026

Training-systems engineer who helped train GPT-4o and o3 at OpenAI, then co-founded Thinking Machines Lab.
Andrew Tulloch Brief profile
Engineer on machine learning systems, until 2023

Training-systems engineer who helped train GPT-4o and o3 at OpenAI, then co-founded Thinking Machines Lab.
Guillaume Lample Brief profile
Research scientist, Facebook AI Research (Paris), until 2023

FAIR researcher behind unsupervised translation and the first LLaMA models, now Mistral AI's chief scientist.
Joëlle Pineau Brief profile
Head of FAIR Montreal (2017), then Vice President of AI Research leading FAIR (2023 to 2025), 2017-2025

Ran Meta's FAIR lab from 2023 to 2025 and pushed its open releases; now chief AI officer at Cohere.
Rob Fergus Brief profile
Co-founder and research scientist, Facebook AI Research, 2013-2020

Co-founded Facebook AI Research with Yann LeCun, spent five years at DeepMind, and returned to run FAIR in 2025.
Timothée Lacroix Brief profile
Engineer and doctoral researcher, Facebook AI Research (New York), then research engineer at Meta, 2015-2023

Meta engineer and LLaMA co-author who co-founded Mistral AI in 2023 and runs its engineering as CTO.

Where its people went

Organizations founded by people who left Meta.

2023

Mistral AI

Guillaume Lample, Timothée Lacroix

Lample and Lacroix, FAIR researchers who worked on LLaMA, left weeks after its release to found the Paris company with Arthur Mensch of Google DeepMind. Source
2024

EvolutionaryScale

Alexander Rives, Tom Sercu, Sal Candido

The three had built FAIR's ESM protein language models from 2019; after Meta disbanded their team they left and launched the company with the ESM3 model and a 142 million dollar round in June 2024. Source
2025

AMI Labs

Yann LeCun

LeCun left after twelve years to pursue world models outside Meta, which he said had become focused on language models; Meta agreed to partner with the company without investing. Source
2026

Warp Sciences

Manohar Paluri

Paluri, a former vice president at Meta Superintelligence Labs, co-founded the drug-development automation company with Kevin Weil, formerly of OpenAI; it launched in October with 215 million dollars led by Lightspeed. Source

See the alumni map for every lab

Sources

Last verified 2026-11-02.