OpenAI, Microsoft, Google and Mistral AI, to name a few, are in a frantic race to develop the best of the major language models. Billions of dollars are devoted to it and new versions of these much talked about AI are released regularly. The next one to appear is likely to attract a lot of attention, since it is Llama 3, developed by the social media giant Meta.
During its Innovation Day, which took place on Wednesday April 10, 2024, the American company announced that its model would be launched in May. At least in its complete version, capable of standing up to some of the best proposals on the market. Until then, to make us want to use Llama 3, Meta will launch “a number of varied models with different capacities“, according to Nick Clegg, president of international affairs of the company. We should therefore discover versions closer to Gemini Nano or Claude Haiku, while waiting to see this great language model deploy its full potential.
If Clegg remained vague on the timetable, The Information suggests that the first drafts of Llama 3 could be released to the public as early as next week.
Significant progress, but still much work to be done
Meta’s next AI is expected to be less powerful than GPT-4, and have up to 140 billion parameters, according to some rumors. However, the important thing for Llama 3 will be to correct the trajectory taken by its predecessor, equipped with safeguards considered too strict by observers.
Ultimately, the vocation of this AI, in all its iterations, will be to power Meta’s products, from Facebook to Instagram via WhatsApp and Messenger. The company has big ambitions for Llama 3, which it wants to project to the rank of “the most useful assistant in the world“, according to Joelle Pineau, vice-president of AI research at the American giant. “There is still a lot of work to get there“, she underlines, and we will have to see if the progress made so far will be enough to convince users who risk being increasingly difficult to impress.