Mark Zuckerberg dislikes the fact that Meta is playing catch up.
Before a decade, the company’s founder and chief executive officer recognized the potential of artificial intelligence and made substantial investments in its development. He recruited Yann LeCun, one of its earliest visionaries, to lead the charge. Now, only a few short months after OpenAI’s ChatGPT debuted on the consumer market, Meta is falling behind in the exact same technology.
After spending years prioritizing academic discoveries and freely sharing them while struggling to leverage on their commercial potential, Meta is now scrambling to refocus its resources on producing AI products and features, including its own chatbots.
That’s a tall order, given that many of Meta’s top AI employees have left the company, and in the midst of the company’s own rounds of cutbacks during what Zuckerberg has termed a “year of efficiency.” According to an analysis by the Wall Street Journal, approximately one-third of Meta employees who co-authored published AI research related to large language models—the complex systems that power AI systems such as ChatGPT—have departed the company in the past year.
Zuckerberg and other senior executives have assumed greater control over the organization’s AI strategy. They established a new generative AI division that reports directly to Chief Product Officer Chris Cox, one of Meta’s most experienced and respected executives. The team is training generative AI models, which generate content such as text, images, and audio, for incorporation into “every single one of our products,” according to Zuckerberg. He has praised LLaMA, Meta’s flagship AI language model, which, after its code was leaked, sparked the development of homegrown tools that could one day compete with the products that Google and OpenAI are attempting to sell.
If Meta is successful in commercializing its AI initiatives, it could increase user engagement, improve the metaverse, and make the company more appealing to young users, who are currently difficult to acquire. If Meta cannot capitalize on this technology quickly enough, it risks losing relevance as competitors, such as a rapidly expanding number of scrappy AI firms, surge ahead.
Meta’s Vice President of AI Research, Joelle Pineau, stated in a statement that the company is not lagging in AI and defended the company’s emphasis on research and structure, stating that it will position Meta for success. Pineau stated that Meta’s AI research unit is “one of the world’s leading destinations for AI researchers and open science,” and that its research output has increased dramatically over the past year. “Our research discoveries have provided a solid foundation upon which to build a new class of generative AI-powered experiences for our app family. We are proud of the contributions that Meta’s past and present AI researchers are making to help influence the future of cutting-edge AI.”
Zuckerberg announced on Friday the Voicebox AI model, which can read audibly text prompts in a variety of ways and correct audio recordings using text prompts to remove background noise, such as a dog’s bark. Meta did not specify when the public will have access to the research endeavor.
This article is based on more than a dozen interviews with current and former Meta employees, evaluations of LinkedIn and social media profiles, and startup news announcements.
Zuckerberg and other executives have referred to artificial intelligence as the third limb of Meta’s stool, considering it crucial to the company’s long-term growth and relevance, alongside global connectivity and virtual/augmented reality. Falling behind in AI risks making Meta appear stodgy and slow, as opposed to the agile, aggressive upstart that coined the phrase “move fast and break things” and set the pace for innovation in Silicon Valley.
The White House did not invite Meta to a summit of AI leaders in May, which was described as a gathering of “companies at the forefront of AI innovation.”
Meta has taken sharp turns in the past when it appeared to be falling behind, such as when it shifted Facebook’s ad business from desktop to mobile-first or when it launched Instagram’s Stories feature in 2016 to lure users away from Snapchat, which had introduced a similar feature a decade earlier.
Meta faces additional strategic, political, and monetary obstacles. Longstanding emphasis on original research in the AI division of Meta disincentivizes work on generative AI systems, such as ChatGPT, which generate humanlike text and media. In designing the hardware required to operate such AI programs, executives made a mistake, which the company is now attempting to rectify. As a result of years of scrutiny regarding the company’s management of user data and human-rights violations, some executives are hesitant to launch new AI products for consumers.
Meta began investing in AI in 2013. Zuckerberg and then-CTO Mike Schroepfer sought to hire one of the foremost AI experts to head a new research division dedicated to advancing the technology. They found their lieutenant in LeCun, a professor at New York University whose pioneering research in the discipline was renowned.
Deeply rooted in academia and basic research, LeCun was instrumental in establishing a company culture that reflected his priorities: employing scientists over engineers and prioritizing academic outputs, such as research papers, over product development for the company’s end customers. The company’s ability to commercialize its innovations was hampered by the strategy, according to individuals with knowledge of the situation.
It also encouraged a diffuse, bottom-up approach to directing research and allocating resources. People reported that researchers pursued independent initiatives in a variety of directions rather than a unified enterprise-wide strategy. Meta divided hardware into small pools for each project. if given more computer chips than they needed, some researchers would engage in unnecessary tasks to avoid giving them up, according to some individuals.
Meta was tardy to equip its data centers with the most potent computer chips required for AI development. Even as the company acquired more of these processors, it lacked an effective method for distributing them to engineers and researchers. At times, tens of thousands of coveted and expensive hardware remained idle, according to some of the individuals.
Meta is currently revamping its data centers, which may have contributed to the bottlenecks. According to a company blog post, Meta’s most recent supercomputer for AI initiatives has 16,000 of these chips as of May.
As large language models began to demonstrate increasingly impressive capabilities in 2020, tension grew within Meta’s AI research division between those who urged the company to invest seriously in the industry’s new direction and those, including LeCun, who believed such models lacked scientific value. Internally and publicly, LeCun’s opposition to large language models (he believes they don’t bring AI closer to human-level intelligence) made it difficult for researchers with opposing views to amass the support and vast resources required for such projects, according to some of the people.
The Open Pretrained Transformer (OPT) language model was created in 2022 with approximately 1,000 chips, and the premier LLaMA model was created in 2023 with approximately 2,000 chips. In contrast, the industry standard is 5,000 to 10,000 processors. Executives cite the LLaMA leak as a prominent example of Meta’s desire to share its AI technology. LLaMA was initially made available to a select group of outside researchers before it was leaked online.
In the past year, Meta has lost a large number of AI researchers who worked on these and other important generative AI projects, with many citing exhaustion or a lack of confidence in Meta’s ability to remain competitive. According to their LinkedIn profiles and sources with knowledge of the situation, six of the fourteen authors listed on the LLaMA research paper have departed or announced they will be leaving. Eight of the 19 co-authors for the OPT paper have also departed.
Since OpenAI released ChatGPT in November of last year, the rate of departures has accelerated. Some have been enticed by the AI startup craze, which has prompted personnel changes across Silicon Valley, including at Google. The number of job postings on LinkedIn that mention GPT has increased by 79% year-over-year as of March, according to the professional social network.
A spokesperson for Meta stated that the company has continued to recruit and acquire new AI talent.
Following the launch of ChatGPT, Zuckerberg and Cox joined Chief Technology Officer Andrew Bosworth in supervising all AI-related company efforts. The three executives now devote hours per week to AI, attending meetings and approving AI-related initiatives.
Instead of conducting scientific research, the new generative AI division focuses solely on the development of useful products and tools. It received more than 2,000 internal applications and has quickly amassed hundreds of individuals from various organizations. Hardware resources have been transferred from the AI research division and are now being used to train new generative AI models, according to individuals with knowledge of the matter.
In March, Zuckerberg stated that “advancing AI and building it into every one of our products” was the company’s largest investment. Zuckerberg stated at Meta’s annual shareholder meeting in May that the company aims to extend the technology to the metaverse.
Earlier this month, at a town hall meeting with employees, Zuckerberg announced a number of generative AI products that the company is currently developing, according to a spokesperson for Meta. The initiatives include AI agents for Messenger and WhatsApp, AI stickers that users can generate from text prompts and share in chats, and a photo generation feature that will enable Instagram users to modify their own photographs using text prompts and then share them in Instagram Stories.
Zuckerberg also presented some internal-only generative AI tools for employees, such as Metamate, a productivity assistant that draws information from internal sources to complete tasks at the request of employees. A significant number of employees were recently given access to Metamate as part of a trial run, according to a spokesperson for Meta.
During a town hall, Zuckerberg stated, “In the past year, we’ve witnessed some truly incredible breakthroughs — qualitative breakthroughs — in generative AI.”
Meta still confronts significant obstacles. After seven years of intense government and media scrutiny for its user-privacy practices, the company’s increasingly low tolerance for risk has created friction regarding how and when to introduce AI products, according to people familiar with the matter.
In the past, when developing and publishing large language models that are prone to producing incorrect answers or offensive comments, Meta had to consider its public reputation.
Several years ago, AI researchers were reportedly working on a chatbot codenamed Tamagobot, which was based on an early iteration of a large-language-model system. The team was impressed by its performance but determined that it was not worth launching while the company was facing intense criticism for allowing misinformation to thrive on its platform during the 2016 presidential election, according to one of the individuals.
When Meta released its BlenderBot 3 chatbot in August 2022, the concern over public scrutiny was also evident. BlenderBot 3 was criticized within a week of its release for making fraudulent statements, offensive remarks, and racist comments. Additionally, the system described Zuckerberg as “creepy and manipulative.”
The spokesperson for Meta stated that the project remained online for more than a year until the conclusion of the research, and that the company maintained an open and transparent approach throughout its life cycle. He added that Meta has released and completed numerous other initiatives that demonstrate the company’s willingness to take risks.
In November 2022, however, the scenario was repeated when the corporation released Galactica, a science-focused large language model. Due to its incorrect and biased answers, the system was shut down by Meta within three days of its publication, following a wave of criticism by scientists.
After two weeks, OpenAI released ChatGPT.
