Meta unleashes its Llama

The company's open-source AI is now roughly on par with rivals from OpenAI and Google. Should we worry?

Casey Newton

Jul 23, 2024 — 8 min read

Screenshots for "Imagine me," a new AI selfie feature powered by Llama 3.1, which Meta released today. (Meta)

Today, Meta released the largest ever open-source large language model to power generative artificial intelligence applications. Llama 3.1 represents an ambitious effort to shape the development of the AI industry in ways that favor Meta. It is also likely to spark new conversations around whether open source models are likely to generate more harm than their closed counterparts.

In January I wrote about Meta’s unusual plan to make available for free a technology that it has spent more than a decade and tens of billions of dollars to build. Unlike its peers, including Google and OpenAI, Meta is not selling subscriptions to its AI for individuals or teams, and says it has no plans to do so.

In the short term, Meta says that open-sourcing Llama helps its own systems get better more quickly and cheaply than they otherwise would. Inviting a global ecosystem of developers to iterate and improve on its models, and funnel those improvements back into Meta’s core systems, accelerates the development of Meta’s own products and systems and can save it money in doing so.

In the long-term, owning one of the most powerful LLMs could provide the company with a basis for all manner of money-making products. It could also — and Meta lets this part go unsaid — thwart the development of competitors like OpenAI by giving away their core business for free. (One way to think of Llama 3.1 is that it is the free Google Docs to Microsoft’s paid Office 365.)

To make that happen, of course, Meta’s AI needs to be among the best in its class. The company said today that it is making strides in that direction: Llama 3.1’s largest, 405-billion parameter model outperformed OpenAI’s GPT-4 Omni and Claude 3.5 Sonnet on some benchmarks, it said.

“If you compare them to the GPT family, if you compare them to the Claude family, if you compare them to Gemini — I'll let the numbers speak for themselves,” Chris Cox, Meta’s chief product officer, said in an interview with Platformer. “But we're pretty proud now that some of them we’re beating, and then on some of them we’re just in range of best-in-class models. So that's really cool.”

And while developers have only had it in their hands for a few hours, the early reception has been positive. On Hacker News, some developers found that Llama 3.1 performed at or near GPT-4o’s level, according to comments posted there.

People with extremely high-end hardware may be able to host the largest Llama 3.1 model on their own hardware. But most developers will want to use a hosted service, and to that end today Meta announced partnerships with 25 companies that will make Llama 3.1 available to use through their cloud computing platforms, including Amazon Web Services, Google Cloud, Microsoft Azure, Databricks, and Dell.

The fact that an open-source model now rivals closed alternatives speaks to the way that every major AI developer’s LLMs are converging on one another in quality. Seemingly every few weeks now, one of the big AI players releases a new model, or variation of a model, with slightly improved cost, performance, or other attributes. And for the most part, at least in my testing of the major players, the resulting products are mostly indistinguishable from one another.

That’s bad news if you’re selling a subscription for $20 a month, as OpenAI, Google, and Anthropic all are. But it’s great news for Meta, which can expedite the development of its own systems and undercut its rivals at the same time, all without damaging its core advertising business.

If there are risks to Meta here, they lie in the potential for regulation and damage to its reputation. That’s because open-source systems, by their nature, allow anyone to take them and repurpose them for their own ends, no matter what they might be. Fears that the open development of superhuman systems could create existential risk for humanity are among the reasons that OpenAI abandoned that approach in favor of a closed one.

Over the past two years, open-source vs. closed development has become one of the most hotly debated issues in tech. The Biden administration has not definitively favored one approach over the other. But in its executive order last year, the administration did call for projects that use past a certain computing threshold to disclose that fact to the government and to perform safety testing. Some open-source advocates argue that the administration is laying the groundwork for further restrictions that will effectively outlaw open-source, and in so doing create a tiny cartel of vastly powerful AI companies that will reap most of the benefits of the technology.

Among the voices making this argument is Marc Andreessen, the venture capitalist and longtime Meta board member. It was among the reasons that he and his business partner Ben Horowitz offered a full-throated endorsement of Donald Trump for president last week.

Zuckerberg has reached a similar conclusion. In a kind of manifesto published today, Meta’s CEO argued that open source development is the best approach not only for Meta’s business, but for the world at large.

He writes:

It’s worth noting that unintentional harm covers the majority of concerns people have around AI – ranging from what influence AI systems will have on the billions of people who will use them to most of the truly catastrophic science fiction scenarios for humanity. On this front, open source should be significantly safer since the systems are more transparent and can be widely scrutinized. Historically, open source software has been more secure for this reason.

Zuckerberg also suggests that there are few harms that can be surfaced through AI today that you can’t already Google. “We must keep in mind that these models are trained by information that’s already on the internet, so the starting point when considering harm should be whether a model can facilitate more harm than information that can quickly be retrieved from Google or other search results,” he writes.

These arguments seem sound enough today, when the models are barely capable of doing more than regurgitating their training data. It’s less clear how they will sound in a future, when open-source software can reason, scheme, and execute complex plans — no matter what those plans might be.

And Meta is among the companies building toward that future. Cox told me that the next radical advancements in LLMs will come when they can master three skills: solving problems that require multiple steps; personalizing themselves to their users and developing a more complete sense of who their users are and what they need; and successfully taking action on your behalf online.

Should LLMs develop those skills, and become part of the Llama of the future, Meta will surely attempt to put safeguards around it to prevent it from being misused. But the nature of open source technology is that the moment Llama 3.1 goes online, it is out of Meta’s hands. And where its rivals will retain broad power to shut down rogue systems, it’s less clear that Meta will be able to build a similar killswitch.

Open source development has served Meta’s interests well so far, and the relatively benign AI tools we have today give us little reason to fear what Llama 3.1 can do. But it’s hard to make good policy for a technology that is improving in exponential leaps every few years. As AI grows more powerful, Meta’s case for open-source development will be worth revisiting.

Correction: This post originally said Cox is Meta's chief privacy officer. He is its chief product officer.