Meta's LLaMA AI Model Leaks to Public Via 4chan
Facebook's proprietary large language model, LLaMA, leaks onto 4chan, marking the first time a major tech firm's private AI escapes to the public.
Facebook's powerful large language model, known as LLaMA, leaks online and becomes available to anyone after a 4chan user shares a torrent file. This unprecedented event marks the first time a major tech company's proprietary AI model escapes to the general public, bypassing the strict approval process that normally limits access to approved researchers and government officials.
Unlike competitors such as Google, Microsoft, and OpenAI that keep their newest models locked behind APIs or consumer interfaces, Meta previously shared LLaMA with a select group of researchers. Meta confirms the leak but defends its approach, stating that sharing state-of-the-art AI models helps the research community evaluate and improve the technology while balancing responsibility and openness.
4chan users report successfully running the LLaMA model on their own personal machines, though the exact implications of this widespread access remain unclear. The AI model operates similarly to OpenAI's GPT-3, using trillions of tokens scraped from sources like Wikipedia, GitHub, and academic papers to predict and generate human-like text across its various sizes.