Meta has made available their own large language model called Llama 2, which they can download for free for both commercial and backup purposes. This new model wants to challenge in an indirect way GPT-4 on OpenAI, which powers popular chatbots like ChatGPT and Bing Chat.
Meta makes Llama 2 available, for generative AI
During Microsoft Inspire, Meta highlighted its growing collaboration with Microsoft and support for Azure and Windows at the event. The introduction of Llama 2 is intended to give companies, startups and researchers a advanced access to artificial intelligence tools. Compared to its predecessor, Llama 1, this new model was trained on a 40% larger data set.
As previously leaked, the Llama 2 training process involved a fresh mix of publicly available online data, and Meta has ensured that the model it has been supervised by safety tests to prevent illegal and dangerous use.
Meta has implemented two different variants of Llama 2 one of which focused on conversations called Llama 2-Chat. This Meta initiative aims to improve and make advanced artificial intelligence solutions accessible.
However, on its blog, Meta explains that the vast majority of data used to train the language model comes from the Anglo-Saxon world. Sergio Donato on DDay points out that 89.70% of the data used comes in English, while only 0.11% is in Italian. This means that the language properties of bots created with this technology may be limited in Italian.
Among the information released by Meta, there is also the ecological impact. Indeed, to train the model had to use 3.3 million GPU hoursfor emissions from 539 tons of CO2.
It remains to be seen which commercial products based on Llama 2 will reach the market: only then will even those who are not engineers or computer scientists be able to appreciate its potential.