LLAMA Large Language Model Release Meta, To Improve AI Capabilities

JAKARTA - Meta Platforms Inc announced on Friday February 24 that it will release a new big language model called LLAMA, abbreviating from the Large Language Model Meta AI, for researchers and institutions affiliated with government, civil society, and academics.

The move heats up competition in the world of artificial intelligence (AI) as big tech companies are racing to integrate this technology into their products and they also want to impress investors.

Big language models are used to simplify information and generate content. They can answer questions with sentences that seem as if written by humans. Meta's LLAMA is trained on 20 languages with a focus on languages that use Latin alphabets and Cyrillic.

"Today's Meta announcements appear as a step to test their generative AI capabilities so they can implement them into their future products," Gil Luria, senior software analyst at DA Davidson, was quoted as saying by Reuters.

"AI generative is a new app from Meta's AI having lacking experience, but obviously important for their business future," he added.

Meta claims that LLAMA requires "far less computing power" than previous models, and can outperform competitors who examine more parameters. LLAMA versions with 13 billion parameters can outperform GPT-3, a new predecessor of the model on ChatGPT.

Meta also claims that their LLAMA model contains 65 billion "competitive" parameters with Chinchilla70B and Google's POLM-540B. The LLAMA model is expected to enhance Meta's capabilities in the generating AI field.

The announcement shows that AI is a promising investment field in the tech industry, whose growth has slowed and led to budget cuts as well as a reduction in experimental risks.

Meta has previously released a large language model OPT-175B in May 2021 which is also aimed at researchers, and the model is the basis for making BlenderBot chatbots.

Then, Meta also introduced a model called Galactica that can write scientific articles and solve mathematical problems, but they then pulled the demo of the model after generating a false response that sounded authoritative.