LLaMA, the artificial intelligence engine developed by Meta, has been circulating on the Internet unsupervised in the form of a torrent for several days, allowing any user to download the language model and use it without supervision from the company.
The leak occurred last week from the well-known and controversial 4chan forum, where several users have been comparing for months the different capabilities of generative artificial intelligences that have been available for months, such as GPT-3 (created by OpenAI) or LaMDA (Google ).
Until now, Meta offered free and full access to this language model, but only by invitation and for users with an academic profile. The objective was to promote research on these types of models to try to find solutions to the most common problems, such as the tendency to fabricate data or the ease with which they sometimes acquire hostile response tones.
Although there are large language models openly available, they often require a complex infrastructure to train or function, limiting the number of institutions that can access them. Its quality is usually not up to the standards of those created by large technology companies such as Google or OpenAI.
“Even with all the recent advances in long language models, access remains limited due to the resources required to train and run these very large models. This limits researchers’ ability to understand how and why these work.” large linguistic models, hindering progress in efforts to improve their robustness and mitigate known problems, such as bias, toxicity and the potential to generate misinformation,” they explain from Meta.
LLaMA, developed by the company’s research team, stands out for being more compact than other models and is available with a variable number of parameters (from 7,000 to 65,000 million parameters) that makes it easier to run on more modest equipment.
Still, it requires a fairly large infrastructure, with a minimum of 30 GB of graphics memory (language models and other applications in the AI ??domain benefit from parallel processing power, which is often the strength of graphics processors)
There are at least two versions of LLaMA circulating through torrents, including the more complete model with 65 billion parameters. Several users have optimized their operation in recent days so that they can be run on conventional home computers.
Like other language models, LLaMA works by taking a sequence of words and predicting the next one to generate long texts. Meta has trained you with millions of texts in the 20 most widely spoken languages, focusing on Latin and Cyrillic alphabets.
According to the criteria of The Trust Project
