It appears like Meta is on its way to making big advances in artificial intelligence. Mark Zuckerberg said on Thursday via Instagram that Meta's huge language model and AI assistant are receiving a number of updates. In addition to additional advancements, Zuckerberg told podcaster Dwarkesh Patel that Meta was eagerly looking forward to bringing multimodality, multi-linguality, and wider context windows'.

Highlights:

• Meta CEO Zuckerberg unveils updates to language model and AI assistant, emphasizing multimodality. 
• Llama 3 models and image generator released, targeting generative AI market. 
• MetaAI integrates into Meta's platforms, for standalone usage, competing with Google, MistralAI.

In addition to releasing early versions of Meta's most recent LLM, Llama 3, the company also launched an image generator that refreshes visuals in real time as the user writes instructions. This is viewed as Meta's attempt to catch up in the generative AI market, which is currently controlled by OpenAI.

The first Llama 3 models, available in two sizes (8B and 70B parameters), have been integrated into MetaAI, the company's artificial intelligence assistant.

MetaAI seeks to soar.

Meta is presently presenting its virtual assistant as a smart AI that outperforms its colleagues in areas such as thinking, coding, and creative writing, competing with models owned by Google and even French AI firm MistralAI. The most recent version of Meta AI assistant will be prominently featured in the Facebook, Instagram, WhatsApp, and Messenger applications. It will also apparently have its own website, making it a direct competitor to OpenAI's ChatGPT. 

When questioned about Meta AI's latest models, Zuckerberg assured Patel that a new version, Llama-3, will be released. He stated that the models are open-source for the use of the developer communities and will power Meta AI. "We now believe that Meta AI is a highly intelligent, publicly available AI assistant that individuals may use. "We're furthermore merging Google and Bing for real-time knowledge," he explained. 

Zuckerberg described the key aspects of Meta AI's latest releases, stating that Meta will make it 'a lot more prominent throughout its apps'. We've introduced a bunch of new creation features that I believe are very fantastic and that people will appreciate. Animation is a good thing. You can animate any image. It now produces high-quality photos so quickly that it generates them as you type and refreshes them in real time," he said to the host.

Zuckerberg offered more information about Llama 3, stating that his organization has been training three versions: the 8B parameter model, the 70B model, and another 405B dense model that is still being trained. The CEO stated that his company is also introducing its 82MMLU, which has top ratings in math and logic. "The 8 billion (model) is nearly equally powerful as the most powerful version of Llama-2 that we released."

At the moment when AI models are becoming more evaluated based on their coding ability, Meta AI models appear to be an outlier. Zuckerberg explained that when they went to work on Llama 2, they did not prioritize coding since he believed that people would not ask Meta AI many coding-related questions on WhatsApp. 

He admitted that over the previous 18 months, coding has become a significant factor in several domains. “Even if people don't have coding questions, teaching the models on programming helps them grow more systematic in resolving the question and helps them think across a lot of various kinds of domains.”