Here’s why the AI ​​open source debate is important for us humans

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

There’s a huge debate in Silicon Valley and among policymakers about whether companies should open source AI. While Google and OpenAI keep everything siloed, Meta has put its entire AI strategy on the open-source path, allowing virtually anyone to access its powerful language models.

Nils Tracy admits that his company likely wouldn’t exist without Llama, the name of Meta’s open-source artificial intelligence software.

“No, I don’t think that’s possible,” Tracy said. “The support Meta has and the strength Meta has to build this model allows us to do it.”

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

Tracy copied the entire Llama source code, made a few tweaks, and then “tuned” the AI ​​to identify personal information in legal documents that needed redaction. The AI ​​can also learn a law firm’s internal writing style.

“We have contracts from law firms or other types of documents they might have that are written in their style,” Tracy said. “We’ll train him on that and get him to learn that style of writing.”

Tracy could try to adopt large language models from OpenAI, Google, or other AI companies. But without their core code, he would have to trust that they weren’t abusing his data, and “tuning” would be harder. Plus, Llama is free.

“This open source model is incredibly valuable for anyone who wants to leverage it to develop a new product or a new app,” said Elizabeth Seger, director of digital policy at think tank Demos.

Because AI models require a huge amount of capital and infrastructure to build, small companies like Blinder simply do not have the resources to build their own original, large language models from scratch.

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

“Let’s say you’re a hot dog manufacturer,” Seger said. “So you make hot dog buns completely free, so anyone can get a hot dog bun. But then they have to buy your hot dogs.”

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

Traditionally, open source software has another advantage: if a security flaw occurs, the global developer community is able to detect it.

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

“Would I rather live in a world where there are actually a lot of practitioners who know how to fix AI models when they are attacked or misused, or a world where I am at the mercy of a few institutes?” Farhadi said.

But not everyone agrees that open-source Sunshine is a cure-all for AI security. Many fear that, unlike previous generations of software, opening up the code base for AI will make it more vulnerable to attacks from bad actors.

Developers have already removed security restrictions from some open-source AI image models to create deepfake pornographic content.

Aviv Ovadya, co-founder of the AI ​​& Democracy Foundation, said he worries about catastrophic consequences in the future, such as bad actors figuring out how to use open-source AI to build biological weapons.

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

Ovadya added that viewing AI as something completely open or completely locked in black boxes is a kind of false division.

The various open-source AI models already published on the internet are in fact open to varying degrees. Some models publish their training data along with their all-important “weights”—the mathematical secret sauce that drives the output of a large language model.