Mark Zuckerberg, the CEO of Meta, has recently come under scrutiny from two U.S senators, Sens. Richard Blumenthal (D-CT) and Josh Hawley (R-MO). The senators raised concerns over Meta’s groundbreaking large language model (LLaMA), which was leaked and, according to the senators, has the potential for misuse in various harmful activities, such as spam, fraud, privacy violations, and harassment. The letter from the senators questions how Meta assessed the risks before releasing LLaMA and the steps taken to prevent its abuse. The senators even accused Meta of “doing little” to censor the model.
For context, it is essential to understand LLaMA’s distinctiveness. It is one of the most extensive open-source Large Language Models currently available. LLaMA has played a key role in the current status of open-sourced LLMs, ranging from humorous chatbots to fine-tuned models with serious applications. For example, Stanford’s Alpaca open-source LLM, released in mid-March, utilizes LLaMA’s weights. Vicuna, another fine-tuned version of LLaMA, matches GPT-4’s performance, further attesting to LLaMA’s central role in the LLM space.
Meta released LLaMA in February, allowing approved researchers to download the model. However, the senators argue that the company did not carefully centralize and restrict access. Shortly after its announcement, the full LLaMA model surfaced on BitTorrent and became accessible to anyone. This wide accessibility seeded a significant leap in the quality of AI models available to the public, raising questions about potential misuse.
Interestingly, the senators seem to question whether there was a “leak” at all, placing the term in quotation marks. The focus on this issue arises in a time of rapid open-source language AI developments by startups, collectives, and academics. The letter charges that Meta should have anticipated the wide dissemination and potential for abuse of LLaMA, considering the minimal release protections in place.
In addition to releasing LLaMA for approved researchers, Meta also made its weights available on a case-by-case basis to academics, including Stanford for the Alpaca project. The weights were eventually leaked, granting global access to a GPT-level LLM for the first time.
While Meta hasn’t responded to a request for comment from Decrypt, the concerns surrounding open-source AI models’ risks and benefits have become apparent. The balance between innovation and risk remains precarious. The open-source LLM community will undoubtedly keep a close watch on the developments surrounding the LLaMA saga.
Source: Decrypt
