Meta's large language model LLaMA leaks online
A torrent of the system was uploaded on 4chan earlier this month and has since been shared across multiple AI communities
LLaMA, the latest AI language model developed by Meta, has been leaked online and is available for download.
Meta made an announcement about LLaMA two weeks ago. The company said at the time that rather than releasing the model as a public chatbot, it would offer it as an open-source package that members of the AI community can request access to.
Meta said that its intention was to democratise access to AI and encourage research into its challenges.
However, Meta's attempts to regulate access to LLaMA seem to have been unsuccessful.
Merely a week after the company began receiving requests for LLaMA access, the model was leaked online.
On March 3rd, a torrent of the system was uploaded on the 4chan forum and has since been shared across multiple AI communities.
Furthermore, instructions on how to download the model have also been published on GitHub.
This marks the first time a major technology company's confidential AI model has leaked to the general public.
Thus far, companies such as Google, Microsoft and OpenAI have kept their latest models confidential, only allowing access through consumer interfaces or an API to limit the risk of misuse.
Some members of 4chan claimed that they have been running LLaMA on their personal machines.
Several AI researchers who have downloaded the leaked system told The Verge that it is legitimate.
AI researcher Matthew Di Ferrante was able to compare the leaked version with the official LLaMA model that Meta had distributed and confirmed that they were identical.
Shawn Presser, an AI engineer who wrote the download instructions on Microsoft's code-sharing platform GitHub, showed screenshots to The Register, demonstrating that he was able to generate text from the model successfully.
Presser believes that it is better to release the model without any restrictions instead of limiting access only to approved academics.
"I think the good will outweigh the bad, by at least tenfold. Probably closer to 100x," he said.
LLaMA is not a single system, but instead comprises four models of various sizes, which are quantified in billions of parameters. These four models are LLaMA-7B, LLaMA-13B, LLaMA-30B, and LLaMA-65B.
Meta asserts that the 13 billion-parameter version (13B) surpasses OpenAI's 175 billion-parameter GPT-3 model on several AI language model benchmarks.
LLaMA-13B can operate on a single A100 GPU, an enterprise-level system that is comparatively accessible and can be rented for a few dollars per hour on cloud platforms.
Di Ferrante said that anyone who is knowledgeable about configuring servers and development environments for complicated projects should be able to operate LLaMA if they are given enough time and proper instructions.
Stella Biderman, director of EleutherAI, a non-profit AI research laboratory, and a machine learning researcher at Booz Allen Hamilton, stated that the model's computational requirements would be the most significant limitation on its effective use.
"Most people don't own the hardware required to run [the largest version of LLaMA] at all, let alone efficiently," Biderman said.
Even though the model has been leaked, Meta said it would continue to share LLaMA with approved researchers.
"It's Meta's goal to share state-of-the-art AI models with members of the research community to help us evaluate and improve those models," a spokesperson said.
"LLaMA was shared for research purposes, consistent with how we have shared previous large language models.
"While the model is not accessible to all, and some have tried to circumvent the approval process, we believe the current release strategy allows us to balance responsibility and openness."