(Credit: Meta)
Meta today rolled out an updated version of its AI assistant running Llama 3, the newest version of the country's large language model (LLM), which it is also open sourcing.
"With this new model, we believe Meta AI is now the most intelligent AI assistant that you can freely use," Meta CEO Mark Zuckerberg said in an Instagram video announcing the update.
Meta AI will be integrated into the search boxes on WhatsApp, Instagram, Facebook, and Messenger, as well as meta.ai on the web. Zuckerberg also says it will pull "real-time knowledge" from Google and Bing.
"We also built some unique creation features, like the ability to animate photos," Zuckerberg says. "Meta AI now generates high-quality images so fast that it creates and updates them in real-time as you're typing. It'll also generate a playback video of your creation process."
Under the hood is Llama 3. Meta today open-sourced the first set of Llama 3 models at 8 billion and 70 billion parameters, and it's currently training a 400 billion model (in theory, the more parameters, the more nuanced and complex the model).
Meta says it used a training dataset seven times larger than Llama 2 and made improvements to pre-training and post-training to reduce the number of errors while improving reasoning skills significantly. It also uses a tokenizer with a vocabulary of 128K tokens "that encodes language much more efficiently."
Llama 3's datasets cover 30+ languages, although Meta says it performs the best in English.