So much of the digital world relies on open source software, where the underlying code is published on the internet so developers can freely modify it or build on it 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 fundamentally run the internet.
There’s been a big debate in Silicon Valley and among policymakers about whether companies should open-source their artificial intelligence. While Google and OpenAI have kept everything isolated, Meta has bet its entire AI strategy on the open-source route, making its powerful large-scale language models accessible to just about anyone.
Nils Tracy acknowledges that his business probably wouldn’t exist without Llama, the name of Meta’s open-source AI.
“No, I don’t think we can do that,” Tracy said. “It’s only with Meta’s support and Meta’s ability to build that model that we can make this happen.”
Tracy is founder and CEO of Blinder, a North Carolina-based startup that offers a suite of tools to help law firms securely interact with AI, which can do things like redact documents and automatically apply for copyright protection.
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 how to write internal law firm documents.
“We have law firm contracts and other types of documents that are written in law firm style,” Tracy says, “so we train them on that and help them learn how to write them.”
Tracy could have adopted large language models from OpenAI, Google, or other AI companies. But without those companies’ underlying code, he would have to trust that his data wasn’t being misused, and “fine-tuning” would be more difficult. Also, Llama is free.
“The open source model is incredibly valuable to people downstream who want to use it to develop new products and new applications,” said Elizabeth Seger, director of digital policy at the think tank Demos.
Creating AI models requires significant funding and infrastructure, so small companies like Blinder simply don’t have the resources to build their own large-scale language models from scratch.
You might wonder why Meta would spend tens of billions of dollars to develop an AI and then plan to give it away for free. Seger says one possible explanation is an economic strategy called “commoditizing complements.”
“Let’s say you’re a hot dog manufacturer,” Seger says, “you make hot dog buns completely free and everyone can get hot dog buns, but they have to buy your hot dogs.”
In this case, Meta’s free hot dog buns will be the company’s open source AI. AI development is still in its early stages, so Mark Zuckerberg can’t know exactly what his profitable hot dog will be, but the hot dog of the future could be a dataset that Meta sells for training, or the Meta hardware that the AI runs on.
Traditionally, open source software has another benefit: if there are security vulnerabilities, a global community of developers will find them.
Ali Farhadi is CEO of the Allen Institute for Artificial Intelligence, a nonprofit that has released a fully open-source AI model called OLMo.
“Would you rather live in a world where there are actually a lot of experts who know how to fix AI models when attacks and abuses occur, or one where we’re at the mercy of a few institutions?” Farhadi said.
But not everyone agrees that open source disclosure is the solution to AI security: Many worry that, unlike previous generations of software, exposing the underlying code to AI could make it more vulnerable to bad actors.
Programmers have already removed safety restrictions from some open-source AI image models in order to create deepfake porn.
Aviv Ovadia, co-founder of the AI & Democracy Foundation, said he worries about dire future outcomes, such as bad actors figuring out how to use open-source AI to create biological weapons.
“When AI systems are public and inspectable, it’s much easier to improve and learn,” Ovadia said, “but it’s also much easier to weaponize AI, and once it’s out there, it’s irreversible.”
Thinking of AI as either completely open source or completely black boxed is something of a false dichotomy, Ovadia added.
In fact, a wide range of open-source AI models are already available on the internet, with varying degrees of openness, with some models even exposing their training data along with the all-important “weights” – the mathematical secret sauce that drives the outputs that large language models produce.
There’s a lot going on in the world, and Marketplace is here for you.
Marketplace helps you analyze world events and bring you fact-based, easy-to-understand information about how they affect you. We rely on your financial support to keep doing this.
Your donation today will power the independent journalism you trust. Help sustain Marketplace for just $5 a month so we can continue reporting on the things that matter to you.
