Much of the digital world relies on open source software, in which the underlying code is published to the internet and made free for developers to tinker with and build upon as they see fit. Mozilla’s Firefox browser, the publishing platform WordPress and the operating system Linux are just three examples, along with many other lesser known programs that essentially keep the internet running.

There’s a big debate within Silicon Valley and among policymakers over whether companies should make artificial intelligence open source. While Google and OpenAI are keeping everything walled off, Meta has staked its entire AI strategy upon the open source route, allowing pretty much anyone access to its powerful large language models.

Nils Tracy admits his business probably wouldn’t exist without Llama, the name for Meta’s open source AI.

“No I don’t think it could,” said Tracy. “The backing of Meta and the power that Meta has to build that model enables us to do this.”

Tracy is the founder and CEO of Blinder, a North Carolina-based startup that provides a suite of tools for law firms to safely interact with AI. Blinder can do things like redact documents or automatically file for copyright protections.

Tracy’s company copied all of Llama’s source code, made some tweaks and then “fine-tuned” the AI to identify personal information in legal documents that needed to be redacted. The AI can also learn in the in-house writing style of a law firm.

“We have contracts from a law firm or other types of documents they might have that are written in their style,” said Tracy. “We will train it off of that and we will get it to learn that style of writing.”

Tracy could have tried to adapt large language models from OpenAI, Google or other AI companies. But without their underlying code, he’d have to trust they weren’t misusing his data and the “fine-tuning” would be tougher. Also, Llama is free.

“That open source model is extremely valuable for anyone downstream who wants to use it to be able to develop a new product or a new application,” said Elizabeth Seger, director of digital policy at the think tank Demos.

Because AI models take such massive amounts of money and infrastructure to create, small companies’ like Blinder simply don’t have the resources to create their own original large language models from scratch.

If you’re wondering why Meta plans on spending tens of billions developing its AI, only to give it away for nothing? Seger says one possible explanation is an economic strategy called “commoditizing the complement.”

“Let’s say you are a hot dog producer,” said Seger. “So you make hot dog buns completely free, so everybody can get their hands on hot dog buns. But then they need to buy your hot dogs.”

In this case, Meta’s free hot dog buns would be its open source AI. And while it may be too early in AI’s development for Mark Zuckerberg to know exactly what his money-making hot dog would be, those future hot dogs could be datasets Meta will sell you to train it or Meta hardware the AI runs on.

Traditionally with open source software, there’s another benefit too: If there’s a security vulnerability, the global developer community is there to spot it.

Ali Farhadi is CEO of the nonprofit Allen Institute for Artificial Intelligence, which has released a completely open source AI model called OLMo.

“Would I rather live in a world where there is actually a large number of practitioners who know how to fix AI models when an attack or misuse happens, or a world in which I’m at the mercy of a couple of institutes?” said Farhadi.

But not everyone agrees that open source sunshine is the cure for AI security. Many worry that unlike previous generations of software, making the underlying code available for AI will make it more vulnerable to bad actors.

Programmers have already removed safety restrictions from some open source AI image models to create deep fake pornography.

Aviv Ovadya, co-founder of the AI & Democracy Foundation, said he worries about future catastrophic outcomes, like bad actors figuring out how to use open source AI to build a biological weapon.

“It’s much easier to improve and learn about AI systems if they’re open and available to inspect,” said Ovadya. “But it’s also much easier to weaponize them, and there’s no undo once they’re out in the world.”

Thinking about AI as completely open source or completely black boxed is kind of a false binary, Ovadya added.

Different open source AI models already published to the internet are, in truth, varying degrees of open. Some models publish their training data along with crucially important “weights” — the mathematical secret sauce that guides the output a large language model produces.