Are Meta’s statements about the LlaMA models exaggerated?

Are Meta’s statements about the LlaMA models exaggerated?

San Francisco, 3/6/2023

Meta Platforms has released its new LLaMA (Large Language Model Meta AI) – but for now only for researchers.

Meta-AI researchers have developed a collection of LLaMA models that they claim can outperform OpenAI’s GPT-3 despite being smaller.

A Meta spokesperson explained that these new language models will not be used in Meta’s products like Instagram or Facebook. The models range in size from 7 to 65 billion parameters, which is significantly smaller than the 175 billion parameters of GPT-3.

One of these models, LLaMA-13B, for example, is ten times smaller than GPT-3 and, according to one project member, should still outperform it in “most benchmarks”. All LLaMA models were trained using publicly available datasets such as Wikipedia and Common Crawl.

Meta says it will make the models available to academic researchers “on a case-by-case basis.” It goes on to say that the researchers will be “connected with organizations in government, civil society and academia, as well as with industry research laboratories.”

Hits: 5