In an exclusive interview ahead of Meta’s AI lab’s 10th anniversary, Yann LeCun discusses the artificial intelligence surge as the tech giant unveils new tools for generating audio, researching video, and identifying biases in language models.
Yann LeCun, Meta’s chief scientist, recalls his reaction when OpenAI, a then-emerging competitor, launched ChatGPT last November. This chatbot, capable of generating cover letters, jokes, scripts, and virtually any text form, propelled AI into the mainstream, sparking a new frenzy in Silicon Valley.
For Yann LeCun, who founded Meta’s AI research lab a decade ago, the astonishment wasn’t about the technology itself—Meta had been developing similar large language models, the technology underpinning ChatGPT, for years. Just weeks before, Meta had introduced Galactica, a chatbot designed for writing scientific papers, which faced criticism on X, formerly known as Twitter. Hence, when ChatGPT was hailed as revolutionary.
“The shock lay in the realization that when such a tool is offered by a smaller company, it garners significant admiration and usage,” Yann LeCun said. “Its impact on public perception and the broader public domain was a major surprise, not just to us but to OpenAI as well.”
This kind of reaction is a common challenge at Meta, the parent company of Facebook, which has been mired in controversies including election meddling, misinformation, and impacts on teen well-being. These scandals have made Meta a prime target for politicians and regulators in Washington, and undermined trust among its platform users. Did this affect how Meta’s research was perceived?
“Yes and no,” Yann LeCun admitted. “Meta is gradually overcoming its image issues,” he added with a laugh. “There’s still some negativity, but it’s improving.”
Regardless, the AI revolution spurred by ChatGPT has ushered in a new era for Meta’s FAIR lab. Established as Facebook AI Research in 2013 (the ‘F’ later changed to “Fundamental” after rebranding to Meta), the unit aimed to lead breakthroughs in AI. A decade later, as AI dominates the tech scene, the lab has become pivotal within one of the world’s most influential corporations.
“The surprise was the public’s overwhelming response to such tools, especially when not from a major tech firm.”Yann LeCun
Over the past year, Meta has shifted towards integrating AI as the core of new products, moving beyond its background support role, LeCun noted. While the lab initially concentrated on image recognition, the rise in generative AI has redirected its focus. In February, Meta launched LLaMA, a large language model competing with OpenAI’s GPT, which has already been adopted by companies like Shopify and DoorDash. Amidst this, FAIR has emerged as a staunch advocate for open-source research, amid industry debates over balancing AI safety with the benefits of collaborative progress. With these developments, Meta aims to roll out more products, particularly aligned with CEO Mark Zuckerberg’s substantial investment in the metaverse.
“The metaverse represents the ultimate, immersive, generative experience,” explained Joelle Pineau, Meta’s vice president of AI research. She envisions digital worlds inhabited by AI-generated characters, their dialogue powered by LLMs, and environments crafted through AI-generated animation.
Zuckerberg, too, has frequently promoted this vision, though he acknowledges it may be a decade away. The current iteration of the metaverse has faced mockery and skepticism. “Mark remains a staunch believer in this long-term vision and the potential of AI,” Pineau stated.
A ‘signal’ of importance
Zuckerberg and former Facebook CTO Mike Schroepfer founded FAIR in December 2013, appointing NYU computer science professor Yann LeCun as its leader. The lab’s ambitious goal was to foster significant advances in AI, inspired by breakthroughs like AlexNet, a neural network architecture essential for image recognition.
Initially, the AI team was strategically placed near Zuckerberg at Facebook’s headquarters, signaling the importance of AI to the company’s future. Schroepfer, who resigned as CTO in 2022 (FAIR also operates in New York, Montreal, Tel Aviv, and London.)
Over the years, FAIR has achieved numerous milestones. PyTorch, launched in 2016, is a machine learning framework that facilitated the development of generative AI models. FastMRI, an AI-driven technology, significantly accelerated MRI processes. Another project, Few-Shot Learner, aimed to quickly eliminate harmful Facebook content, such as vaccine misinformation.
One of FAIR’s notable early projects was a chatbot called M. In August 2015, Facebook introduced this service to about 2,000 Californians, promoting it as a versatile assistant capable of various tasks, like arranging flower deliveries or making restaurant reservations.
“Mark continues to be a firm believer in AI’s long-term prospects.”Joelle Pineau
However, M’s capabilities were somewhat overstated. While Facebook’s AI algorithms addressed some user requests, most tasks were executed by human contractors in traditional ways—like calling florists or manually booking tables online. “The reality was that the software wasn’t even close to handling 80% of user requests, as one might hope. It was far less effective.”
M was initially designed to test what users would expect from chatbots. Yet, Facebook discontinued the project in 2018, making it seem rather modest in the era of ChatGPT.
M exemplified the intersection of research and product development, a focus for the lab from its inception, Schroepfer noted. He expressed regret over not implementing more GPU chips for data training earlier.
Newproducts of FAIR
At its 10th anniversary event, FAIR showcased new products and tools. One of these, Audiobox, allows users to input voice or text prompts to generate corresponding sound clips based on preferred vocal characteristics. The lab also introduced Seamless, a suite of language translation models, including one that retains the original tone, cadence, and volume of a speaker’s voice.
Additionally, FAIR announced a new open-source video dataset, shot from diverse perspectives. This dataset could be used to create AI-driven instructional videos featuring multiple viewpoints. This project is a collaboration between Meta and several universities, including Carnegie Mellon and the University of Pennsylvania.
Furthermore, Meta is releasing ROBBIE, a tool for detecting biases in generative language models. This tool compares AI-generated responses from various LLMs, identifying biases across categories like race, body type, and political ideology. The goal is for developers to evaluate potential harms and balance trade-offs, according to a research paper.
‘Dirty stuff inside your closet’
One of FAIR’s major releases was LLaMA in February, a robust language model rivaling OpenAI’s GPT and Google’s LaMDA and PaLM models. Uniquely, LLaMA 2 is open-source, a decision championed by Zuckerberg following internal debate.
This stance has positioned Meta and FAIR as prominent advocates for open-source research, with the company publishing its code for public and research community use. “The default approach is to aim for open-sourcing every project,” said Pineau. “Keeping things closed can lead to unchecked issues.” However, open-sourcing isn’t a strict policy at FAIR, as demonstrated by their selective approach with the Audiobox model.
The tech industry’s debate over open-source AI involves a trade-off between transparency, scientific collaboration, and safety concerns. Critics, like Connor Leahy of Conjecture, argue that Meta’s open approach is reckless. In contrast, companies like Google’s DeepMind maintain more secrecy to manage risks and prevent abuse.
LeCun dismisses such criticism, stressing the benefits of open-source AI despite its nascent stage. His stance contrasts with DeepMind, which has become less transparent as Google integrates it more closely into its operations. LeCun believes this approach could hinder overall progress in the field.
Meta’s open-source advocacy has prompted speculation that the company is mainly driven by business interests and rivalry with Google and OpenAI. However, LeCun denies this, asserting their motivation is progress and collaboration within the research community.
DeepMind spokesperson Amanda Carl highlighted their contributions to open source, mentioning transformative technologies like the transformer and TensorFlow.
For Meta, the evolving AI landscape brings back familiar challenges regarding safety, content moderation, and well-being, as noted by a former Facebook employee who worked with FAIR. These issues are expected to intensify with the advancement of AI.
Despite these challenges, AI’s role at Meta is expanding, garnering increased interest from Zuckerberg. “He was deeply involved in establishing FAIR a decade ago,” Pineau observed. “After a period of lesser involvement, he’s now engaging more actively again.”