LLaMA, the artificial intelligence designed by Meta, was put online by a user, without the knowledge of the company, on the 4chan forum. He bypassed the group’s approval system to get his hands on the language model. This AI is the alternative to ChatGPT.
In the wake of Google and Microsoft, Meta has also embarked on the race for artificial intelligence. Mark Zuckerberg’s group has announced a language model called LLaMA, for Large Language Model Meta AI. Designed as an alternative to GPT, the linguistic model behind ChatGPT, LLaMA is able to generate text in response to its interlocutor.
Initially, Meta reserves its generative artificial intelligence for researchers approved by it. To test the AI, they must submit an application. Meta grants access “on a case-by-case basis to academic researchers, those affiliated with the government, civil society and academia organizations, and industry research laboratories around the world”. To submit a request, the researcher must explain what he intends to do with LLaMA and provide examples of studies carried out by him.
At the moment, therefore, access to the linguistic model is very limited. It is not yet planned that all Internet users can exchange with LLaMA, as is currently the case with ChatGPT or Prometheus (Bing’s chatbot).
Meta’s LLaMA AI ends up on 4chan
Unfortunately for Meta, its language model was uploaded without its knowledge. On Thursday, March 2, 2023, the linguistic model indeed appeared on the 4chan forum. A user has shared a torrent file to install LLaMA on a computer.
Owing to this, some Internet users claim to have installed the model on their PC. Many testimonies support this. Interviewed by The Verge, several AI experts explain that they launched the model on a machine using the torrent. According to them, the files correspond well to the documents provided by Meta to university researchers accepted into the program.
Contacted by Vice, Meta confirms in half words the leak of LLaMA. The Menlo Park company remains evasive, but suggests that an individual has circumvented the system in place to determine which researchers can gain access to AI:
“LLaMA was shared for research purposes, consistent with how we shared previous large language models. Although the template is not available to everyone, and some have tried to circumvent the approval process, we believe the current release strategy allows us to balance accountability and openness. ”
At this point, we don’t know much about the leak
On Reddit, some netizens are claiming the files were uploaded by a researcher who got approval from Meta. In any case, this is what the unique identification code, provided by Meta to approved researchers, suggests.
Before putting the files online, the user forgot to remove the code. In theory, Meta was, therefore, able to trace the origin of the leak by simply consulting the torrent files.
The files contain all versions of the linguistic model. The firm has indeed designed four versions of the model, ranging from 7 billion parameters to 65 billion parameters. The version limited to seven billion requires less computing power. This is why it can run on a simple computer. 4chan netizens were able to run the model on a PC with an RTX 3060 or Ryzen 7900X CPU. Note that it still takes a bit of technical expertise to take advantage of the leak.
READ ALSO:
Type “Google Gravity” And click Pn “I’m feeling lucky”: You Will Never Believe What Is Happening!
A great first
This is the first time ever that a proprietary language model has been publicly disclosed through a leak. So far, companies involved in the race for AI, such as OpenAI or Google, have managed to padlock the use of their tools to avoid abuses.
For example, OpenAI only grants access to its chatbot through an interface for the general public, i.e. its website, or an API for enterprises. In this way, the American start-up is always able to supervise the use of its robot, with its rules and guidelines.
Unsurprisingly, Meta is doing its best to prevent its linguistic model from spreading on the web. The Californian group has ordered several websites to remove the installation files. They had, for example, found themselves on the open-source platform Hugging Face, focused on the design of apps based on machine learning.
Meta asked the platform to remove the files, claiming they are an “unauthorized distribution” that “constitutes copyright infringement or misuse”. Despite Meta’s relentlessness, many copies of the LLaMA files ended up online, including on GitHub and Reddit.