Meta Launches AI Language Model Llama 3

Meta on April 18 announced early versions of its latest large language model, Llama 3, along with an image generator that can update images in real time as users type in commands, a move aimed at keeping up with OpenAI, the current leader in the AI ​​market.

These models will be integrated into its Meta AI virtual assistant, which the company touts as the most advanced of its kind among free products. Meta AI compares performance on topics such as reasoning, programming and creative writing to compete with products from rivals including Alphabet’s Google and French startup Mistral AI.

The updated Meta AI virtual assistant will be more prominently integrated into Meta's Facebook, Instagram, WhatsApp, and Messenger apps, and will also have a new standalone website, putting it in more direct competition with ChatGPT - OpenAI's darling backed by Microsoft.

A welcome page for users on the new website will ask them to test out the assistant's features, such as making travel packing lists, playing '90s music quizzes, or helping with homework.

Challenge the leading position

Meta has been working to bring innovative AI products to its billions of users in a bid to challenge OpenAI's technological lead, including overhauling its costly computing infrastructure and merging previously separate research and product teams.

The tech giant has publicly released its Llama models for use by developers building AI applications as part of its efforts to catch up, as a robust free option could hamper its rivals' plans to monetize proprietary technology.

This strategy has raised security concerns from critics, who are wary that malicious actors could use the model to build anything.

Meta product manager Chris Cox said in an interview that Meta equipped Llama 3 with new computer coding capabilities and fed it images as well as text during this training, although the model currently only outputs text.

More advanced capabilities, like the ability to plan longer steps, will come in future versions, he added. Versions slated for release in the coming months will also have “multimodal” capabilities, meaning they can generate both text and images, Meta said in blog posts.

“The ultimate goal is to help you take the burden off your shoulders, to make your life easier, whether it’s interacting with businesses, writing text or planning a trip,” says Cox.

Meta product manager says that incorporating images into the Llama 3 training process will build on an update released this year to Ray-Ban Meta smart glasses, a product developed with eyewear maker EssilorLuxoticca. This allows the Meta AI to identify objects the wearer sees and answer questions about them.

Developers complained that the previous Llama 2 version of the model failed to understand basic context, confusing queries about how to “delete” a computer program with requests for instructions on murder. Rival Google has faced similar issues, recently suspending its use of the Gemini AI image-generating tool after it was criticized for creating inaccurate images of historical figures.

Meta says it's trying to fix those issues in Llama 3 by using "high-quality data" to help the model recognize subtleties. Meta doesn't reveal details about the datasets it used, only saying that it fed seven times more data into Llama 3 than it did into Llama 2.

Meta shares rose 1.8% at the end of the day on April 18 (US time).

Additionally, Meta also announced a partnership with Google to bring Google's real-time search results into its virtual assistant's responses, adding to its existing deal with Microsoft's Bing search engine.

The Meta AI virtual assistant is expanding to more markets beyond the US with this update, including Australia, Canada, Singapore, Nigeria, and Pakistan.

Meta is “still exploring the appropriate way to do this in Europe,” where privacy rules are stricter and the upcoming AI Act could impose requirements like disclosing model training data.