I still remember that in October 2021, Facebook announced that it would change its name to Meta, saying that it would work hard to build the Metaverse, but the results disappointed investors and consumers. Although the virtual reality platform "Horizon Worlds" launched by Meta is a step towards realizing the metaverse, the graphics are poor, the world is boring, and there are only a few users. Meta Chairman and CEO Mark Zuckerberg's virtual double in the Horizon Worlds was even ridiculed by everyone, and people couldn't help but think that Zuckerberg had chosen the wrong battlefield.
Unexpectedly, in 2023, Meta released an open source large-scale language AI model called Llama to catch up again. According to Fortune, the model has been downloaded more than 600 million times on websites such as the open source AI community Hugging Face, and more than 90% of open source models are based on Llama.
Meta's Llama model has been changed from leaked to free for use
Meta has invested in the AI field in 2013, led by Yann LeCun, the "father of convolutional neural network", in the FAIR (Fundamental AI Research) laboratory. An academic robot named Galactica was launched in November 2022, but the robot often gave wrong answers, shocking many scholars. At the time, it was even criticized as "the most dangerous thing Meta has ever created."
Twelve days after Meta shut down Galactica, OpenAI released ChatGPT, which caused a whirlwind around the world; FAIR concentrated on developing the Llama model. Worried about being criticized again, Meta initially only provided access to a few people. Unexpectedly, the model was leaked a few weeks later. Meta originally planned to require relevant websites to remove leaked models, but then Zuckerberg announced on July 18, 2023 that Llama 2 would be free for research and commercial use, and this became the key to Meta's comeback.
"Open source" becomes the key for Meta to quickly catch up with its opponents
Many scholars called on Meta to make Llama an open source model at the time. Yann LeCun said that many people believe that open source is extremely valuable, allowing outsiders to gain access to test and inspect the operation of the model, which will help various industries create new ecosystems. FAIR researchers believe that the open source Llama 2 can make the model more powerful faster at a lower cost. After Zuckerberg learned about it, he assessed that open source would not affect the company's operations, and he did not want to have a situation where only a few companies control AI. He ultimately decided to give researchers, developers, and other users access to the underlying program and the ability to determine how the model processes information in order to improve the model. The speed and accuracy of the Llama 3 model, released in 2024, is on par with, and even better than, the strongest models from OpenAI and Anthropic.
Khajuria described Zuckerberg's use of Llama as an open source model as a "genius move." Through open source, it can help attract top talents, allow the platform to accelerate innovation and develop new revenue sources, and extend the life of the platform. She said that with thousands of developers helping to correct errors and improve the Llama model, Meta can quickly catch up with OpenAI, Google and Anthropic. "Without open source, it (Llama) might have taken longer to be on par with other advanced models," Khajuria said.
Open source Llama is powerful but its business model is unclear
Schweta Khajuria, an analyst at Wolfe Research speculates that the Meta AI function developed by Llama can be used to help brands place advertisements. In addition, there is speculation that the Llama model may implement a payment system in the future to assist Meta's revenue. However, the reality is that Meta has spent billions of dollars building the Llama model and estimates that capital expenditures in 2024 will be as high as 40 billion US dollars, and expenditures may increase significantly in 2025. It is still provided to Goldman Sachs, AT&T and other thousands of companies use it for free, and the profit model still needs to be developed. Abhishek Nagaraj, associate professor at Haas School of Business expressed his confusion about the commercial interests of Llama, because from a purely economic point of view, Meta invested costs to build the model but provided it to enterprises for free, making it difficult to make a profit. Khajuria also warned that if Meta's revenue does not accelerate in 2026, "investors will lose patience."
In addition, some people are worried that open source will cause a crisis. China has used Llama to develop military AI tools. And Donald Trump is re-elected as President of the United States, he certainly does not want other countries to obtain advanced AI models made in the United States. The outside world is worried that it may intensify geopolitical tensions and arms races. Facebook co-founder Dustin Moskovitz also said that he is not opposed to open source Llama, but Meta should not release more powerful versions.
Source: Fortune
