Mark Zuckerberg keeps saying that his AI model is Open Source. He is abusing the term

Mark Zuckerberg can say mass. Llama models are not Open Source. Not at least in the strict sense of the definition, because As they explained As I wrote in the Open Source Initiative some time ago, “Meta confuses ‘Open Source’ with ‘resources available to some users under certain conditions’ – two very different things.”

These comments, made after the launch of Llama 2, are equally true for Meta’s new family of generative AI models. In both cases, as in the rest of the industry, we are witnessing these models taking everything they can from the public internet (and probably some of the private internet), as well as using, as in this case, the term Open Source a little too freely. Let’s see why.

Infinite voracity

Dubbed Llama 3.1, these models are promising in terms of performance and can even surpass GPT-40 or Claude 3.5, but in addition to highlighting this power and versatility, Mark Zuckerberg highlighted in an open letter how “Open Source AI is the way forward“.

The tweet with the image, posted by Elon Musk a few months ago, is no longer available on Twitter (X).

It is the same speech that, for example, he made after the presentation of the models in an interview. on Bloombergalthough there he did admit that Meta keeps secret the data sets they have used to train Llama 3.1. “Although it is open, we are also designing this ourselves.“, he noted, and only indicated that Facebook and Instagram posts had been used, as well as proprietary data sets licensed from others, without specifying further.

This lack of transparency is common in the industry: we don’t know exactly how other models like GPT-4 or Claude 3.5, which are completely closed and proprietary, have been trained. It is likely that in these and other cases surprising data has been collected: one such dataset, for example, contains 5,000 “tokens” from my personal blog.

The voracity of these models seems endless. This has led to controversy and lawsuits, but also to agreements for content companies to license their texts, images and videos to train them. Sometimes they don’t even ask for permission: OpenAI ran out of data to train its AI, so it transcribed a million hours of YouTube to train GPT-4, for example.

“Open Weights” is not the same as “Open Source”

It is true that the model is freely available on GitHuband that’s certainly remarkable: as was the case with Llama 2, these models can be used by companies and independent developers to create AI models derived from Llama 3.1.

Is something similar to what happens with GNU/Linux distributionswhich start from a Linux kernel and a series of components to which they then add their own additional elements.

The Llama 3 license certainly allows this way of working, but it also imposes a key barrier. on your license: Models derived from Llama 3.1 are free unless they are very successful. If the model ends up being used by more than 700 million active users per month, a license will have to be purchased from Meta.

But as in other cases, What Meta also shares are the so-called “weights”which provide information on how their calculations are performed. This allows anyone to download the files of the already trained neural network, and then use it directly or refine its operation for their own use cases. Doing something like this makes these models, rather than being Open Source, considered “Open Weights”.

As they explain on Ars TechnicaThis contrasts with what happens in proprietary models such as those of OpenAI, which do not share these weights and monetize the models through subscriptions to ChatGPT Plus or through an API.

That use of the term “open” by many AI projects, including Llama 3.1, is drawing increasing scrutiny (just ask OpenAI, which uses it as part of its company name).

This is what, for example, is highlighted by an interesting research by a team at Radboud University in Nijmegen, the Netherlands. The project leaders have analyzed various AI models, rating a series of parameters that allow them to assess whether the models are more or less open.

Source: Radboud University Nijmegen.

The result is a fantastic board where we can quickly check two things. First, that no model is perfect in this regard. And second, that Meta models are very low in this rating, and it is therefore very difficult to consider them as Open Source.

Simon Willison, co-creator of the Django programming environment and an expert in this field, commented that Mark Zuckerberg’s open letter was a “fascinating” and “influential” document, but also stressed that “it seems, however, that we have lost the battle in terms of getting them to stop using the term Open Source incorrectly.”

Indeed, Zuckerberg’s influence makes it difficult for the general public not to accept that Meta’s models are indeed open source when they are not entirely. As Willinson explained in comments on Ars Technica:

“I consider Zuck’s prominent misuse of “Open Source” It is an act of small-scale cultural vandalism. Open source should have a consensual meaning. Overusing the term weakens that meaning, which makes the term less useful overall, because if someone says ‘it’s open source,’ that doesn’t tell me anything useful anymore. And then I have to dig around and find out what they’re actually talking about.”

Yes, indeed. This widespread use – and not just by Zuckerberg – has weakened the concept, in part because there is no universal, accepted definition of what open source actually is in general, and what an open source AI model is in particular.

Image | Black011 with Midjourney