Regarding Meta’s Release of LLaMA 2

July 18, 2023

Earlier this year, Meta (Facebook) released an AI model called LLaMA. LLaMA is a GPT-like language model with similar capabilities. It was released under a closed license that was only available to those approved by Meta, though out the time they advertised it as “Open Source”. The model quickly leaked, and became a popular community AI tool with many variations built on top of it and even a custom (now arguably industry standard) open source software framework built to run it called llama.cpp.

There had been indications that an open version that amongst other things would be licensed for commercial use, would be released. Unfortunately Meta chose to release the model instead under a restricted license. I wrote the following earlier on Linkedin:

A grim development for software freedom today as Meta eschews open source licensing and releases their LLaMA model under a license that only lets you use it for what they want. What would the world be like if Linux was only available for uses approved by big tech companies? This is a somewhat esoteric issue but it's going to be the difference between an open technology ecosystem and one where big tech companies and special interests dictate how we can use our computers. I wrote the linked article1 a while ago but it's even more relevant now.

What you can do is be vocal about calling out companies like Meta who falsely use "open source" as a brand (before LLaMA they were a great contributor to the community don't get me wrong), and prioritize using and building truly open source AI models that respect user freedom.

And on HackerNews:

It's disappointing that you're stuck using LLaMA at Meta's pleasure for their approved application. I was hoping they would show some leadership and release this under the same terms (Apache 2.0) as PyTorch and their other models, but they've chosen to go this route now which sets a horrible precedent. A future where you can only do what FAANG wants you to is pretty grim even if most of the restrictions sound benign for now. The real danger is that this will be "good enough" to stop people maintaining open alternatives like open-LLaMA. We need a GPL'd foundation model that's too good to ignore that other models can be based off of.

Open source licensing can be a strategic tool to perpetuate open and access to technology. Meta could have been a leader here, but unfortunately chose to go with terms that benefitted them commercially and politically. Get in touch if you want to know more or take action.

See also:

Motivations and strategy for open sourcing AI model weights

Andrew Marble, andrew@willows.ai


  1. http://marble.onl/posts/software-licenses-masquerading-as-open-source.html↩︎