Meta Platforms on Thursday announced an early version of its latest large-scale language model, Llama 3, and an image generator that updates images in real time while users type prompts, as it races to catch up with generative AI market leader OpenAI. was released.
These models will be integrated into the virtual assistant’s Meta AI, which the company touts as the most sophisticated of its peers’ products available for free, and will be used for topics such as inference, coding, and creative writing. In terms of performance, it compares its performance with rival products from Alphabet Inc.’s Google and Google. French 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 equipped Llama 3 with new computer-coding capabilities to provide images as well as text for this training, but for now the model will only output text, Chris Cox, Meta’s chief product officer, said in an interview. Told.
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.
Cox said incorporating images into Llama 3’s training will enhance updates rolling out this year to the 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 respond. Questions about them.
Meta stock closed up 1.5% on 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.” Privacy rules are stricter in Europe, with the next AI law expected to impose requirements such as disclosure of model training data.
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 (8 billion parameters and 70 billion parameters) outperform other free models in performance benchmarks commonly used to assess model quality. He said that he had a favorable score against. 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.
Mehta said Llama 3 has mitigated these issues by using “high quality data” to make the model aware of nuances. He did not elaborate on the dataset used, but said he fed Rama 3 with seven times the data he used for Rama 2.
