Move over OpenAI’s ChatGPT, Google’s Bard, and Microsoft’s Prometheus—there’s yet another large language model-powered artificial intelligence in town. Meta, the company formerly known as Facebook, introduced its own AI today, called LLaMA. The animal-esque acronym predictably stands for Large Language Model Meta AI.
CEO Mark Zuckerberg described his company’s contribution to the buzzy AI technology sphere in, what else, but a Facebook post.
“Today we’re releasing a new state-of-the-art AI large language model called LLaMA designed to help researchers advance their work,” Zuckerberg wrote on his social media platform. He added that “LLMs have shown a lot of promise in generating text, having conversations, summarizing written material, and more complicated tasks like solving math theorems or predicting protein structures.” But the Meta exec did not explain exactly which (if any) of those tasks LLaMA could currently accomplish.
In fact, the only detail that Zuckerberg offered on the large language model in the Friday announcement is that his company is “committed to this open model of research and we’ll make our new model available to the AI research community.”
In a company blogpost published Friday afternoon that included a link to a full research paper on the AI and its Github model card, Meta offered significantly more information. “Like other large language models, LLaMA works by taking a sequence of words as an input and predicts a next word to recursively generate text,” the company wrote. Meta added that LLaMA was trained on text from 20 different languages. It’s model card notes that training data included publicly available text from CCNet, C4, Wikipedia, ArXiv, and Stack exchange.
Meta described its AI as a “smaller foundation model,” that “requires far less computing power and resources,” than other large language models, and said LLaMA will be available in multiple sizes. The company further emphasized its commitment to transparency and responsible AI development, and reiterated that only AI researchers will be given access.
“To maintain integrity and prevent misuse, we are releasing our model under a noncommercial license focused on research use cases. Access to the model will be granted on a case-by-case basis to academic researchers; those affiliated with organizations in government, civil society, and academia; and industry research laboratories around the world,” the company wrote.
So far, at this stage in the research, LLaMA has not been incorporated into any of Meta’s products or platforms (including Instagram and Facebook), a company spokesperson told Gizmodo via email. The spokesperson further noted that the company has nothing to share, at this time, about a public preview or expanded, public access.
At the moment, it’s hard to say how LLaMA will stack up against other tech giants’ AI attempts in a rapidly crowding field. However, just about every recent large language model launch so far has come with its own share of snafus. Google and Microsoft both integrated AI-text generators into their search platforms, and both ended up unwittingly advertising inaccurate information. Then, there’s the unresolved questions of copyright, privacy, whether or not AI can develop “feelings,” and how to manage the workarounds many on the internet are finding to bypass restrictions and generate offensive content or even malware.
Meta acknowledged some of these pitfalls in its blogpost. “There is still more research that needs to be done to address the risks of bias, toxic comments, and hallucinations in large language models. Like other models, LLaMA shares these challenges,” the company wrote. However it seems Meta believes its open, research-oriented approach is a step forward towards resolving AI’s numerous, ongoing issues.
Those hoping to access LLaMA for themselves can apply via this request form.
Update 2/24/2023, 2:16 p.m. ET: This post has been updated with additional information from a Meta spokesperson.
Update 2/24/2023, 2:04 p.m. ET: This post has been updated with information from a Meta company blog.