meta platform races to catch up with generative AI market leader OpenAI, with an early version of Llama 3, its latest large-scale language model, and an image generator that updates images in real-time while users type prompts. has been released.
These models will be integrated into virtual assistant Meta AI, which the company touts as the most sophisticated of its peers' products available for free, for topics such as inference, coding, and creative writing. , citing performance comparisons with rival products from Google and France. Startup Mistral AI.
The updated Meta AI assistant will receive more prominent billing within Meta's Facebook, Instagram, WhatsApp, and Messenger apps, as well as a new standalone website and collaboration with Microsoft-backed OpenAI's blockbuster ChatGPT. You will be able to compete more directly.
The landing page that greets the site's visitors offers things like having your assistant create a vacation packing list, playing 1990s music trivia with you, helping with homework, and the New York City skyline. I encourage them to try drawing a picture.
Meta involves an expensive overhaul of its computing infrastructure and the merging of previously separate research and product teams to create dozens of generative AI products to challenge OpenAI's leading position in the technology. It is desperately trying to push it out to 100 million users.
The social media giant is rolling out the Llama model, which can be used by developers building AI apps, as part of a catch-up effort, as a strong free option could thwart rivals' plans to monetize their proprietary technology. It is openly released. This strategy has raised safety concerns from critics who are wary of what malicious attackers might build using this model.
Meta has added new computer-coding capabilities to Llama 3, and while the training involved inputting images as well as text, the model currently only outputs text, said Chris Cox, Meta's chief product officer. said in an interview.
multimodality
More advanced inference (such as the ability to create longer multi-step plans) will follow in subsequent versions, he added. The version expected to be released in the coming months will also be capable of “multimodality,” meaning it can generate both text and images, Mehta said in a blog post.
“The ultimate goal is to take the hassle out of work and make your life easier, whether it's dealing with businesses, writing, or planning a trip,” Cox said.
By including images in Rama 3's training, Cox enhances an update rolling out this year to Ray-Ban Meta smart glasses, a product with eyewear manufacturer Essilor Luxoticca, that allows Meta AI to identify objects visible to the wearer. He said he would be able to answer questions. About them.
Read: Meta’s extraordinary comeback
Meta stock rose 1.8% late Thursday.
Meta also announced a partnership with Google to include real-time search results in Assistant responses, complementing an existing arrangement with Microsoft's Bing search engine.
This update expands the Meta AI Assistant to more than 10 markets outside the US, including Australia, Canada, Singapore, Nigeria, and Pakistan. Cox said Meta is “still figuring out the right way to do this in Europe,” where privacy rules are stricter and the upcoming AI law will impose requirements such as disclosure of model training data. He said it was planned.
The voracious need for data for generative AI models has emerged as a major source of tension in technology development.
Meta CEO Mark Zuckerberg gave a nod to the competition with OpenAI in a video accompanying the announcement, in which he called Meta AI “the most intelligent AI assistant at your disposal.”
Zuckerberg said that the two smaller versions of Llama 3 currently available have 8 billion parameters and 70 billion parameters, which are performance benchmarks commonly used to assess model quality. He said it scored more favorably than other free models. The largest version of Rama 3 is still being trained, he said, and has 400 billion parameters.
Nathan Benaich, founder of AI-focused venture firm Air Street Capital, said while these results were “absolutely impressive,” the performance gap between the free and proprietary models has widened. He said that it shows that
Developers complained that previous Llama 2 versions of the model did not understand basic context, confusing queries about how to “kill” a computer program with requests for instructions to carry out a murder. states. Rival Google has encountered similar issues, recently suspending the use of its Gemini AI image generation tool after it drew criticism for producing a large number of inaccurate depictions of historical figures.
Read: The meta of bringing in-house chips to power AI
Mehta said Llama 3 has mitigated these issues by using “high quality data” to make the model aware of nuances. Although he did not elaborate on the dataset used, he said that he fed Rama 3 with seven times the data that he used in Rama 2. Katie Paul and Jeffrey Dustin, (c) 2024 Reuters