The success of Meta AI hinges on the capabilities of Llama 3, the next iteration of Meta's foundational open-source large language model. Image: Tech & Startup/Dall-E
Meta is launching a major update to its Meta AI assistant and its underlying large language model, Llama 3. The Meta AI assistant, first introduced last September, is now being integrated across Meta's core platforms including Facebook, Instagram, Whatsapp and Messenger. Users can interact with the assistant through search bars, message inboxes and even directly within the Facebook feed. Additionally, a standalone web app for Meta AI has been launched for the first time.
The success of Meta AI hinges on the capabilities of Llama 3, the next iteration of Meta's foundational open-source large language model. Meta claims Llama 3 outperforms competitors in key benchmarks and excels at tasks like coding. Two versions of Llama 3, with 8 billion and 70 billion parameters respectively, are being released for both internal use within the Meta AI assistant and for external developers. A much larger multimodal version is planned for release in the coming months.
Meta AI boasts several unique features. It is the only known chatbot to integrate real-time search results from both Google and Bing. It can also generate image animations and high-resolution images on the fly. The assistant offers prompt suggestions to guide users on its capabilities, aiming to make it more user-friendly.
While initially launched in the US, Meta AI is now expanding to English-speaking countries including Australia, Canada, and several African and South Asian nations. This rollout brings the assistant closer to Meta's vast user base of over 3 billion.
The strategy echoes Meta's past successes with Stories and Reels, features inspired by competitors but implemented at a larger scale. CEO Mark Zuckerberg acknowledges that Meta AI is not yet a major player in the AI assistant field, but believes this broader rollout will introduce it to a wider audience.
Alongside the consumer-facing launch, Meta is releasing two versions of Llama 3 as open-source models for developers. This exemplifies the rapid development in the field of AI models. The largest version of Llama 3 boasts over 400 billion parameters and is trained on a dataset 7 times larger than its predecessor.
A key focus for Llama 3 was reducing "false refusals" - instances where the model declines a harmless prompt. Zuckerberg highlights the ability to answer previously refused queries, like creating a drink recipe or offering breakup advice. The company is yet to decide on open-sourcing the largest version of Llama 3, but Zuckerberg downplays safety concerns as a reason for withholding it.
Meta acknowledges the lack of standardised benchmarks for evaluating AI models. The company remains tight-lipped on the specifics of the data used to train Llama 3, although it confirms using a mix of public internet data, synthetic AI-generated data, and excluding user data.
The rapid advancements in AI models make it difficult to predict future dominance. While Meta positions itself as a leader with Llama 3, rumours suggest OpenAI is developing GPT-5, potentially surpassing current models. Zuckerberg emphasises Meta's focus on long-term development, already planning for Llama 4 and 5. The company's goal, he states, is to "compete with everything out there and to be the leading AI in the world."
