Facebook’s ChatGPT rival leaks online
Facebook’s large language model, which the company’s parent Meta Platforms is positioning as a potential rival for OpenAI’s ChatGPT, has leaked online.
LLaMa, which stands for Large Language Model Meta AI, is officially only available for non-commercial use by certain Meta-approved users.
These can include AI and machine learning researchers, civil society organisations, and government officials.
But a torrent file to download the tool has appeared on multiple websites in the past few days.
The file was initially uploaded to 4chan before spreading to other places on the Internet, including a GitHub repository.
According to a report from 80.lv the first person to spot its emergence was Replit CEO Amjad Masad.
According to Vice, several users who downloaded the LLaMa tool could run it locally on their machines.
Other prominent language models from Google, Microsoft, and OpenAI are only available to the public through purpose-built consumer interfaces or APIs, which come with limitations to prevent abuse.
However, some malicious users have been able to bypass some of ChatGPT’s security mechanisms and get it to create phishing emails and malware code.
It is unclear what the ramifications of a leaked LLaMa could be, but the tool presumably won’t have the types of mitigations in place with the current version Meta provides to legitimate users.
Some developers have pointed out that the leak is not that big of a deal, as it was fairly easy to get legitimate access to the tool.
“All I did was put down that I’m a student studying machine learning and wanted to test the model, no proof required. Got access in a few days,” one Redditor said.
Others have said they struggled to get approval from Meta, with their applications seemingly being rejected for no reason despite falling into one of the qualifying testing categories.
Now read: Google’s Universal Speech Model a “critical first step” towards 1,000 languages
Loading ...