Meta is expanding its commitment to open-source artificial intelligence, making its large language model LLaMA available for download, according to a report from The New York Times. The model competes with OpenAI's GPT-4, which is not open source.
Meta first released LLaMA's underlying code in February through an email-gated process, requiring users to request access and be deemed safe by the company. The code was leaked to 4Chan almost immediately, according to the NYT report.
Researchers at Stanford used the leaked model to build an AI system and study its behavior. They found it could generate what the report described as seriously problematic text. Stanford researcher Moussa Doumbouya reportedly told colleagues that making the technology widely available would be like making "a grenade available to everyone in a grocery store," according to the NYT.
Meta's Yann LeCun Disagrees on Risk
Meta's chief AI scientist, Yann LeCun, pushed back on that assessment. "You can't prevent people from creating nonsense or dangerous information or whatever," LeCun told the NYT. "But you can stop it from being disseminated."
The report noted that at-scale content moderation is already imperfect, even without generative AI, suggesting LeCun's argument faces practical challenges.
Meta's open-source strategy contrasts with the approach taken by OpenAI and Microsoft, which have kept their models proprietary. Which approach will prove more effective remains unresolved as competition in AI intensifies.
Editorial Notes