The models will be integrated into virtual assistant Meta AI, which the company touts as the most sophisticated of its free peers. The assistant will get higher billing within Meta’s Facebook, Instagram, WhatsApp and Messenger apps, as well as a new standalone website that will position it to compete more directly with Microsoft-backed OpenAI hit ChatGPT .
Elevate your tech prowess with high-value skills courses
| College offering | Course | Website |
|---|---|---|
| IIT Delhi | IITD Certificate Program in Data Science and Machine Learning | Visit |
| MIT | MIT Technology Leadership and Innovation | Visit |
| Indian School of Business | ISB Professional Certificate in Product Management | Visit |
The announcement comes as Meta strives to bring generative AI products to its billions of users to challenge OpenAI’s leading position in the technology, involving an overhaul of IT infrastructure and team building previously separate research and products.
The social media giant has equipped Llama 3 with new computer coding capabilities and this time provided it with images as well as text, although for now the model will only produce text, said Chris Cox, chief product officer of Meta, in an interview.
More advanced reasoning, such as the ability to develop longer plans in multiple stages, will follow in later versions, he added. Versions scheduled for release in the coming months will also be capable of “multimodality,” meaning they will be able to generate both text and images, Meta said in blog posts.
“The ultimate goal is to help make your life easier, whether that’s interacting with businesses, writing something, planning a trip,” Cox said.
Discover the stories that interest you
Cox said including images in Llama 3’s training would improve an update rolled out this year for Ray-Ban Meta smart glasses, a partnership with eyewear maker Essilor Luxoticca, allowing Meta AI to identify the objects seen by the wearer and answer questions. about them. Meta also announced a new partnership with Alphabet’s Google to include real-time search results in Assistant responses, complementing an existing deal with Microsoft’s Bing.
Meta AI Assistant is expanding to more than a dozen markets outside of the United States with the update, including Australia, Canada, Singapore, Nigeria, and Pakistan. Meta is “still working on the right way to do this in Europe,” Cox said, where privacy rules are stricter and the next AI law is poised to impose requirements such as data disclosure training models.
The voracious need for data from generative AI models has become a major source of tension in the development of the technology.
Meta has released models like Llama 3 for free commercial use by developers as part of its catch-up effort, because the success of a powerful free option could thwart rivals’ plans to earn revenue from their proprietary technology . The strategy also sparked security concerns from critics wary of what unscrupulous developers might use to build the model.
Meta CEO Mark Zuckerberg gave a nod to the competition in a video accompanying the announcement, in which he called Meta AI “the smartest AI assistant you can freely use.”
Zuckerberg said the largest version of Llama 3 is currently being trained with 400 billion parameters and already scores 85 MMLU, citing metrics used to convey the strength and quality of AI models’ performance. The two smaller versions deployed now have 8 billion parameters and 70 billion parameters, and the latter scored around 82 MMLU, or Massive Multitask Language Understanding, he said.
Developers complained that the previous version of the Llama 2 model failed to understand basic context, confusing queries about how to “kill” a computer program with requests for instructions on how to commit murder. Rival Google has faced similar problems and recently suspended use of its Gemini AI image generation tool after coming under fire for producing inaccurate depictions of historical figures.
Meta said it reduced these issues in Llama 3 by using “high-quality data” for the model to recognize nuances. He did not give details on the datasets used, although he said he introduced seven times more data into Llama 3 than for Llama 2 and leveraged “synthetic” or AI-created data to strengthen areas such as coding and reasoning.
Cox said there is “no major change in posture” when it comes to how the company sources its training data.
