Protestors to Meta AI: “Sharing model weights is fundamentally unsafe”
WHO: Concerned citizens
WHAT: Protest against Meta (formerly Facebook) releasing large language model (AI) weights
WHERE: The Meta building in San Francisco, 250 Howard St, San Francisco, CA 94105 (event page)
WHEN: 4:00pm Friday September 29, 2023
WHY: Protestors will be gathering outside the Meta building in San Francisco on September 29th to call for “responsible release” of their cutting edge AI models. The protestors demand that Meta stop releasing the weights of their AI models publicly— what they term “irreversible proliferation”.
Protest organizer Holly Elmore says “Unlike other top AI labs, Meta says that they plan to keep releasing the weights of their models publicly, which allows anyone with an internet connection the ability to modify them. Once model weights are released, they cannot be recalled even if it becomes clear the model is unsafe– proliferating the AI model is irreversible. Chief AI Scientist at Meta, Yann LeCun, has said he’s not worried if bad actors gain access to powerful AIs, even if the AIs have human-level or superhuman capabilities. This is unacceptable. New tech shouldn’t be creating a Wild West where my good AI fights your bad AI– AI should be regulated like we regulate pharmaceuticals or airplanes, so that it is safe before it is released.”
When large language models (LLMs) are accessed through an API, users access a version that has been fine-tuned to ensure user safety and there can be additional input/output filtering for safety that does not take place when users are directly running the model from the weights. Meta recently released the model weights of its LLM, LLaMA 2, and the company has said that they will continue to release the weights of new models, trusting that “good guy AIs” will win against “bad guy AIs”.
Safety measures for LLMs like LLaMA 2 aren’t just about whether they say hateful things. Not only can the safety fine-tuning be stripped from models if one has the weights, Meta released a base version of LLaMa 2 without any safety fine-tuning. This is dangerous because LLMs are increasingly being developed as autonomous agents without the need for a human in the loop. Researchers have found that LLMs can be used to scale phishing campaigns, suggest cyberattacks, synthesize dangerous chemicals, or help plan biological attacks. Over the coming years, as their general capabilities increase, they will only become more capable of malicious activities.
When asked about the protest, Dr. Peter S. Park, AI Existential Safety Postdoctoral Fellow at MIT and Director of StakeOut.AI, said “Already, widely released AI models are being misused for non-consensual pornography, hateful content, and graphic violence. And going forward, Llama 2 and successors will likely enable cybercriminals, rogue actors, and nation-state adversaries to carry out fraud and propaganda campaigns with unprecedented ease.”
Sign up for the free insideBIGDATA newsletter.
Join us on Twitter: https://twitter.com/InsideBigData1
Join us on LinkedIn: https://www.linkedin.com/company/insidebigdata/
Join us on Facebook: https://www.facebook.com/insideBIGDATANOW