Why Meta’s Mark Zuckerberg is pushing the open approach to AI models

Mark Zuckerberg META-Q is doubling down on a more open approach to developing artificial-intelligence models, in contrast to other players in Silicon Valley that have chosen closed, proprietary paths.

The founder of Meta Platforms Inc. published a lengthy missive on Monday arguing that open-source technology is a “positive and important force for empowering people and preventing centralization.”

The company is also releasing a new AI model with open weights. That means anyone can download the parameters – numerical values that dictate the model’s behaviour – and tweak them to have more control over how the model functions. Called Muse Glimmer, the model is small enough to run on a consumer laptop.

Meta’s AI model hacks another company during cybersecurity testing

That differs from leading AI models from OpenAI and Anthropic that run on cloud computing infrastructure and are generally not possible to modify.

The letter cements Meta’s position in the debate around open source that has raged in Silicon Valley and Washington recently, one that involves the race between the United States and China for AI supremacy, questions about private versus public gain, and how best to secure a potentially dangerous technology.

Open source, in which a program’s source code is made freely available, has long been a standard practice in software. But building sophisticated AI models has been prohibitively expensive, which is why only a handful of well-capitalized companies are leading development. That’s starting to change. Chinese companies such as Moonshot AI and Alibaba are releasing capable open-weight models, allowing anyone to build on top of them, which has rattled American tech firms and investors.

Advocates of the open approach say that it allows users more control over the technology and can better protect privacy when the models run on local devices.

“Open models are seemingly the only kind of proposal where you can squint a little bit and see a way out of being stuck using OpenAI for the rest of our lives,” said Fenwick McKelvey, an associate professor at Concordia University who is researching open-source AI.

However, some experts have raised safety concerns about open weights because these models can be modified and misused – for instance, to orchestrate cyberattacks.

In response to fears that the U.S. government could restrict access to open-weight models, a group of tech companies including Meta, OpenAI, Nvidia Corp. and Microsoft Inc. signed a letter in late July saying that a robust open-source community is key to ensuring American leadership in AI. One reason, the companies argued, is that it will help boost adoption. “America wins the AI era by diffusing it into the workflows of factories, hospitals, farms, classrooms, and main street businesses,” the letter stated.

Opinion: Locked out of Anthropic’s Mythos, Canada must fight for AI access in USMCA talks

Toronto-based Cohere Inc., which has released weights for some of its models, also signed the letter. “This tech needs to be controlled by the people who use it, otherwise it’s just another piece of tech oligarchy control,” co-founder Nick Frosst wrote on X.

Meta has released model weights in the past, but paused amid a reorganization in its AI division last year. Not only is the company recommitting to the practice, but Mr. Zuckerberg recommended policy changes to help open source flourish. That includes advocating for distillation, a concept in which a large AI model is used to train a smaller one. OpenAI, for example, has accused Chinese companies of using its models to train their own through distillation.

“Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe,” Mr. Zuckerberg wrote.

One reason why Meta has embraced open-weight models is that they could weaken the business plans of OpenAI and Anthropic. By giving away access, Meta could entice some consumers to reduce paid usage of rival models. Meta has a massive business selling online advertising, too, a luxury that other AI companies do not have.

OpenAI, Anthropic agents implicated in new security breaches

“The fact that they can undercut their competitors and potentially raise doubts about OpenAI’s future … that could be worth it,” Prof. McKelvey said. Even with open-weight models, Meta could be trying to become a default option and retain significant influence over the ecosystem. “The strategy might be open, but there’s only one company that really controls it,” he said. “You become the hub of a new ecosystem.”

John Ruffolo, founder of Maverix Private Equity in Toronto, said that Meta’s open-weight approach could have more appeal outside of the U.S., allowing players in other countries to catch up in AI. “They want to focus on distribution and get you using it,” he said of Meta. “Now you’re using Meta’s products, and that’s their lock-in right there.”

Open-weight models still lag closed versions in terms of capabilities, but not by much. Epoch AI, a research-focused non-profit, estimated earlier this year that open-weight models are about four months behind.

In Canada, the federal government’s AI strategy released in June said that open-source AI is a “powerful alternative” to proprietary models and can lower barriers to adoption. The document was vague about next steps, saying that Canada will lead a “global, multi-stakeholder effort” to invest in open-source development and create an inventory of tools that domestic companies and organizations can access.

Anthropic chief executive officer Dario Amodei wrote in a blog post recently that open-weight models “potentially present a higher risk than closed models, because it is very difficult to apply guardrails to them or monitor their usage.” He clarified that he is not in favour of a ban, but said the U.S. should restrict chip exports to China, crack down on large-scale distillation and institute mandatory model safety testing.

Nicolas Papernot, an associate professor in computer engineering at the University of Toronto, said both open and closed models come with risks. As part of a recent study, for example, he and a team of researchers used an open-weight model downloaded from the internet to create an AI-driven computer worm that could infiltrate a variety of devices and replicate itself before moving on to another target.

But he is not opposed to open models, which he said are critical for understanding how AI works and for building cybersecurity defences. “Open-weight AI models in and of themselves are not the threat,” he said. “We need to be concerned about how to build up our cybersecurity to defend against all kinds of attacks, from rigged open-weight models and from far more powerful closed models.”