Meta spends a lot on artificial intelligence – only to offer the latest cutting-edge AI models to the public for free.
The tech giant recently announced the rollout of Llama 3, the open-code large language model underlying the Meta AI assistant tool that is now widely accessible on its flagship social media platforms. Meta aimed to “build the best open models comparable to the best proprietary models currently available,” it wrote in a blog post.
The model quickly received support from technologists – but not from investors.
Meta stock tumbled Wednesday after CEO Mark Zuckerberg touted a slew of money-losing projects during his first-quarter earnings call. Zuckerberg expressed his confidence in investing even further in advanced open-source AI models.
“I’m very happy with the way Llama 3 has come together so far,” he said on the earnings call, adding, “I expect our models will only continue to improve with open source contributions.”
“Open source” is generally used to describe software that is made available to the public, who can use and improve or build upon the code. Several fast-growing AI startups, such as Mistral AI and Hugging Face, have found success offering open-source models and tools.
In AI, models – such as large language models and fundamental models – are the complex pieces of software that use algorithms, trained on relevant data, to perform tasks such as recognizing patterns and making predictions or decisions. OpenAI, Google, Anthropic, and many more companies are constantly developing, training, and adapting their models to create increasingly capable AI programs.
So what’s the catch? Why would Meta spend so much time and money developing AI models only to make them open source?
Alex Ratner, CEO of data-focused company Snorkel AI and assistant professor of computer science and engineering at the University of Washington, wrote in an email that Llama 3 is a “much bigger step change than many people realize.” He said he suspects AI developers will benefit massively, while Meta can also reduce costs, attract talent and increase potential revenue.
When a company or organization wants to integrate AI technologies into its operations, it must use its own internal data to properly fine-tune those AI models. But using closed-source models means giving away that data – often the most valuable asset – to an AI model vendor, who ultimately owns the final model.
However, using a model like Llama 3 gives these companies a way to maintain ownership of their own data and the output of their work. That makes the use of open-source models much more attractive to developers and data scientists, he wrote, potentially giving Meta a better chance to attract top talent in the AI space.
“Enterprise AI is becoming less about the largest and most powerful generalist model, and more about the specialist trained on your data,” Ratner wrote, “and Meta has positioned Llama 3 to potentially be the epicenter of that.”
Others in the tech community have offered similar assessments. Bindu Reddy, CEO of generative AI startup Abacus.AI, recently said on X that Meta’s move with Llama “was a strategic masterstroke and serves her business interests.”
Aaron Levie, CEO of cloud computing company Box, said on X: “It’s unbelievable that Meta is advancing open source AI in such a big way here. A great time to build software when you have hundreds of billions of dollars from the big technology players engaged in advanced R&D and infrastructure, which directly benefits you as a developer.”
Unlike many of its fellow tech giants, Meta has a history of investing in open source projects, from machine learning library PyTorch to JavaScript library React to data query language GraphQL, among others. The earlier versions of Llama were also advertised as open source.
But there are limits to Llama’s openness.
While Meta allows most developers to use Llama 3 commercially, those who want to use it in products with more than 700 million monthly active users – essentially just major tech platforms like Google, TikTok and Snapchat – must apply for licenses from Meta. Also, the data used to train Lama 2 and 3 has not been made public, making it difficult to assess possible bias.
And the company’s license for Llama 3 has not been approved by the Open Source Initiative, a widely recognized nonprofit organization that sets rules for defining open source software, making labeling Meta’s LLM “open source” somewhat controversial in the technical world.
Still, Zuckerberg detailed in another earnings call this year that Meta has strategic interests in positioning itself as a leader in open-source software infrastructure. Not only is open source software a more popular option among developers and researchers, he said, but it is also more likely that such software will become an industry standard.
The advantage of closed-source AI models – such as OpenAI’s GPT, Anthropic’s Claude and Google’s Gemini – is that the AI providers behind the LLMs retain more control over their own algorithms and technologies, making it harder for competitors to replicate them. But Zuckerberg has said he’s not worried.
“The short version is that open sourcing improves our models,” Zuckerberg said on the call. “And because there is still a lot of work to turn our models into products, and because there will be other open source models available anyway, we find that there are particular advantages to being the open source leader, and that this will drive differentiation of our products does not take away. a lot.”