Shulou(Shulou.com)11/24 Report--

Beijing, June 19 (Xinhua) in the battle for artificial intelligence (AI), Mark Zuckerberg (Mark Zuckerberg) got up early in the morning, but caught a late episode. Now, he is doing something he doesn't like: catch up.

As early as a decade ago, the Facebook founder saw the future of AI and invested heavily to promote the development of his own AI. He hired Yann LeCun, one of AI's early visionaries and known as the "godfather of AI", to lead the business. Now, just a few months after OpenAI's ChatGPT entered the consumer market, Meta is lagging behind in the same technology.

A decade of first-mover advantage in 2013, Facebook (now renamed Meta) began investing in AI. Zuckerberg and Mike Schroepfer, then chief technology officer, personally sought to hire a leader in AI to lead a new research division of the company and advance AI development. The deputy they found was New York University professor Yang Likun, who made a prestigious breakthrough in the field of AI.

Yang Likun is deeply rooted in academic and basic research, advocating a culture that reflects his priorities: hiring scientists rather than engineers, and emphasizing academic achievements, such as research papers, rather than developing products for the company's end users. Over the years, this strategy has made Meta's basic AI research lab attractive to top talent, but it has also challenged the company's ability to commercialize its technological achievements, according to people familiar with the matter.

Meta encourages a decentralized, bottom-up approach to research and resource allocation. Researchers push their own agendas and pursue independent projects in different directions, rather than developing a unified company-wide strategy, according to people familiar with the matter. Meta splits hardware resources into small chunks and allocates them to each project. Some researchers get more computer chips than they need, and they use them for unnecessary tasks to avoid giving them up.

Meanwhile, Meta has been slow to equip its data centers with the most powerful computer chips needed for AI development. Even if the company gets more of these chips, it doesn't have a good mechanism to send them to engineers and researchers. Some say there are sometimes thousands of coveted pieces of expensive hardware lying idle.

Meta is overhauling its data center, which may be causing some dysfunctions. As of May, Meta's latest AI supercomputer had 16000 such chips, according to a Meta blog post.

Tensions within the Meta AI research department are also rising as the big language model begins to show more and more capabilities in 2020, according to people familiar with the matter. The two camps are divided: one side urges the company to seriously invest in the new direction of the industry, while people, including Yang Likun, see these models as a short-lived frenzy and lack of scientific value. Some people have revealed that Yang Likun is strongly opposed to the big language model both within the company and in public, believing that they will not bring AI closer to human intelligence, which makes it difficult for dissenting researchers to get the support and resources needed to develop the big model.

As a result, some Meta researchers can only use fewer resources to continue to develop large models. They developed a large language model called the Open pre-training Model (OPT) using about 1000 chips in 2022 and Meta's flagship model LLaMA using about 2000 chips in 2023. By contrast, it takes 5000 to 10, 000 chips to develop such a large model by industry standards. Initially, Meta allowed a limited number of outside researchers to access LLaMA, but then it was leaked online, sparking a wave of innovation. Meta executives see this as a major example of Meta's desire to share its AI technology.

Since then, Meta has lost many AI researchers who worked on these and other key generative AI projects last year. Many people feel exhausted or lack confidence that Meta can keep up with its competitors. Six of the 14 authors who wrote research papers for LLaMA have left or announced they are leaving, according to LinkedIn and people familiar with the matter. Eight of the 19 co-authors who wrote the paper for OPT also left.

After OpenAI released ChatGPT in November last year, the turnover of Meta employees began to accelerate. Some people have been tempted to leave by the AI startup boom, which has driven across-the-board personnel changes at Silicon Valley companies, including Google. The number of GPT jobs mentioned on LinkedIn rose 79 per cent year-on-year in the year to March, according to LinkedIn.

A spokesman for Meta said the company has been continuing to recruit and bring in new AI talent.

After the ChatGPT launch, Zuckerberg, Meta chief product officer Chris Cox, and chief technology officer Andrew Bosworth were in charge of all the company's AI-related work. The three executives now spend hours a week studying AI, attending meetings and approving AI projects.

Previously, Meta had spent years prioritizing academic findings and sharing them for free, but was unable to capitalize on the commercial potential of these findings. Now, it is rapidly refocusing resources on developing available AI products and features, including its own chat bots.

Meta has set up a new generative AI division that focuses on building available products and tools rather than scientific research. The department received more than 2000 internal applications and quickly gathered hundreds of employees from different teams. Hardware resources have been transferred from AI research and are being used to train new generative AI models, according to people familiar with the matter.

Zuckerberg said in March that "promoting AI and integrating it into each of our products" was the company's biggest single investment. At Meta's annual shareholders' meeting in May, Zuckerberg said the company also wants to expand AI into the virtual world.

At a town hall meeting with employees earlier this month, Zuckerberg announced a range of generative AI products currently being developed by the company, including AI agents for Messenger and WhatsApp, AI stickers that users can generate and share in a chat based on text prompts, as well as a photo generation feature, a Meta spokesman said. The photo generation feature allows Instagram users to modify their photos using text prompts and then share them in Instagram Stories.

However, Meta still faces a wide range of challenges. After seven years of government and media scrutiny of its user privacy practices, the company is becoming less tolerant of risk, causing friction over how and when to launch AI products, according to people familiar with the matter.

Denying lagging behind, the White House held a summit of AI leaders in May, calling it a meeting for "companies at the forefront of AI innovation". But Zuckerberg was not invited.

Joelle Pineau, vice president of Meta AI research, said in a statement that the company was not lagging behind in AI development and defended its previous focus on AI academic research and structure, saying it laid the foundation for Meta's success.

Meta's AI research division "is one of the world's leading destinations for AI researchers and open science, and its research results have increased significantly in the past year alone," Pinault said. "our research breakthroughs have laid a solid foundation for this as we bring a new class of generative AI experiences to our application family," she says. "We are proud of the past and present contributions of Meta's AI researchers who have helped shape the future of advanced and state-of-the-art AI technologies." Phoenix New Media science and technology "AI outpost" will continue to pay attention to this.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.