Facebook’s language model, dubbed LLaMa, has leaked to the controversial 4chan platform, embarrassing Mark Zuckerberg’s group.
In the battle of artificial intelligence, the Meta group, owned by Facebook, is experiencing a serious setback. While its “in-house ChatGPT”, LLaMa (for Large Language Model Meta AI) was intended only for approved researchers, government officials or certain members of civil society, the language model was posted on the forum of British 4chan pictures. This community is particularly frequented by certain hackers and tormented individuals, and unsavory content circulates there historically.
A proven LLaMa leak…
Our American colleagues from Vice explain that last week, an Internet user posted a torrent file on 4chan that allows you to download Facebook’s LLaMa tool. It is not a chatbot, but a new language model that can be used as an infrastructure to build a ChatGPT-like conversational robot.
Several 4chan users claim to have downloaded the California giant’s AI model and started using it on their machines. But the contours of this rather disturbing leak for Meta are blurred. A spokesperson for the company also spoke on the subject.
The model ” LLaMa has been shared for research purposes, consistent with how we have shared previous large language models. Although the model is not available to everyone and some have tried to circumvent the approval process […]the current strategy allows us to balance responsibility and openness “, explains the company.
… which would not bother, on the surface at least, the Meta group
In short, Meta recognizes that there has been a leak, but is not overly alarmed. Nevertheless, the company would have launched various requests for the withdrawal of the leaked model, so as to control its distribution which was not envisaged at this stage. It may already be too late…
Still, the language model for AI LLaMA will be available in several formats (7, 13, 33 or 65 billion parameters). The 13 billion parameter model alone could be even more efficient than GPT-3. LLaMa was created using data from academic articles found on ArXiv, books, but also based on Wikipedia and GitHub databases, or Stack Exchange.
Source : Vice
1