Meta released their flagship text-only model Llama 3.1, the best performing open-source artificial intelligence (AI) model so far, on Tuesday (23 July), which may very well be the company’s last one available to EU users.
Meta, the parent company of Facebook and Instagram, has faced pressure from EU regulators regarding its privacy policies around AI, as well as its “pay or OK” model for advertising, in recent months. In response, it has decided not to bring its latest AI products to the bloc.
Last week, Meta said it would not release its newest multimodal AI models in the EU, citing “the unpredictable nature of the European regulatory environment,” a spokesperson told Euractiv. Previously, the company rolled back plans to launch Meta AI, which would train AI with users’ publicly available posts.
Multimodal models can handle content formats like video, image and sound in addition to text.
Under current conditions, Meta will not be releasing any future multimodal models in the EU, and it is unclear whether Meta will apply the same strategy to text-based models, a Meta insider with knowledge of the matter told Euractiv.
Meta argues that cutting off the EU from its models, which the firm must do due to regulatory unpredictability, ultimately hurts European businesses and consumers.
It is creating a “gap in the technologies that are available in Europe versus” the rest of the world, Rob Sherman, Meta’s deputy privacy officer and vice-president of policy told the FT. Sherman also hinted that no more multimodal or text models might be coming to the EU.
Developers can download open-source models, often for free, and deploy them for their own applications, which also gives them the ability to change and audit the model. Closed-source models like OpenAI’s ChatGPT are available through subscriptions or similar interfaces.
Many companies employ these open-source models, to reduce costs and get more control over their own software and data.
These models are released under open-source licenses, which allow for the models to be downloaded and used, with certain restrictions.
To block EU users from using Meta’s multimodal model, Meta could set some geographical restrictions on the license or block access for European IP addresses.
Not open enough
Just how open Meta’s AI development is is a matter of debate in the industry. Though Meta shared more technical details on Llama 3.1 than closed-source developers, the company did not disclose specifics about what data the model is trained on.
“Holding back training data is not really in the spirit of open-source development,” Zuzanna Warso, director of research at the non-profit Open Future, told MIT Technology Review in March. Warso is taking part in the Open Source Institute, the de facto arbiter of open-source matters.
The idea behind open-source software is to make technology available to a plethora of developers outside the original entity that created something but to also enable developers to make the software better.
Meta’s motives in opening its models to developers are not entirely altruistic.
“I expect that our models are just going to improve further from open-source contributions,” Meta founder and CEO Mark Zuckerberg told investors in April.
As Meta’s AI models get better, the company may eventually have “people pay to use bigger AI models and access more compute,” Zuckerberg said.