In February of last year, Meta—which owns Facebook, Instagram, and WhatsApp—threw a metaphorical grenade into the market for artificial-intelligence software by announcing a new model that it called Llama. Meta’s offering is similar to those offered by OpenAI and Google in that it is an AI engine known as a “large language model.” (I’ve written on a number of occasions about this kind of software and the risks it poses.) But Llama is unique in one critical way: Meta isn’t charging companies to use it or to integrate it into their products, as its competitors do with their tools. Instead, it is giving the software that powers Llama away free of charge, with a so-called “open source” license that allows anyone to use and modify the software for their own purposes (with some restrictions).
Software analysts say that this approach is a risky strategy for Meta—it costs billions of dollars to build an AI engine—but the potential benefits of getting the upper hand in the emerging industry seem to have convinced Mark Zuckerberg, the company’s founder and CEO, that it is a risk worth taking. And Meta has been doubling down on this strategy at a rapid pace: last week, it released the third version of Llama only a few weeks after announcing the second. In a blog post, Zuckerberg said Llama 3.0 (its technical name is Llama 3.1 405B) is the most powerful version of the software yet, and reiterated why he thinks that making the software open-source is good not just for Meta but for the AI industry—and the world—in general.
Zuckerberg also compared his open-source AI engine to the development of Linux, an open-source operating system for computers that has become the standard for cloud servers and mobile devices that run on Android. In the early days of computing, Zuckerberg wrote, major tech companies invested heavily in developing their own closed-source versions of Unix, an operating system for large computers; at the time, he notes, it was hard to imagine “that any other approach could develop such advanced software.” But, even though it was not backed by any multibillion-dollar corporations, Linux grew in popularity, in part because it was free but also because it allowed developers to modify the underlying code to suit their own purposes, and because it was supported by volunteers who fixed problems without asking to be paid. Zuckerberg wrote that he expects AI to “develop in a similar way,” and noted that Meta has historically open-sourced a lot of the software behind its services, through efforts such as the Open Computer Project.
Industry analysts have pointed out that Meta’s interest in an open-source AI model is driven by more than altruism. Unlike competitors such as Google, Microsoft, and Amazon, the company doesn’t own a cloud-computing service that it can use to both provide the horsepower for its AI offerings and make them available at a low price. It’s thus in Meta’s interests to promote an open approach, one that allows it to form partnerships with cloud providers such as Google Cloud and Amazon’s AWS to host and run Llama. (Anyone can run Llama—but it takes a lot of computing power.) In the Platformer newsletter, Casey Newton compared this to Google offering a free, cloud-based document management system to compete with Microsoft Office.
Open-source advocates have also pointed out that while Meta’s Llama engine can be downloaded and used for free, it doesn’t technically qualify as “open source” because the underlying data that powers the engine—the database of aggregated information that Meta used to train the AI—isn’t itself open. Users who get access to Llama 3.0 can use it and can modify the way it operates and the terms on which it generates results, and they can see what are called the “weights” or values assigned to certain types of data. But they can’t see the data underlying those operations, nor can they modify it. In addition, there are restrictions on who can use the model: if a company has more than seven hundred million users, it has to request a license from Meta directly, and Meta can refuse to let the company use the software if it wants to.
So how does Meta make money if it’s giving away its software? This is a question that some technology investors have been asking themselves ever since the company said that it spent close to ten billion dollars on AI research and development last year, and that it plans to continue spending. That announcement, which came along with Meta’s annual financial results in April, helped fuel one of the largest single-day stock-price drops in the company’s history, a decline that shaved more than two hundred billion dollars off its market value. (The share price has recovered somewhat since then, and Meta had a strong earnings report yesterday.)
Open-source software advocates say there are a number of ways in which companies like Meta can generate revenue even though they are giving their software away. Red Hat—which sells a for-profit version of Linux—and WordPress, the website publishing company, both provide added features and/or services and support for a fee. In Meta’s case, the company could do much the same thing, by offering a vanilla version of Llama for free and then adding features or related services for those who are willing to pay extra for ease of use and some support.
Of course, there is already a large AI company with the word “open” in its name—OpenAI, the creator of ChatGPT—but it is not open at all, not even to the extent that Llama is. OpenAI was originally set up to create an open-source version of AI software, but its founders decided that it had to become a for-profit venture, in part due to the costs; in the end, it became a for-profit inside a nonprofit, a confusing structure that exacerbated tensions last year, when the board voted to oust Sam Altman, the CEO, in part because he wanted to be more aggressive in developing ChatGPT. (Altman was reinstated; I wrote about all this at the time.) One of the other factors that helped push OpenAI away from building a truly open engine was the fear that it could be used for nefarious purposes. OpenAI’s original mission statement stipulates that the goal is to develop safe AI for the benefit of humanity, not for profit, and states that if OpenAI’s engine achieves what is called “artificial general intelligence” (loosely, humanlike intelligence), it will stop licensing that model to others. Some observers similarly fear that Llama could be used for evil.
Zuckerberg wrote in his blog post that a world where AI is widely deployed is a better one, since “larger actors can check the power of smaller bad actors.” But not everyone agrees that the benefits of open source outweigh the disadvantages. Geoffrey Hinton, one of the founders of modern AI, quit working for Google last year because he wanted to speak out about the potential dangers of the technology. He told Wired that if AI software is freely available, “cyber criminals everywhere will be delighted,” adding that while open source works well for developing operating systems and other types of software, AI engines are fundamentally different because they can’t be scrutinized in the same way. “People fine-tune models for their own purposes,” Hinton said, “and some of those purposes are very bad.”
Llama isn’t the only open-source model in the AI ecosystem, although it is undoubtedly one of the largest. Others who have taken the open-source approach to artificial intelligence include a French platform with the unusual name of Hugging Face, which is dedicated to hosting and sharing open-source models and tools. (A staffer at the company once told me that the odd name was the result of trying to translate a French term into English.) A different company, Stability AI, is known for creating open-source image-generation models such as Stable Diffusion (which has come under fire in the past for generating obscene images, including of child sexual abuse) and EleutherAI, a community-driven research lab that is focused on the development of open-source AI models.
One of the more interesting parts of Zuckerberg’s argument for Llama being open source amounts to a broadside at Apple: he writes that developers, including companies like Meta, are frustrated by “the way they tax developers, the arbitrary rules they apply, and all the product innovations they block from shipping.” It would be better, Zuckerberg argues, if competitors like Apple were not “able to constrain what we could build.” It’s hard to think of Meta, a globe-spanning colossus worth more than a trillion dollars, as an underdog nipping at its corporate masters. Whether a large enough number of true underdogs will buy into this defense of open-source AI is an open question. As is whether it’s a good idea.
Other notable stories:
- Yesterday, Ismail Al-Ghoul and Rami Al-Refee, respectively a correspondent and camera operator for Al Jazeera in Gaza, were covering the assassination of Ismail Haniyeh, a senior leader of Hamas, near his home in the territory when Israel ordered the area evacuated; according to Al Jazeera, the pair were driving away when they were killed in an Israeli strike that the broadcaster alleges was targeted. (At time of writing, Israel had yet to comment; it has generally denied targeting journalists.) As the Committee to Protect Journalists notes, Al-Ghoul had previously reported being assaulted in Israeli custody. By CPJ’s count, he and Al-Refee are at least the sixth and seventh journalists affiliated with Al Jazeera to have been killed since October 7, with the overall toll of the conflict on journalists now standing at a hundred and thirteen deaths.
- Also yesterday, Donald Trump sat for an interview at the National Association of Black Journalists’ annual convention. As CJR’s Feven Merid reported ahead of time, the invite sparked controversy among the group’s members, and Trump’s appearance did nothing to quell it; he attacked one of the interviewers—Rachel Scott, of ABC—and questioned whether Kamala Harris is actually Black. Karen Attiah, who stepped down as co-chair of the convention before the interview, reported during it that the audience in the room was “boiling with anger and disappointment”; later, MSNBC’s Joy Reid said that the NABJ had been “played.” CNN’s Abby Phillip countered that the Trump invite was “grounded in journalism” and had produced “one of the newsiest interviews with him this year.”
- As we noted yesterday, Will Lewis, the beleaguered publisher of the Washington Post, was back in the headlines this week after plaintiffs in a phone-hacking case against Rupert Murdoch’s UK media company alleged that he “fabricated a fake security threat” to delete evidence of wrongdoing when he worked for Murdoch in the early 2010s. That alleged security threat involved convoluted allegations that Gordon Brown, a former British prime minister, was trying to steal a Murdoch executive’s emails. Brown has strongly denied this; now he has revealed that police in London are looking into Lewis’ conduct, though he is not under criminal investigation. The Guardian has more details.
- As we also noted earlier this week, police in London charged Huw Edwards, a former top news anchor at the BBC, with three counts of making indecent images of children. (Edwards was first suspended by the broadcaster last summer after a complicated—and ultimately unrelated—set of allegations against him were reported by The Sun; we covered them here.) Edwards has since pleaded guilty to the charges, admitting in court that he received WhatsApp messages containing the images from a convicted child sex offender. Edwards, who no longer works for the BBC, has yet to be sentenced.
- And police in New York arrested a man who damaged the license-plate covers of Secret Service agents assigned to protect Ella Emhoff, Harris’s stepdaughter—though the man appears to have been motivated not by Emhoff’s identity but by a broader campaign to stop motorists evading tolls and traffic laws by obstructing their plates. Gersh Kuntzman, a New York journalist who has popularized that campaign, told the AP that he felt some responsibility for the incident, adding that, while he supports “cleaning up” state property that is supposed to be readable, he has “never messed with the Secret Service.”
ICYMI: The National Association of Black Journalists booked Donald Trump. Many had strong feelings.
Mathew Ingram is CJR’s chief digital writer. Previously, he was a senior writer with Fortune magazine. He has written about the intersection between media and technology since the earliest days of the commercial internet. His writing has been published in the Washington Post and the Financial Times as well as by Reuters and Bloomberg.