Facebook (now called “Meta”) has released LLaMA, a generative AI model (Large Language Model Meta AI). A blog post claims the model will be made public in an effort to accelerate progress in artificial intelligence research and expand participation in it. Meta has pledged its support for open science, and researchers can explore smaller and more performant models like LLaMA without investing heavily in the infrastructure needed to do so.
The report’s authors claim they trained the model on trillions of tokens using only freely available information, eschewing the use of any proprietary or otherwise inaccessible training data. Like other large language models, LLaMA predicts the next word to generate text recursively; it was trained on text from the 20 languages with the most speakers, concentrating on those employing the Latin and Cyrillic alphabets.
Meta anticipates that the public unveiling of LLaMA will inspire researchers to dig deeper into the topics of ethical AI and massive language models. The company recognizes that more excellent investigation is warranted into the potential for prejudice, toxic comments, and hallucinations in large language models. The entire AI community must work together to establish clear criteria for the responsible use of AI and large language models.
The release of LLaMA coincides with a flurry of activity among technology firms to publicize and incorporate cutting-edge AI techniques into commercial goods. According to CNBC, the new Meta model will be distinguishable from other offerings since it will come in a range of sizes, from 7 billion to 65 billion parameters. Essentially, it claimed that a 13-billion-parameter implementation of LLaMA might surpass GPT-3, a more recent predecessor to the model upon which ChatGPT is based.
It claimed that the size of its 65-billion-parameter LLaMA model was “comparable” with Google’s Chinchilla70B and PaLM-540B models, which are considerably more significant than the model Google used to flaunt its Bard chat-powered search.
Meta may be less familiar with using artificial intelligence, known as “generative AI,” but this field is crucial to the company’s development.
The tech industry’s slowing growth has resulted in significant layoffs and a reduction in experimental ventures, but AI has emerged as a bright light for investments.
Meta claimed that LLaMA might perform better than its rivals by considering fewer criteria.
Researchers will be able to request access to the model on a case-by-case basis. The LLaMA source code will be publicly available so that others can test methods for reducing or eliminating these issues in big language models. This is a departure from Google’s previously secret LaMDA and ChatGPT models.
Remember, a LLaMa that reads a lot is wool-versed.