Meta's LLaMA AI Model Leaks Online Via 4chan Torrent
Meta's restricted LLaMA language model leaks onto the internet through 4chan, allowing unauthorized public downloads. The leak includes the 7B and 65B model weights and mocks Meta's official application process.
Meta's newest large language model family, LLaMA, surfaces online illegally as its model weights become freely accessible via torrents. The AI model is originally designed as an open-source package exclusively for approved researchers, government officials, and civil society members who complete an application process.
An anonymous 4chan user uploads torrent files for the LLaMA 7B and 65B models on the platform's technology board. The leaker also submits the torrent link to the official LLaMA GitHub page as a sarcastic pull request titled 'save bandwidth by using a torrent to distribute more efficiently' alongside Meta's official Google Forms access link.
The leaked model functions like other large language models by predicting the next word in a sequence to generate text. LLaMA comes in various sizes, with the larger versions training on 1.4 trillion tokens gathered from diverse sources including Wikipedia, GitHub, Stack Exchange, and academic papers.