Meta today released a compact version of its latest open source AI model, Llama 3, as well as a new update for Meta AI, its assistant, which can now answer questions using real-time information from Google and Bing.
Why it matters: Meta is demonstrating that its open source AI models can continue competing with Google, Anthropic, Microsoft, and OpenAI's closed models.
- Meta is betting that the convenience of having a powerful AI assistant within their apps could help preserve their high-margin, ad-based business model — keeping users engaged instead of going elsewhere for AI benefits.
How it works: The Meta AI assistant is available in the search box at the top of WhatsApp, Instagram, Facebook, and Messenger, and via a new website at meta.ai.
- The company claims Meta AI is the "most intelligent" free AI assistant.
- A new feature in its AI image generator called "Imagine" allows users to create images from text in WhatsApp and the Meta AI website.
- CEO Mark Zuckerberg said in a statement that the service will update images in real-time as you type in more detailed prompts, and can create custom animated GIFs.
- Outside of the U.S., Meta AI will now be available in Australia, Canada, Ghana, Jamaica, Malawi, New Zealand, Nigeria, Pakistan, Singapore, South Africa, Uganda, Zambia and Zimbabwe.
Case in point: ""Let's say you're planning a ski trip in your Messenger group chat. Using search in Messenger you can ask Meta AI to find flights to Colorado from New York and figure out the least crowded weekends to go – all without leaving the Messenger app," the company wrote in a blog post.
What's inside: Llama 3 is pre-trained on over "15 trillion tokens that were all collected from publicly available sources. Our training dataset is seven times larger than that used for Llama 2, and it includes four times more code."
- Around 5% of the training data is "high-quality non-English data that covers over 30 languages," though Meta does not expect the "same level of performance in these languages as in English."
- For any chatbot built with Llama 3, including Meta AI, Meta's researchers found that if a user asks "a reasoning question that it struggles to answer, the model will sometimes produce the right reasoning trace: The model knows how to produce the right answer, but it does not know how to select it."
What they're saying: "It's pretty wild," CEO Mark Zuckerberg said in a statement, adding that the smaller of the two models released today "the 8B is already nearly as powerful as the largest Llama 2 model."
What's next: Meta is still training a version of Llama 3 that is nearly six times more powerful than the biggest of the two models released today — based on 400 billion parameters compared to 70 billion.
- "Our goal in the near future is to make Llama 3 multilingual and multimodal, have longer context" and improved "reasoning," according to the blog post.
- Meta AI will be available on Meta Quest headsets later in the year.
Reality check: While open AI models have benefits for economy-wide innovation, and make it easy for others to make customized versions of the models, Stanford researchers say that any attempt to impose conditions on users of open models are "easy for malicious actors to ignore."
