Open source isn’t working for AI

Clearly, we need to do something about how we talk about open source and openness in general. It’s been clear since at least 2006 when I rightly got smacked down for calling out Google and Yahoo! for holding back on open source. As Tim O’Reilly wrote at the time, in a cloud era of open source, “one of the motivations to share—the necessity of giving a copy of the source in order to let someone run your program—is truly gone.” In fact, he went on, “Not only is it no longer required, in the case of the largest applications, it’s no longer possible.”

That impossibility of sharing has roiled the definition of open source during the past decade, and it’s now affecting the way we think about artificial intelligence (AI), as Mike Loukides recently noted. There’s never been a more important time to collaborate on AI, yet there’s also never been a time when doing so has been more difficult. As Loukides describes, “Because of their scale, large language models have a significant problem with reproducibility.”

Just as with cloud back in 2006, the companies doing the most interesting work in AI may struggle to “open source” in the ways we traditionally have expected. Even so, this doesn’t mean they can’t still be open in meaningful ways.

According to Loukides, though many companies may claim to be involved in AI, there are really just three companies pushing the industry forward: Facebook, OpenAI, and Google. What do they have in common? The ability to run massive models at scale. In other words, they’re doing AI in a way that you and I can’t. They’re not trying to be secretive; they simply have infrastructure and knowledge of how to run that infrastructure that you and I don’t.

“You can download the source code for Facebook’s OPT-175B,” Loukides acknowledges, “but you won’t be able to train it yourself on any hardware you have access to. It’s too large even for universities and other research institutions. You still have to take Facebook’s word that it does what it says it does.” This, despite Facebook’s big announcement that it was “sharing Open Pretrained Transformer (OPT-175B) … to allow for more community engagement in understanding this foundational new technology.”

That sounds great but, as Loukides insists, OPT-175B “probably can’t even be reproduced by Google and OpenAI, even though they have sufficient computing resources.” Why? “OPT-175B is too closely tied to Facebook’s infrastructure (including custom hardware) to be reproduced on Google’s infrastructure.” Again, Facebook isn’t trying to hide what it’s doing with OPT-175B. It’s just really hard to build such infrastructure, and even those with the money and know-how to do it will end up building something different.

This is exactly the point that Yahoo!’s Jeremy Zawodny and Google’s Chris DiBona made back in 2006 at OSCON. Sure, they could open source all their code, but what would anyone be able to do with it, given that it was built to run at a scale and in a way that literally couldn’t be reproduced anywhere else?

Back to AI. It’s hard to trust AI if we don’t understand the science inside the machine. We need to find ways to open up that infrastructure.